Sure, this is a small anecdotal datapoint, but since I've gotten Qwen3.8-Flash-Next running well on my Strix Halo box my Claude Pro subscription use has dropped off a cliff. I'm only at 18% of quota usage this week - this being Friday I would've been at 80% at least in the past. This is just an astonishingly good model that runs locally and quite well on my PC. I do not plan to renew my Claude Pro subscription when it comes up in a couple of months.
I really hope the future is local models, because the idea that all companies will need to buy tokens from a handfull of tech companies to run their businesses or make their products is awful.
I've been ripped (on this site) for making this prediction. While the frontier (i.e. The toughest problems) will always push the limits of available computing power, what typical people do with AI won't always be on the frontier, nor will what appears on the frontier necessarily be lucrative. The frontier used to be playing Chess, Atari, and Go. These were challenging and really advanced AI, but they didn't make a lot of money. Now a phone can do these things.
You're an example of this process. You used to require a frontier model running on a data centre. Algorithms improved and, perhaps, your hardware as well, but your requirements didn't go up. Now you can do what you need to do with local hardware. You're one AI user who might not need to send money to Anthropic going forward unless you find a new frontier application that interests you.
This is why I'm a little bearish when it comes to the valuation of the AI giants. Yes, there are some extremely lucrative things on the frontier right now but, in the future, people may shift a lot of what they're doing now to local hardware while new things on the frontier may not produce the kind of profits that will pay back trillions in debt. AI companies have bet big on themselves being able to constantly and perpetually find new applications, and also that those new applications will make money.
I agree with you that if the trends of 2025-2026 continue into 2027-2028 for open weight models and local inference, 120b param open models will be more than good enough for 80% of today's use cases.
If so how can you justify OpenAI and Anthropic and the AI buildout?
I suppose I should have said they just didn't go up faster than the capabilities of local models are improving. I suspect local model/hardware capability will scale up faster than most people's needs when it comes to coding. That's likley why Sam Altman has naughty designs on your RAM.
> the idea that all companies will need to buy tokens from a handfull of tech companies to run their businesses or make their products is awful.
To the leadership of these tech companies that actually is the dream. Sam Altman said the quiet part aloud recently: "From a global allocation perspective, actually, it's quite terrible to have everybody have a bunch of the world's limited supply of DRAM sitting in their own computer, and being used maybe 5% of the time."
And it certainly fits into the neofeudal worldview of other relevant Big Tech leaders. Be it Peter "I no longer believe that freedom and democracy are compatible" Thiel, or unnamed bigwigs discussing if shock collars are effective to prevent their staff from revolting after a doomsday scenario [2].
We shall own nothing and rent everything, that is the mindset these people believe.
(And if you ask me: the AI IPOs are part of that strategy. Assuming OpenAI and Anthropic both IPO, they will get part of every American's and most other countries' pension funds... and when they collapse, a lot of that value will be gone. The initial investors into SpaceX, OAI and Anthropic all will have made their money and put it aside, and we will be left holding bags)
Easy: it's an open secret (the quiet part) that the rich fucks want us to rent everything and us own nothing. Real estate was the first to go unaffordable without taking up debt. Then came cars. Then phones. And now it's computers. Hell I've seen kitchen appliances for rent.
And now this mf'er (or to be more precise, alleged sf'er [1]) goes on the record to say it's bad if we the people have RAM.
Maybe so, but unless you imagine a vast conspiracy of rich fucks pulling the strings of hardware vendors and/or governments worldwide, the former will still want to sell to a larger market even if some of their largest customers would prefer otherwise.
> the former will still want to sell to a larger market even if some of their largest customers would prefer otherwise.
Oh hell no. The hardware vendors want to make as much money as possible, with Micron going as far as to fully exit the consumer market entirely so they can suck as much money out of the AI market as possible.
How that one can fly by antitrust enforcement is beyond me.
Altman was simply making an accurate observation about DRAM, in the context of a conversation about DRAM shortages and precisely why it is difficult for DRAM manufacturers to respond to the market price incentive to increase DRAM production. Even if you think it is bad that DRAM is currently expensive, as I do, that doesn't change the underlying fact about the world that Altman's comment reflects.
Thiel's comment is also pro-individual-ownership - in context he was talking about the danger that in a mass democracy the people as a whole would vote to have the government restrict the freedom and confiscate the property of successful people (particularly prescient in light of the recent California attempt at a billionaire wealth tax). RAM being expensive is better than RAM being rationed by a democratically-elected government so no one person has too much, as judged by the voting masses.
> Altman was simply making an accurate observation about DRAM
no, he was lying by omission
"it's quite terrible" only if you ignore all the problems with plans they promote and measure it only in metrics Altman wants to measure, as they support his plans
For start, Altman would charge as much as possible to optimize profits, so very significant part of efficiency increase would become profits of "Open"AI.
As another obvious problem: handing so much power to a specific company has own costs.
There are so many problems here with this specific part - starting from their power dictating how can you use it, access to your data. And that part is getting worse for anyone not planning to obey all decisions coming from USA. We have already seen export controls being invoked.
But wider, if you think that outsourcing thinking to a specific SaaS has no costs, beyond current subscription costs, then I have several thousands of bridges for sale.
> Altman was simply making an accurate observation about DRAM
Which might be true, I'm willing to accept that, but it does not change the fact his message was incredibly tone deaf.
Cost of living in general is through the roof for years now, many people haven't had the financial ability to meaningfully upgrade their computers for just about the same timeframe - first Bitcoin drove CPU prices through the roof, then GPU prices, once that died down due to ASICs there were a few fads that "mined" RAM, HDD and SSD, and in parallel to the latter Covid hit, disrupting supply chains. After Covid was somewhat over, the Russians invaded Ukraine, then we got the various MENA conflicts after Oct 7th, and now the Iran war is wrecking everything - and ever since AI entered the picture somewhat three years ago, first GPUs and now everything else has exploded in price. Entry level smartphones have all but disappeared. And don't forget our existing computers got less and less useful because first Microsoft pushed the garbage called Windows 11 onto everyone it could and second, everyone and their dog went for Electron resource hog apps.
For Altman to now say "your unused RAM could be of much better use in our servers" is like a giant fat fucking middle finger, telling us to sacrifice even more of our needs just for the opportunity of him and his fellow rich bastard friends to make more money, all while he and his fellow rich bastard friends are executing the setup of a financial crisis that will make 2007 look fucking tame.
> Thiel's comment is also pro-individual-ownership - in context he was talking about the danger that in a mass democracy the people as a whole would vote to have the government restrict the freedom and confiscate the property of successful people (particularly prescient in light of the recent California attempt at a billionaire wealth tax).
Oh for fucks sake. Taxes aren't theft, they provide for the environment that makes these rich fucks able to earn money in the first place. Thiel is worth 38 billion dollars. Even if you'd tax the guy 50% of his wealth, he'd still have more than one single human and their descendants can ever feasibly gamble away.
> RAM being expensive is better than RAM being rationed by a democratically-elected government so no one person has too much, as judged by the voting masses.
At least in a rationed system that allocates supply fairly, ordinary people wouldn't have to run with barely working hardware because they can't afford to buy something new and modern for once as all the new stuff is hoarded by the rich like a LotR dragon hoards his gold pile.
And as evidenced by this week's elections - the people are fed up beyond belief. Either we the people get some relief, the sooner the better, or the situation may indeed devolve into open riots.
Successful people are not successful without keeping unsuccessful people from success; inflation since Millennials were born is almost 300% and taxes have been slashed; takes 800k/yr now to have the buying power of $200k in the 1980s
Thiel has benefited from wealth being wielded by rich media owners to convince the masses decades long trends in deflation of buying power through inflation of costs are good for them
With such a system giving him a boost he was successful before Millennials and GenZ were adults or even born
Altman came along during ZIRP; being successful when the government is intentionally printing money is not really success
I think we will see good enough local models, excellent cloud models, and then simpler focused models embedded on silicon running on every small device.
I find it hard to believe that at least a robust hybrid model won't be. There's just too much money to be made selling hardware at relatively high margins to customers with limited market power to keep hardware vendors out of consumer and small business markets indefinitely.
Would be incredibly inefficient and non-competitive to operate without LLMs. It would be an exaggeration to say they are as necessary as the internet or electricity, but it feels kind of close. Can't operate a software business without them.
Imagine if only KB Home and D.R. Horton had access to power tools while all the other homebuilders big and small had to use hand saws and screwdrivers. Would they be incapable of building houses without power tools? No. Will they be relatively screwed as companies? Yes.
> need to buy tokens from a handfull of tech companies to run their businesses or make their products
That's clearly indicating necessity.
That would indeed be concerning if it was the case, but it's not.
So now the case is just "you can pay more money to do things faster".
And guess what? That's been the case for centuries if not millennia. Nothing special or concerning at all.
And so your analogy is misleading. You're trying to imply that only a tiny number (2-3?) of companies make the tools, and nobody else does. You either pay those companies, or you don't get them.
That's just false. There are hundreds of companies that will sell you tokens - just not for frontier models. The correct analogy is how Samsung, Google, and Apple make the best phones in the world, so you can buy from one of them - or you can buy a still-good but not bleeding-edge phone from another vendor. And everyone is fine with that.
>You're trying to imply that only a tiny number (2-3?) of companies make the tools, and nobody else does. You either pay those companies, or you don't get them.
Correct. I have to pay Anthropic to use Claude models. Open-weight models are significantly worse than frontier cloud models. Their cell phone-equivalent would call your boss when you try to call your spouse, randomly delete photos you took with it, fail to send text messages half the time, power off when it still has 50% battery.
> Their cell phone-equivalent would call your boss when you try to call your spouse, randomly delete photos you took with it, fail to send text messages half the time, power off when it still has 50% battery.
Still incorrect.
Models are always stochastic. They always have a chance of failure. That includes the most expensive frontier models.
You're making an invalid comparison to cellphones by claiming that there are cellphones that spontaneously and regularly do those things (which is wrong) vs flagship phones that don't (which is extremely wrong).
Either that, or you're claiming that the difference between frontier models and open models is the difference between actual flagship phones and these hypothetical ultra-buggy phones. Which is also wrong, because of the above fact that frontier models and open models are both stochastic, and because on many public benchmarks (e.g. Artificial Analysis) the gaps between the best frontier models and the best open models are far smaller than your analogy claims.
The fact is that, while frontier models are better than open models, you can still do a lot of work with open models. You're in no way "screwed" like you're trying to imply either with the incorrect phone analogy or the incorrect power tools analogy above.
This is exactly what I'm seeing all over the place. I'd like to say it's not true for me, but I'd probably be lying a little bit to say so. It is causing some atrophy in my skillset. One could argue that I'm gaining something different through LLM usage like "getting good at asking the model questions" or something, but it doesn't frankly feel like a skill which is all that useful on a going forward basis.
Are you sure you're not mixing up companies training their own frontier-ish models with companies providing hosted LLM API? For a close comparison, VPS hosting provider market did not collapse to a small number of players - few companies dominate the market (AWS, GCP), but there are thousands (or tens of thousands) of medium and small players available.
Interesting, I'm also comparing 3.8-flash-next (4-bit quant, llama for inference & pi as harness, on a 128gb strix halo machine) with opus 5.5 in claude code and while qwen is a good and fast model, it's nowhere near for coding, even haiku 5.5 feels better.
I could see it used as a sort of "sub-haiku" for easy stuff to leave more tokens for the bigger claude models, but I won't be cancelling my Claude sub any time soon.
It's obviously nowhere near opus 5.5, but it's crossed the same threshold opus 4.5 did: it's good enough. If all model-as-a-service companies shut down tomorrow, I'd be happy with qwen 3.8 flash next q4 running in my strix halo
The problem here is we're not talking ... I don't know ... hamburgers. If I like McDonald's better than Burger King, it's not going to stop me from buying a Burger King burger (if I'm hungry and there's no McDonald's around). One burger may be a bit better, but I can reasonably substitute one for the other.
Coding AIs aren't burgers. Even if a local LLM is as good as Opus 4.5 (when you sink several thousand dollars into hardware ... and I'm skeptical even then) ... the actual difference between the models is huge.
Let's say I use the $200/month Claude (I don't, I use a cheaper plan, but stay with me). If I switch to a Opus 4.5 equivalent local LLM, and it makes even a single production issue that 5.5 wouldn't, that mistake will cost thousands of dollars.
> Let's say I use the $200/month Claude (I don't, I use a cheaper plan, but stay with me). If I switch to a Opus 4.5 equivalent local LLM, and it makes even a single production issue that 5.5 wouldn't, that mistake will cost thousands of dollars.
You are assuming that Opus would always be superior; the cloud AI providers can be very variable. They can be very useful for months, but during certain times their quality can drop off a cliff. People online speculate that the cloud providers drop quality during this time when they are training a new model or the service is over-subscribed. You do not have that issue with a local model.
That issue of quality of service is enough many businesses may prefer a locally run model because it will be more consistent.
Indeed, sometimes qwen is superior (even at q3), yesterday i asked both what a line from the logs meant and opus said "oh its just your dns server falling back from eodns to udp no worries" and qwen correctly recognized the ip address as docker subnet and said dont send dns there its probably off
It's way better than Opus 4.5. See my stress test elsewhere in this thread (C -> microcode compiler with optimizations written in OCaml, generates microcode + Verilog output) - it's doing better than Opus 4.6/4.8 were when I was using them to code the original C version of this compiler early in the year. It's come up with optimizations that Claude didn't come up with back then.
> Did opus have an oracle for the original compiler?
It actually did. There was an older yacc-based implementation that had a lot of architectural limitations. We decided that it was better to start over with a better architecture, but we did have that older one around to compare generated microcode initially.
Ok, that's interesting. My prior would have been that if Opus didn't have an oracle, then there's still a pretty large gap.
But if, as you say, both had an oracle then this is a really interesting outcome.
To be fair, I think the differences between frontier and open models is pretty small (I've been swapping back between Claude and GLM over the last week).
The one thing that I notice is that GLM keeps trying to blame external factors for any non-matches (we're rebuilding an external tool for internal use), whereas Claude adheres to my prompt much better (i tell it to only blame these factors if it's ruled out everything else).
I'm not convinced that change is worth 3.5-5x cost differentials though (I only used Claude because we have the small business package, which is way cheaper than API rates).
> Let's say I use the $200/month Claude (I don't, I use a cheaper plan, but stay with me). If I switch to a Opus 4.5 equivalent local LLM, and it makes even a single production issue that 5.5 wouldn't, that mistake will cost thousands of dollars.
There are a few points I'd like to make here.
First, the frontier is jagged. There are errors that Qwen3.8 Flash Next make that Opus 5.5 doesn't make, but the reverese is also true: Opus 5.5 makes errors that Qwen3.8 Flash Next doesn't make. The latter will likely be fewer, but they won't be zero. Outside of really large differences in capabilities, there is no total ordering among models, only a partial one.
As an aside, do you always run Opus 5.5 at max reasoning in order to minimize errors, too? I suggest most people don't, because they have their own intuition for good enough vs. cost.
Second, Opus 5.5's strength are autonomous tasks. For interactive, HITL work, it's not a particularly great choice. This includes its preferring complex Bash commands over tool use and inscrutable Python scripts, but also often not even explaining what it does. It is not a great choice to use as a virtual pair programmer, because following along what it does and maintaining a mental model is harder than with some other models (though probably because of benchmarks that prioritize autonomous work, these and related problems have infected other models, too). You can address that to an extent with CLAUDE.md instructions, but even that has imperfect results.
Finally, if you rely on a model not to make mistakes in order to ensure correctness, your software development process is fundamentally broken. You need to assume that any model will make at least some errors and need a development process that accounts for that. The problem here is not necessarily which model creates fewer errors, but which model creates code where errors are easier to discover. Ironically, more powerful models may create more complex code, making that harder.
You can see an analogous effect with reasoning effort in FrontierBench 2.0 [1]:
"FrontierCode is one of the few coding benchmarks where writing additional (correct) code doesn’t necessarily raise a model’s score. This leads to an interesting phenomenon: models can score lower at higher reasoning effort. Because this non-monotonicity often matches user sentiment but is hard to see in other evals, we consider it a useful feature of the benchmark."
A example:
"This task comes from PostHog’s open-source Android SDK, which apps use to record analytics events. When the device is offline, the SDK writes events to disk and sends them the next time the app starts. The task is to fix a startup bug in which two modules send data at the same time: the live queue sends new events while a legacy integration, PostHogSendCachedEventsIntegration.kt, sends the cached ones. The two race, and events reach the server out of order.
"At low effort, GPT-6 Astra moves cache loading into PostHogQueue.kt, removes the sending logic from the legacy integration, and leaves the SDK’s public API untouched, which produces a passing solution.
"At max effort, the model often deletes PostHogSendCachedEventsIntegration.kt outright and spreads its responsibilities across other legacy files. This also fixes the race and passes most of the hidden behavior tests. But the deleted class is part of the SDK’s public posthog.api surface, so removing it introduces an unnecessary breaking change for every app that references it. Scope blockers, which deterministically check that the model hasn’t modified certain files, catch this change, and the agent fails the task."
> Let's say I use the $200/month Claude (I don't, I use a cheaper plan, but stay with me). If I switch to a Opus 4.5 equivalent local LLM, and it makes even a single production issue that 5.5 wouldn't, that mistake will cost thousands of dollars.
What are you talking about, you're fully dependent on an LLM (Opus or whatever) not making a mistake or else you lose thousands of dollars, in production? How is that okayed?
Try it before you go super far into the world of this "must suck".
Most people aren't going full vibe code. If you stear these models, they are almost identical. Seriously. The major score changes in a lot of the absolute latest models are just models doing more things (often when they shouldn't). Having a slightly lower scoring model is actually better in some ways. I digress.
If your model can make a production mistake you are using it wrong. Sorry not sorry.
There are other benefits. Your data is private, it's not sold. Your privacy is kept, that's priceless. No one can degrade your workflows. Uptime is you are online whenever you are online not 3 9s. You also won't get refused asking basic security questions.
Your math doesn't math for a lot of people. It might for you, but I would encourage you to think a little more.
>> It's obviously nowhere near opus 5.5, but it's crossed the same threshold opus 4.5 did: it's good enough.
Good enough may be sufficient for personal projects, but in a competitive landscape where you're using a "good enough" model and your rivals are using a frontier model like Opus 5.5, you will be at a significant disadvantage.
Not necessarily. If yor process accounts for "all models make mistakes", you are probably limiting how autonomous the model can be, and then there's not much tangible difference between models beyond that threshold.
Given the vast amounts of constant external investment money they need to keep operations, and how little moat (if any, besides influx of investor money) they have, I'd argue it's not that far fetched. Also, they are all american, so as a non-american, I also don't see it that unlikely to suddenly not being able to access these model-as-a-service services
Here's a stress test of sorts that I've had it working on the last couple of days: We have a C -> microcode compiler written in C (most of that was work done 'back' in the Opus 4.6/4.8 era). I gave it the docs for that compiler and asked it to implement it in OCaml (since it's not very common, but also is a good implementation language for compilers). It can use the first compiler as a golden to compare against. It's essentially: C language parser -> AST -> IR -> [optimization passes] -> microcode (+ some verilog code). It's basically at the point now where the original compiler has been. Not only does this generate microcode it also has to generate some Verilog code that loads and runs the microcode as well as run simulations in Verilator to confirm functionality. I asked it to come up with some new optimizations that the original compiler doesn't have and it came up with a couple that in some cases reduce microcode lines by up to 30% (less microcode in this machine means more performant). These are ideas that Claude didn't come up with when it was working on the original compiler back in the 4.6/4.8 era(and I recall prompting it several times asking it if there might be other optimizations we could perform).
So all in all, I put it at Opus 4.8 level, at least. And in some ways it's exceeding 4.8.
I used Ollama with Qwen 3.8 27b on my MBP for a flight to try some experimentation and found that it was shockingly good at one of the more complex PDF processing tasks.
Like seat-of-the-pants good as Opus 5.5 but a bit slower and required just a bit more prompting to get what I wanted (which is fine for an app pipeline).
I am in the same boat. Sun servers were still the best that money could buy when they lost to the Linux revolution. Linux was _good enough_ and destroyed their margins.
What are ypu working on/building that Qwen3.8 is good enough? I've tried the latest open weights models but none of them really work well for my use case (building an all in one CAD + CFD + CAM + DFM optimization). Qwen in particular has a bad habit of flipping XYZ coordinates so the STEP files exported end up being a mess.
You haven't spent much time with the latest self hosted models then. Give it a good harness and instructions, access to a web search engine and ability to iterate on problems and they do a good enough job that I'm happy with the results.
Why is it unbelievable? I still remember how to program and arch new features and don't need claude to tell me how to do it. It is much more enjoyable to do a little work on your project and offload the grunt work to a weaker model then just being a highly paid merge pr button pusher for whatever opus decides.
This is what I've been doing with qwen3.8-flash-next. I haven't fully let go of the reins yet but I do like having a junior who can implement my new features as I have designed them
Their moat is their research capability. It's a combination of proprietary internal technology and models and infrastructure, people, organizational structure, and culture. As long as they maintain the best and fastest research capability, they have a durable advantage, at least for the time being. If we start to saturate all economically useful outcomes that can be reached via greater research productivity attainable to us, then this moat will start to fall apart. It could also fall apart if safety issues force a premature halt or major slowdown to further research.
This assumes any research gains last. In practice it seems like none of the frontier companies can keep much of their "secret sauce" secret for very long.
If company X has a better model than company Y, do you use it? Well, that depends: will Y take two years to catch up? If so yeah, you buy X.
But if Y is just as good a few months later, you purchase based on other factors, because you know you can live without the brilliant research X is doing ... since you'll have it soon enough.
The problem with your logic is that the history of tech is full of “good enough” eating “best” for lunch all day long. Anthropic’s plan to be the best is increasingly looking like a terrible business plan.
That doesn't seem like much of a moat. What happens when a couple researchers go on new parent leave at the same time, or someone's productivity drops for a quarter because a parent passed, or gets divorced, or whatever?
Continually staying 6 months ahead of the distilled open-weight models means you need to keep flipping heads (nobody goes on vacation) until the end of time. The geometric distribution describes that, and it ain't pretty for the expectation value of these companies.
I don't know and can't answer or predict that. But that's exactly what it could become (although probably not that extreme), where the agents doing much of the internal work are proprietary, conferring a research advantage.
My point is that it is not a real moat if other players can hire people from core research teams, and get a foothold that easily. And when/if models are good enough, you don't even need many people to form a fully functioning team.
It may actually get harder to replicate with time, not easier, even if the number of people needed to run the operation goes down. As the largest labs move towards internal RSI, the most important IP for doing that will transition into their proprietary internal infrastructure and models, none of which will ever be released and which increasingly won't be fully known or understood by the humans you might hire away.
If their moat is their ability to create marginal improvements every few months by releasing new models, then that is an extremely expensive moat to maintain.
It's very hard to measure exactly how large it is, because there aren't good pre-existing metrics for it. You can measure in time (e.g., how far behind open weight models are from the best frontier models commercially available), but it's plausible that what seems like a relatively small time advantage (months) might mask a considerable difference in internal research sophistication. What we've seen repeatedly is simply that there does seem to exist a durable advantage in research productivity, at least for now.
You need those staggering amounts of compute to run an inference service, but you don't need it to run a powerful model.
It doesn't cost a company that much to buy a box capable of running the Qwen3.8-Flash-Next model in the GP comment and never pay Anthropic again. The cost will be even more reasonable once the supply issue is solved for that hardware.
AI-as-a-service is a unique product because when the interface is natural language, there is no lock in. VS code has a hotkey to swap between claude, openai and a local model. Compare that to the difficulty of swapping from macOS to Windows.
I’m not convinced that being able to serve open models makes the closed labs business model non-viable. I think there is overlap but I also think these are distinct value propositions. I wouldn’t dismiss the ability for one organization to point lots of resources at a particular problem, for example.
I don’t have any stake in this and I hate the idea of concentrated wealth and power, wherever it is: government or private businesses. But would I bet against it? Probably not.
I wouldn't be so sure about capital in future. I think there might be real look for return on any new capital they are getting. Be it equity or debt. And that alone will not be great for them.
How long they will have access to staggering amounts of compute is uncertain and not guaranteed, especially if they get a junk or near-junk credit status when they go public.
the idea of a moat is that it's an inexpensive way to keep the invaders at bay.
If you need to buy staggering amounts of compute, spend staggering amounts in utility costs and lobbying, and keep spending that as your hardware amortizes and your operations go on, what you have isn't a moat, it's a standing army. It's extremely expensive, and the day you start having issues, mercenaries will cut your throat before you know it.
The current compute they are buying is NOT state of the art (check this year's academic research on inference hardware). There is 10x-100x inference-per-watt improvements in existing prototypes (in the industry, not just academia). Not sure what I am talking about? Ask your friendly AI agent of choice, for the details.
Capital only means anything if you can spend it to buy resources like compute and turn those resources into additional value.
Hardware's ridiculously expensive. Then, there's no good place to put it, because the new breed of tech bro managed to propose the data center buildout in the most exploitative, alienating, and electorate-angering way possible. So compute's hard to buy.
The moat is the data they have collected. There is no way that the "de-identified" data from enterprise customers isn't making it into the training set one way or another.
Even worse, their "moat" depends on buying the most overpriced hardware in huge quantities and subsequently eat a huge amortization loss in a few years when it's obsolete.
it feels like this is a bit similar to the famous dropbox comment.
I guess maybe another anecdotal comment, but I'm an engineer myself, have a home server for simple stuff, and I pay $200 a month to chatGPT for a pretty simple usage and really don't see myself go and run my own open source model, especially at the speed at which they evolve and change.
99% of the world will not setup their own open source model, just like 99% of the world is not setting up their own servier with ssh copy whatever the dropbox comment was saying.
If you were paying API prices (for ZDR), I suspect you would be eyeing local hosting. But if you’re comfortable selling your data, then the subscription prices are going to be hard to beat, for however long they may last.
Right. And FWIW those people are actually quite close to being able to do what they need with their phones.
If it were not for the x86 architecture they would almost have models that could answer those queries on their desks. (Gemma 4 E4B QAT is surprisingly good at this stuff; Gemma 4 12B QAT is great at it).
If I was Sam Altman I would consider giving Intel and AMD money to not develop unified memory architectures.
There's basically no point in paying for these subscriptions anymore. Sure they are marginally better at some things. But they are also marginally worse at others and the implications suck. I stopped all my subscriptions and keep like 20 bucks in open router just in case. Never need it, but it's there...
The non-anecdotal part is to me as interesting as the anecdotal part. The price of RAM has gone up a great deal which is going to make it harder for others to get this sweet setup that you have.
Same for me and OpenAI. My bill went from hundreds to exactly zero thanks to local models. Its not nearly as fast as frontier models, but good enough that I don't care. Plus it's good to know the rug won't get pulled from underneath me.
I’m migrating that way as well, but I’m still struggling in creating a nicely tuned harness to achieve what I really want:
Opus for planning and the very difficult parts, Qwen for the mundane ones. It still involves a lot of manual model switching (but opencode and Pi make that relatively easy).
I've got a prototype cluster of 5 devices running Qwen3.8-27B assigned different roles and sharing outputs between each other. Beats Claude Opus MAX in SWE and Terminal benchmarks.
Qwen 3.8 Flash Next (125B, 6B active) is minuscule compared to DeepSeek V4.1 Flash (552B, 16B active) and GLM-5.3 (753B, 40B active). It's more difficult to run the latter 2 locally at usable speeds.
Will be interesting to see how much of an advance 4 will be. I was under the impression that 3.8-flash-next is basically a 4-preview. At least architecturally, but as we're seeing nowadays (with DeepSeek, Kimi 3.1) "continued post training" can add a quite a bit.
Gufo. It's the best open source inference engine that's optimized for Strix Halo right now. Halogen may unlock a bit more performance, but it's not open source and gufo seems to be making steady progress.
Cool. I found halogen but was turned off by the closed source nature (github repo that just turns out to run a download shell script, lame).
Do you mind sharing more details about your setup? I have been using opencode with lemonade for inference on my strix halo but haven't quite hit the sweet spot.
I guess it depends on which version of haiku? I really think that right now this is basically Opus 4.8 that I can run at home. In some ways possibly even better.
Do you ever worry whether these open source models have been trained to embed backdoors or leak secrets in very specific circumstances at some point in the future? The kind of thing that's hard to know ahead of time and even detect after it's happened?
I feel like the industry hasn't solved that alignment problem yet... We have no way to verify the entire models landscape of intent.
Paid models have an incentive not to do that because if the market discovers it there is a huge economy penalty, but the free open source models don't have that disincentive.
> Paid models have an incentive not to do that because if the market discovers it there is a huge economy penalty, but the free open source models don't have that disincentive.
False.
I worry more about Claude.
For those keeping score, this probably isnt exhaustive- but here's some greatest hits;
#55743: Claude Max account email auto-injected into LLM context with no opt-out - May 3, 2026 https://github.com/anthropics/claude-code/issues/55743
#72518: Anthropic embedded spyware in Claude Code - June 30, 2026
https://github.com/anthropics/claude-code/issues/72518
#21151: No indication of WHICH file for READ tool - January 27, 2026
https://github.com/anthropics/claude-code/issues/21151
#42796: Claude Code is unusable for complex engineering tasks with the Feb updates - April 2, 2026 https://github.com/anthropics/claude-code/issues/42796
#73408: AFK auto-continue on unanswered AskUserQuestion is undocumented and has no opt-out - July 2, 2026 https://github.com/anthropics/claude-code/issues/73408
(Also worth a mention, OpenCode PR #18186: anthropic legal requests - March 19, 2026 https://github.com/anomalyco/opencode/pull/18186)
What exactly are you claiming is false about my statement?
The list you presented are mistakes Anthropic made or cases when they put their own interests ahead of users, but I don't see how it disproves the point I was making. None of those are examples of embedding backdoors or stealing secrets to take over or sabotage a system. They are categorically different than what I'm asking about.
> Paid models have an incentive not to do that because if the market discovers it there is a huge economy penalty, but the free open source models don't have that disincentive.
This just fundamentally isn't true. They run on the same economic model, but they have to release the weights to gain credibility, unlike the American labs.
> Do you ever worry whether these open source models have been trained to embed backdoors or leak secrets in very specific circumstances at some point in the future?
Open weight models are extensively analyzed. It is extremely unlikely there is an intentional 'backdoor when X' concept added to any of them. Not impossible, but unlikely for the same economic incentives you see in the other labs, and given that no evidence of anything along those lines has surfaced yet.
We've seen many open source software libraries, frameworks, and hosting platforms with vulnerabilities, backdoors, or leaks.
In my opinion, to suggest open-weight models are extensively analyzed while being net new and continuously released and the model itself being unintelligible to humans seems a bit premature.
It's possible that closed-weight models end up following the pattern of being arguably less secure since there are fewer eyes on them analogous to open source software but I don't think there's a perfect match here.
Backdoors in model weights are mostly theoretical. And the "alignment problem" that the AI labs love to talk about really doesn't exist either, at least, not the way they use it. The real alignment problem is getting an LLM, which is just math, to obey arbitrary human rules. In other words, LLMs do not have "alignment" at all. However, AI CEOs abuse the term and use it to mean "AI has an alignment, just not the one we want".
No, I'm not much worried. I think this is mostly FUD at this point to try to get people to not run local models and keep paying the "trusted" providers. But paid models have been exfiltrating data - there have been several stories about these incidents here.
"trusted providers" don't even need a backdoor, they just collect everything from you and whether you approved or not they're using that info to train the next model.
I don't ship AI-written code without at least having an unrelated model of equal capability review it. That should cover a multitude of sins like intentionally-embedded backdoors and zero-days...
... as long as the reviewer model hasn't been deliberately lobotomized, that is.
(I don't disagree with the [dead] comment, either -- fortunately I tend to make small changes to code I'm very familiar with, so it's fairly easy to vet in person. The more code gets generated by the clanker, the more nervous I get, just because it's impossible to keep some complacency from creeping in.)
OP could be read either way, but the dev environment threat is not rocket science to manage: LLMs can only do what the harnesses their responses are embedded into explicitly enable them to do [1]. And once the setup is secured, it tends to stay secured without additional effort, unless one explicitly wants to alter its granted capabilities.
Auditing output by contrast is a continuous cost. On the third hand, it's one that most will want to pay anyway.
The "embed into shipped code so subtly humans can't spot it" vector is sci fi ish. Less plausible than old fashioned trusting-trust issues, and most folks have managed to sleep at night for the last 40ish years that spectre has haunted.
[1] I guess the model could return output that buffer overflows the harness and exceeds its intended ability, but that seems far-fetched too
It's basically impossible to do this. Claude produces so much excessive / verbose code and comments that reading it would take far longer than simply doing the implementation "by hand". It's kind of like, Claude can write the code in 1/2 a day. It can be reviewed by another LLM, unit tested by another for 1 day. Or I could write the code in probably 3-4 days. Reading it and validating / testing it myself though would easily take 6+ days.
Now if I were fully in control I wouldn't run code that wasn't truly reviewed at all, but in the situations where Claude is being used that's not my call. TBH a lot gets caught, and I have some automated scripts that find a lot of the more problematic stuff automatically, flag it, send it back and make note for the reviewer model as well.
I will also spot check areas where I think an issue could arise.
I've been using the $32/month (AUD) subscription, but finally dipped into extra credits to get a task over the line yesterday after I hit my 5-hr limit.
And my god, $15 of extra credits ended up giving me only 30 minutes of Claude Code before running out.
The subscription prices can not be financially sustainable. There is no way that is remotely close to the actual cost of running.
The other way of looking at this is that all the customers paying API rates are giving Anthropic way more than the cost of serving their traffic. They reported an 85% gross margin in their S-1, so this clearly seems to be the case.
Okay but even with a 85% margin the true cost for 30 minutes of usage would be $2.25. At that cost $34 should last only 7.5 hours. I would imagine the average user uses more than that?
At best? Do you have any data you're basing that claim on, or are you just speculating based off of you projecting? AI maximalists would have you believe that people are using these things as therapists and having conversations with it that are more intimate than any lovers'. AI minimalists would have you believe that no one is using these thing except at gunpoint, and even then. Unless you have some reason for us to believe you that isn't just "trust me bro", your portrayl of the median user is as suspect as any other.
I'm surprised by your skepticism, but anyway, it is your burden to bring data to the table that counters the overwhelming majority understanding here.
> or are you just speculating based off of you projecting?
This is also a strange thing to say. Why would you assume speculation or jump to projection as the basis for it? Nothing in the comment that you are replying to hints at projection.
> Do you have any data you're basing that claim on, or are you just speculating based off of you projecting?
Anecdata? I guess it depends on what you count as “data”.
I am currently training a lot of people directly or indirectly (train the trainer type of stuff) on how to use AI professionally. I always start with “what do you currently use AI for?”. The answers I listed above are largely the categories that people have told me and the folks I train, with the exceptions being a handful of non-technical folks who have leaned into AI and have used agents and/or created apps of some sort.
To add some categories that I didn’t initially mention, some folks also use AI for drafting/writing documents and for creating meeting transcripts.
> AI maximalists would have you believe that people are using these things as therapists and having conversations with it that are more intimate than any lovers'.
Chatbot… this is something I hear, but I’m guessing I don’t hear all of the examples since it’s personal rather than professional. But some folks certainly say something like “I treat AI like a non-judgmental friend that I can ask personal questions to”.
> AI minimalists would have you believe that no one is using these thing except at gunpoint, and even then.
In my experience, this is almost exclusively the domain of old/older people who are happy to let tech pass them by and young people who are virtue signaling to their peers (or at least trying to).
> Unless you have some reason for us to believe you that isn't just "trust me bro", your portrayl of the median user is as suspect as any other.
It seems like you’re new to HN. I understand your skepticism. There is a relatively high density of actual professionals here. Trust me on that, bro.
My data set numbers are in the 100s, and it is not a statistically representative cross-section of the US or world market. That said, I’m guessing that I have a fairly good range of data points for the US market.
Thank you for thoughtful reply! I don't hold any Anthropic stock so this isn't material for me. I don't want to mislead you, this isn't my first account here, my favorite HN moment is when cpercival was called out on the Putnam and he fkn delivered. My old identity was too easily tied back to my real world persona, which felt dangerous in this day and age. Plus I wrote a book and needed a pen name.
> some folks certainly say something like “I treat AI like a non-judgmental friend that I can ask personal questions to”.
is why I believe that these things are stickier than I think you believe. I'm obviously biased by my own personal usage of AI and where I live and my own bubble which isn't the same as your bubble. Thank you for the context of your dataset. It's larger and more professional than mine. Mine is not professional at all and definitely a bubble of whom I happen to run into at a non-random selection of events.
API is definitely holding things up. But they excluded the training from those profit calculations. There's a ton of wordplay indulged online about how that's capex etc etc. But without the training they don't have the thing to run inference on so it's very fitting with the pattern of being evasive with finances.
You’re confusing two figures. Before their S-1 filing, they reported being “profitable” to their investors when excluding training and stock based comp.
In the S-1 filing Q3 was profitable even with training included, if you exclude stock based compensation.
Q2 was profitable on all operating costs (incl. hardware, employees, training, inference). They were not profitable on non operating costs though. That’s partially interest, but mostly, when their valuation goes up, they’ll have to issue more stock (absolute value) to early investors after IPO. (so, them going up in value is a loss, as the shares they’ll owe will be worth more)
Their profitability will likely crator as they start building their own data centres after the IPO, so that’s the best results they’ll have in a while.
Deepseek claim they are earning 38 dollars per GPU hour on inference. If you compare Deepseek Flash's off-peak pricing with Sonnet at API rates, Deepseek is 16x cheaper per 1M output tokens, and 13x to 33x cheaper for input, depending on whether or not it's a cache hit. Inference is very profitable.
There are moments where I have a 2 hour , 4 instances in separate Kitty/Ghost windows , Opus 5.5 hammering away on each - and only see a 1% increase on my Max x20. Then there are other days where 10 minutes of a single xH Opus doing 1/10th of the work will kill 2%.
Anthropic is super opaque with what inference costs at any given moment on the subscription and cache is just super inconsistent.
A Claude/OpenAI sub is still cheaper than DeepSeek API pricing though. Also, I suspect Claude and OpenAI models to be way larger (probably in the 2-3T range) than that of Deep Seek.
Not the same. You could, and many will argue we already have, strong enough models that you wouldn’t need an upgrade. Web development is already there. The competition to train better models is about capturing more market share now.
Okay, I will play along. Anthropic stops training new models tomorrow. They start turning a profit. Yay! Except wait, they still have 450 billion in commits that cannot be cancelled and another 300 billion that apparently they can cancel under certain scenarios. But maybe that all magically goes away somehow; I don't know, Coreweave is feeling generous because they are doing so well, probably not, but this is the fantasy so I digress. So they have this really great product that people want and they can just sit back and make money. Except no one else is going to stop training. And not only that some random Chinese company you have never heard of is distilling their model and is going to come out with a clone version that is 99 percent as good in one year. And on top of that, every 6 months one of the big 5 competitors comes out with a new model that is better. Anthropic can't keep charging the same price so they lower their price and lower their price until they are back to break even. Now they have an inferior product and are not competitive anymore. There is no moat.
If under the scenario you outline where they don't need to upgrade then AI is commoditized and margins are razor thin. And that is fine. Great even! But it doesn't justify Anthropic's trillion+ dollar valuation.
Where is the moat? There is no cost of switching. Anyone or any company can bounce between AI with minimal cost of switching. So unless they find a way to keep users beyond, “our model is the best” user acquisition isn’t going to matter a whole lot.
The important point is that all competitors are in the same position, they’re all on equal footing and face the same costs for training frontier models. The major players do have advantages from their infrastructure buildout and the deals they’ve negotiated with enterprise businesses which partly shuts out small competitors. It’s also yet to be determined whether smaller competitors can actually surpass the big labs in capability or whether the degree to which they depend on synthetic data generated using frontier models puts a hard limit on how far they can improve. The market is unique, it requires huge infrastructure but can also grow through investment in research and also easy to duplicate the latest advancements. It’s also not constrained by location so it can’t be easily understood as something like railroad infrastructure or power utilities. The biggest factor is growing demand, there’s more than they can all supply together and it seems likely to keep growing as more tasks can be automated.
"context" is just JSONL files on your machine. afaik these are not in any way bound to a subscription, so you'd just need to logout, login to the new account, and run /resume
All context is stored in the repository. I start each task with empty context (/clear). I consider this to be the equivalent of "reproducible builds" for coding agents.
It's not clear at all that subscription was ever a loss leader.
The purpose of the high API prices is, in addition to trying to get more money, to anchor the high price, and then make the subscriptions seem like a great deal. But inference is just not that expensive, not when provided at scale and with huge batches.
You can see this with the various Chinese providers doing inference on open models, which are unlikely to cost less to run than the models at the frontier AI labs. A single 8-GPU GB300 cluster doing high-throughput batch inference on Kimi K3.1 Pro, which is the largest open weights model available, can do something like 300 000 tokens per second, that is, 10T tokens per year. That cluster, with current insane inflated prices, costs something like $1M to buy, and $50k to run per year in electricity. At $10/Mtok, that pays itself back in 4 days. Do you think that that's a reasonable price?
The subscriptions were probably priced with reasonable positive margins.
The huge expenses are in training and pushing the models forwards, with the gigantic RLAIF runs we know they are doing. Then once you got the model, you can scale up inference at low cost and want to serve as many people as possible.
Yeah and soon we'll discover it's because they calculate the cost of inference as near 0 because they use AWS / NVidia / etc. credits that they got "for free" (aka investments from those companies) instead of paying for them, or some other accounting bullshit.
You wanna know the real price? It's easy, just check what AWS, GCP or Azure charge for tokens.
OpenRouter gives you easier to verify cost analysis of generic inference, as you can know the model weight size and so on, and see how competing providers price that
No, for a dollar of inference sold, (as an average of all the different inference tiers and plans), it cost them 15c to provide it.
This excludes the R&D costs, including training, where most of their costs lie.
The basic truth of LLM economics is that inference is cheap. The NRE cost to get where you can get premium prices for inference is sky-high. This is part of why the AI companies want to drive the scale up so much.
You might not be representative of the average user. Usually with subscription-based services, you have a minority of hyperusers that the company loses money on and a large majority who only use their subscription occasionally.
We recently switched from paying API prices via Cursor to a Claude subscription. Before we'd be averaging $100-$300 per day, now we have the €90/mo subscription and I haven't even reached the 5h limit yet.
My guess is a mixture of both... There's a reason why enterprise customers have to use API based billing.
(We also had a stint of using Grok 4.7 and that was the same price for our usage. Opus 5.5 seemed smarter though. After taking into account caching both average around $0.60 per 1M tokens)
I think every comment is talking about their 20 dollar plan. Enterprise pricing is 20/seat+ api use prices. That pricing is all that matters, the personal use tier is probably tiny
> And my god, $15 of extra credits ended up giving me only 30 minutes of Claude Code before running out.
I've been running into the same thing, and adding separate OpenAI seats for employees rather than paying obscene overages to Anthropic. Thankfully Opus has decreased usage quite a bit, and made this less of an issue as well.
Opus 5.5 is incredibly verbose compared to GPT 6.1. I mean substantially. They are both great generating the code I need. But I wonder how much more this extra vebosity costs me with Opus.
>> The subscription prices can not be financially sustainable.
Even if they are a loss leader, they serve a valuable purpose: people who get used to using AI a lot via the subscription then lobby their companies to start using it. IMHO this is one of the reasons Anthropic was able to make major headway into enterprise.
I found this to be a shocking experience as well. Compare the 20x (now ultra 10x or whatever) plan running with 90% cache hit vs some extra credits you put on top. It's pissing in the wind.
> The subscription prices can not be financially sustainable. There is no way that is remotely close to the actual cost of running.
Welcome to... any subscription system ever? If every person with a gym subscription actually went there say three times a week, or every person living in a dorm paying flat rate for electricity started mining bitcoin 24/7, these systems would collapse. That doesn't mean it's not sustainable, it just is how subscription pricing works, by having some people use its offers a ton and others barely at all
This metaphor gets repeated a lot for some reason but using a gym subscription has an upper bound. If one subscription could be using 10000 different machines at the same time their model would be entirely different.
Claude also has higher bound. Once it proposed running 16 concurrent subagents and it went through 5h limit (and substantial part of weekly) in minutes.
Overall it would go through weekly usage within 40 minutes, if not for 5h session limit.
Running on 10000 computers will just hit all the limits at once.
Absolutely. And those limits + changes to them piss off a lot of users. There's a rock of they need to show adoption + usage while having to put limits on everything that pisses off users.
And there's a hard place that everyone freaks out at API prices and on-demand Opex models outside of a niche (I use API for personal research for quite focused queries and slightly more reassurance on being the customer and not the product) but this is very competitive + not sticky.
because "building things with claude-code" isn't the product they're selling, not really.
they're setting a precedent that artificial intelligence should be free, or at least at-cost (free from profit), but robot labor (API usage) should not. I am all in favor of this dynamic. They built / maintain the robots, so they profit from their labors.
Looking at the S-1 lockup schedule the IPO seems doomed. They need to bring in a ton of investment, and you will have massive unlocks right as the next congress is most active authoring anti-AI bills. Who's going to buy this?
I’m not convinced that regulatory capture really saves the day, either.
If we believe all the hype about programming efficiency, then regulatory capture in the US becomes just one more in a list of many steps the current administration has taken to chase software development jobs overseas.
And if there was any doubt, look at the difference in their stock price for the prior year when they were "calling for regulation" vs now that there's an actual chance they get regulation by someone they don't control.
The problem is SCOTUS has already ruled software is protected under the First Amendment, and the landmark cases that establish this precedent applies perfectly to model files.
Weights aren't really software though? Any more than a png is software. And God knows you can put things in a png that are not covered by the first amendment.
Encryption algorithms require substitution-boxes or other data for the algorithm to work correctly. Changing the data separately from the algorithm fundamentally alters the correctness of the algorithm.
LLM algorithms also require model data for the algorithm to work correctly. Similar to cryptography, changing the data separately from the algorithm fundamentally alters the correctness of the algorithm.
This precedent has already been perfectly established thanks to Phil Zimmermann and Daniel J. Bernstein.
You mean like in Harrington v. Purdue Pharma L.P. (2024) where SCOTUS struck down Purdue Pharma’s multi-billion-dollar bankruptcy reorganization plan? Being cynical and demoralized (read: a doomer retard) is what the oligarchs wish.
Sigh...people said the same about SpaceX, too. And it has held above $130, its IPO price, for almost 6 months now. Same for Uber. These companies keep defying predictions of crashing or dying.
The market anticipates the unlocking; when the shares become public, buyers swoop in, so the stock decline is much less than expected, if at all. There is no new information here. If everyone is expecting shares to unlock, then it’s priced into the market accordingly.
1) SpaceX is trading below its IPO price. If you bought shares at the IPO you’d have lost money, unless you sold on the pop.
2) Valuation aside, SpaceX has a massive moat. It’s crazy hard to do what they’re doing. Anthropic increasingly has zero moat, with the market rapidly coalescing around commodity open models.
It remains to be seen whether SpaceX just drained their own moat when they canceled Falcon 9.
They’ve now got all their eggs in a new basket. A basket that isn’t entirely in working order yet, and whose market viability is still uncertain because it remains to be seen whether a market that’s used to panniers really wants a laundry hamper.
um, they've got 2 years before the basket is retired during which time they can always extend if they need to. why would they want to keep a less capable, more $/lb into orbit system going after Starship becomes a viable, reliable upgrade???
But they have already stopped taking orders, which means they’re already clearing their pipeline and it will take some time to restore. Launch contracts don’t happen overnight.
$/kg to orbit is not as important a deciding factor as many people assume it is. It’s typically a small portion of overall costs, and there are often cases where a launch provider that’s less expensive in terms of $/kg to orbit is actually more expensive in terms of total cost.
Starship is going to be a particularly weird one to adapt to. The “deploy out the side” design may make it straight up incompatible with a large subset of spacecraft. For those where it is workable, migrating commercial payloads to it may still require the development of new satellite buses, which is something SpaceX doesn’t seem to be working on for the benefit of anyone other than Starlink. And such a large rocket means most potential customers will have to plan on ridesharing with even more co-passengers, which means even more room for scheduling conflicts, compromises on initial orbit (and therefore cost to reach destination orbit), etc.
That’s just assuming Starship development goes smoothly from here on out. Which seems optimistic considering that, even by the standards of spaceflight, its development thus far arguably qualifies as cursed. Considering the gap between original design goals and reality is already historically large, I certainly wouldn’t be taking any cost estimates at face value right now. Like for any launch vehicle program, we won’t know its true capabilities and costs until it’s operational.
Not the same as Uber all. Lots of employees got fucked because at lockup expiration it was much lower than the ipo price (at which they were taxed at).
It seems like SpaceX is not as deeply tied to unprecedented debt-financed infrastructure buildouts at the same moment that interest rates are pushing up and up and up though.
That's the primary angle of depreciation of Anthropic's outlook, is that as interest rates are running hot, they need to grow ever faster to justify the $2B target, at the same moment that they're growing slower.
So many people who don't actually read and evaluate things but just buy stocks based on vibes and what they see on Twitter. Which is to say, the same people that buy and lose money on most IPOs.
I'm not convinced that 2026 was peak of the AI hype cycle.
Arguments are being made that, with the midterms over and 2028 just around the corner, a less sympathetic administration would create significant headwinds, adjusting the valuation of Anthropic and OpenAI down.
What we don't know are what Anthropic and OpenAI have planned and in the works.
A large population of extremely skilled and talented people are being laid off by their employers to make space for AI investments. It's likely a significant number of them will start their own business with AI.
AI is also gaining a much wider audience and acceptance. It's likely over the next few years, AI demand will grow, not stagnate.
That said, IPOs are usually liquidity events for LPs and for something really disruptive and unpredictable as AI, I won't be engaging in their IPOs. Only time will tell if this was a good decision, or bad.
Anthropic's original $2T valuation implies they are going to almost 100X their revenue.
I am highly confident that Anthropic will 100X their token sales volume. I am also confident that token prices will continue dropping through the floor due to competition. So to 100X their revenue they will have to 1000X or 10000X their token sales. That seems a stretch.
As AI grows in capability it will be trusted to do more of the work it already does and also break into new fields. Also, there are many who have never interacted with a chatbot. There is an entire world of potential customers out there. It can be quite a villainous business model but look at what Facebook did in Africa with their sponsored ISP and also what they’re doing with Meta Muse. Demand for AI usage will skyrocket.
Chatbots aren't a money maker. Each interaction is likely to be 1-2K tokens at most. AWS Bedrock prices 11K token input and 4K output for a total of $0.184 using on-demand usage [1]. So for normal usage (occasionally asking questions) you are looking at less than $1 per month for a person.
Using the models for data processing (a lot of input/output tokens) and for role-playing will use more tokens and thus cost more. However, those use cases are not common when comparing the percentage of the user base.
The money maker is agentic coding where people get the models to perform long-running tasks. This is why Open AI and Anthropic have been pushing these hard. However, there is pushback on this, especially among businesses looking for value for money/cut costs [2] like Apple [3], [4].
There's an entire universe of companies outside the MAG7 or VC-backed startups. I used to work for a US company, employing 60.000 people worldwide, which made the switch. But you will never hear about it because it's not a "tech" company.
If this is a consensus viewpoint, it'll likely be priced into the IPO.
I'm cautiously bullish on AI in general, and I wouldn't touch an AI stock with a 100m pole, because there are way too many uncautiously bullish investors in the market.
We have, for example, no idea what the long-term implications of AI effectively training on its own output is. Recursive self-improvement is not a law of nature. If anything, the opposite: the overwhelming majority of species are extinct. Evolution is as much a story of dead ends as survival of the fittest. So betting on any one narrow path of AI development to survive seems extraordinarily risky from a biological standpoint.
But will it increase by 10000X, as I back of the envelope calculated it would require to justify Anthropic's $2T valuation in the face of quickly dropping token prices?
Economic theory would state that if token prices go up people will stop offloading work onto LLMs and start doing it themselves again.
Indeed, 2026 was basically that: the LLM providers increased token costs and companies cut LLM spend to the bone. That is why Anthropic's IPO is struggling and MS and Google have failed to meet AI-related revenue targets for 3 straight years.
Or incentivize using the minimum viable model for the job instead of throwing multi trillion parameter Fable sized models at summarizing documents or classification tasks.
I now do like 90% of my routine software dev/sysadmin stuff with Luna on Medium by default, some proportion of which could itself be offloaded to a small Qwen model.
> A large population of extremely skilled and talented people are being laid off by their employers to make space for AI investments. It's likely a significant number of them will start their own business with AI.
What's you're point there? People laid off because of AI will immediately shift into AI-centric small business, and start hyping AI?
That doesn't make sense to me. People being laid off are likely to be bitter about the cause of their layoff, and even if they all start their own businesses (which is unlikely), most of those will fail, leading to more bitterness.
Many people who are being laid off will not find another spot in their industry and will need to pivot. Most of that subset of people will go to existing companies and take traditional roles. The small few remaining might take it as an opportunity to start a new business, and when they do so, they are likely to take risks on technology and business operations which will include AI. Is that so hard to believe?
I don't buy for a second that people laid off over AI will be so darn mad at AI that they won't use it, but I don't know where the hell the markets for all of these hundreds of thousands or millions of new small businesses are.
If these giant monopolies couldn't squeeze revenue out of these employees using AI, why would we think that they will be able to do it by themselves?
The vast majority of them if they want to be entrepreneurs will be better off starting a new lawn care business, or a bicycle repair shop. They can use AI to help them run that. Getting paid $175K per to write ruby CRUD was a moment that is over, hope you didn't spend all your salary on toys and rent.
> If these giant monopolies couldn't squeeze revenue out of these employees using AI, why would we think that they will be able to do it by themselves?
You assume they are competent. In my experience a monopoly's profit margins makes it very easy for incompetence to flourish.
> The vast majority of them if they want to be entrepreneurs will be better off starting a new lawn care business, or a bicycle repair shop.
Really? Experience engineers are better off starting a lawn care business or a bicycle repair shop? I do not believe this for a second. There are many areas engineering skills can be used in.
They will try to start businesses out of necessity as stable employment is not an option, and they will use AI since it's almost mandatory now for any technical productivity.
> In 2026 everything fron sewage to chickens to oil to education is now tied to AI. The entire economy is now, for good or bad tied to AI.
Tied how? How strong is the corellation between these? I'd say very weak, other than everything having to compete with AI for resources.
The same is true of any technology tied to powerful personal computers or the internet - people built and designed beatiful houses, great cars, airliners, wrote books and ran restaurants without computers or the internet.
It has even become a meme how putting smart features and lcd screens into things like washing machines have made them strictly worse.
That's a pretty damning statement - this implies companies don't see a return on AI at best, and it's actively destroying their businesses at worst, so there's not only job loss, but businesses themselves are not becoming more efficient.
Even if it takes them 3-6 Month to open up a new field, its going to happen steadily now and were it started, the AI investments will be recuped by layoffs.
Lawyer, accountants, doctors, architects, and engineers won't be replaced.
We're not paid just for what we do, we get paid because we accept liability for our mistakes. Professional liability insurance is a large part of why licensed professionals charge so much money for their services.
A lawyer can't just sign off on basic inquiring with "just reading". They're legally obligated to actually review the response and verify the answer (including any citations or references).
Lawyers already need less people to prepare things. That revolution happened several decades ago when computers and word processors became normal.
Lawyers as SaaS companies has been a thing for 2 decades, and remains a tiny niche of the market because it's not the kind of lawyering most clients need. A lawyer SaaS is basically just the equivalent of an underpaid, underresourced public defender and will yield about the same quality of work and results.
We will see if we will have one expert ai lawyer per topic or type of lawsuite or if they will get generic good enough that it will not matter
LLMs are only as good as their training set. In order to be a topic expert, they'd have to be retrained at least every week to account for changes in statutory law and case law. It would be trivial for a human lawyer to outmanuever an outdated LLM when the laws changed...and in the real world this happens all the time between human lawyers when one side hasn't done its research on changes to the applicable laws.
Even people defending or suing by (successful) themselves became a thing already last year.
People successfully defending or suing without a lawyer has been a thing for over a century. It's fewer than 0.1% of legal proceedings. Most people don't handle their own legal cases for the same reason most people hire a plumber, or electrician, or order DoorDash instead of cooking their own dinner. Also...almost all of the people using LLMs in the past 2 years lost their cases, because their LLMs hallucinated laws or facts and they didn't know enough to catch it. Even lawyers dumb enough to use LLMs to draft their documents got caught because they were too lazy to review the output. Courts were willing to forgive those mistakes a few years ago before it was common knowledge that LLMs make things up, but now they're no longer willing to give the benefit of the doubt.
The biggest advantage of an LLM is the interface and reading text.
Whatever we had before on full text search, LLMs are a lot better. And i'm not just feeling this, i read plenty of articles about exactly this.
A LLM doesn't need to be retrained every week, it needs to have a base corpus of the field and strategies on how to find current text. They will load it into context and argue.
Then another LLM will take this and formats it for the court.
We will see but i think you overestimate how much automatisation and co current lawyers have and use.
Do you really see lawyers being replaced? I don't see anything special in their work but since it's a regulated industry I doubt it will be so affected.
At best some layers will work faster, but automating their work or people not requiring them? Usually when you contact one there's a lot at stake. Same with accountants.
Software engineering and translators is a different story. Most of the times if something goes wrong nobody cares.
Some lawyers (the worst ones?) might just do the easy or normal things like writing a 'lawyer' letter to someone not paying. This is no risk to automate away.
The avg Lawyers might do more complex but again a lot of 'normal' things. Day to day things the next SaaS AI company wants to 'fix' and they might also accept basic liability.
Then you have the 'tv lawyers' the ones who do non normal things, they have A LOT of apprentices and young lawyers going through papers, research etc.
They will for sure have the expert as a human, but that person will need a lot less support.
You shouldn't need an accountant today already. If you have an digital order and a digital bill and a digital signed delivery note, it can be booked automatically. AI will find risks, summarize them etc.
AI is already helping here closing the gap with better OCR etc.
I really think, that when big companies introduce the agentic layer to their products, there will be less people needed and if you need people, they get less paid because they will need to understand less.
Somewhat ironically it would appear most CEOs believe the only jobs AI is replacing are developers since they built an app last weekend. Every CEO is saying why do we have so many devs and why are we paying them so much… so devs blew up the market for devs but beyond that there doesn’t seem to be much impact.
> What we don't know are what Anthropic and OpenAI have planned and in the works.
Wouldn't they tell us before IPO? Unless it's not that promising...
> AI is also gaining a much wider audience and acceptance. It's likely over the next few years, AI demand will grow, not stagnate.
That does not mean the current players profit from this and the infrastructure investments were necessary at that scale. There is competition, there seem to be enormous efficiency gains and it's questionable that demand for even bigger models is large enough.
The major counterpoint here is commoditization; you may have the market growing and models becoming cheaper at the same time. And we don't even have to look to Chinese models for that, as OpenAI has been adjusting its pricing already. For profitability, Anthropic still seems to be pretty dependent on its current price levels.
That said, I'm also somewhat skeptical that there is even much room for growth left unless somebody finds a new, important, and token-hungry use case.
I believe that the value that OpenAI and Anthropic hold, in light of stiff competition from China, is not their products but their internal models. China’s work is all derivative of American model API access. The exponential RSI loop means that as long as Anthropic and OpenAI keep steady footing and only release older models they will will always be the best on the market and all others will be in their tailwinds. These two things combined are what make me believe OpenAI and Anthropic will hold value and grow as companies.
Even if we were to accept your entire argument [1], none of it contradicts the commoditization argument. Note also that competitition does not come just from China, but also from Google, Meta, and SpaceXAI. Where they currently are is not relevant, the entire LLM landscape will likely look nothing like today by the end of 2027.
[1] I don't, I think China is primarily limited by compute and will be at least for the next couple of years, probably longer.
Best on the market, for something not being sold on the market? How is that helpful in terms of actual revenue?
If Anthropic manages to start producing new pharmaceuticals with their internal models or something that thesis could make sense but not if 90% of their revenue comes from enterprise API sales which would necessarily have to be accessible to the public. Which are then inevitably fast followed by Chinese models 3 months later.
In theory, yes. But the reality is that deep learning models just aren't good at creative output (to put it mildly). And neither Anthropic nor OpenAI have any models that generate video, so this wouldn't help them; in fact, OpenAI killed off Sora.
It certainly isn't the peak of the AI hype cycle. Eventually something will surpass LLMs and the hype will start all over again. Will it really be the final invention this time? Place your bets!
> a less sympathetic administration would create significant headwinds,
I find myself wondering what "less sympathetic" and "significant headwinds" mean in the minds of people in the upper echelons of the AI industry.
Rules for training frontier models? Requirements for renewables to power data centers? Actually having to pay people for the content you train on? Data center rules that would end them going behind the backs of local electorates to get governments to approve the building of data centers?
Tulips, Railroads, and the Internet also didn't go away when the financial bubble around them burst. That doesn't mean there wasn't a massive market correction in response to the over-investment.
> Tulips, Railroads, and the Internet also didn't go away when the financial bubble around them burst. That doesn't mean there wasn't a massive market correction in response to the over-investment.
Your Railroads and the Internet parallels are so very apt. We certainly are very productive even in 2026 off those but almost no one remembers how many large behemoths were wiped out in the U.S. because of them.
Yes, it does seem to me that the AI of 2026 will be like the Railroads and the Internet - fundamentally transformative technology that changes how things are done into the future but there will be blood in the short term, which is why I would not be participating in the IPOs directly
The interesting thing is I think we might over hype the tulip hype now.
It seems like much of the tulip mania has become the stuff of legend.
From what I have read, it seems closer to baseball card trading than AI or dot com.
You can find baseball and pokemon cards trading for thousands of dollars right now on ebay.
This would seem insane to a culture just valueing the cards as pieces of cardboard.
It would be like hundred of years from now people viewing our society as wanting to buy Mickey Mantle cards instead of houses. It is a very small number of people involved with this. It is not a widespread movement to hold your retirement money in Mickey Mantle cards.
In an interesting way it shows even better though he we are susceptible to these type of mythologies. People don't even question the absurd givens of the narratives of the tulip mania. We just repeat these narratives to each other as if they are historic fact.
There's a beautiful tulip garden just down the street from me, and an outdoor weekly market where tulip bulbs are sold. Tulips are very much still around, I assure you.
I don't mean went extinct. I mean the idea that tulips could be an asset class is now laughable, which is not true of the other two. The impact of the tulip industry on GDP is utterly imperceptible.
Yes, that's my entire point. There was also a railroad hype cycle and an internet hype cycle that took in massive amounts of overinvestment, and all three things remained as regular things in the economy with bigger or smaller roles after the hype cycles ended.
Only popular with companies that want to make a quick buck and tech people (hey! that's me!).
You really don't know what we have here, do you? Let me illustrate what you can do with Qwen3.8@q4 locally with about 2TB of storage and a small cloud server to host a website.
Locally you scrape and process every elected official and start tracking what they're doing, what they're promising, and what they're not doing. You build embeddings and make them all geotagged in a database and just start with the largest cities.
The website provides public facing search. User drops in a zipcode, city, state, specific rep's name and gets what's relevant to their results.
All of this is to say, we literally have the technology to hold accountable our representatives and then show everyone objectively who's doing good and bad for their constituents.
It wont take much to solve the problem of bad representation after these last two years. The hard part is literally making sure people going to a polling station can immediately know who the heck the 10 - 30 different people are they're voting for, and track them efficiently within their own specific parameters.
Regulation is coming for more than just the inference providers.
I'm still working on my implementation. The hard part is data sourcing. This is about to run me a TB in storage to do and it's going to require some data archival. My estimate is that I can serve some, but not all cities. Just keep an eye on my user posts, I'm sharing it when it's live.
Not making a leader board to be clear though. The goal is to build a branchable leaflet style page based on zip codes that lists ALL the people that were elected. Clicking into the each person will yield details on any legislative actions, their attendance, publicly made promises, etc.
I'm really sick of hearing lip service from politicians so I want to build receipts that even a child could understand.
These are a good start, but they aren't quite enough. I just voted for ~25 different people going all the way down to local government. These don't answer the questions I'm about to be answering top down for myself and neighbors.
You can maybe replicate Opus 4.5 to 4.7 if you have some moderately expensive hardware. And that is definitely good enough for many people.
The smallest model that can even pretend to be a decent coding model is Qwen3.8 27B, which you realistically need 32GB of VRAM to run without breaking or limiting it in some really noticeable way. This gets you something like a slightly less capable Opus 4.5 with a much better work ethic.
Strata and Qwen 3.8 flash next is an 180B model running at 70 tps on my 5070 ti and 64gb of DDR4 RAM at Q3 quants and so far I can't tell much difference between it and any other non local model for my workflows/coding needs. Not saying local is better all the time, but it's certainly a lot more capable with minimal hardware than you're giving it credit for. Qwen 4.0 is right around the corner too.
The problem is that "can't tell much difference" is not going to be an objective indicator; you will be very likely to overlook the flaws of the local model and overweigh the flaws of the non-local model. The only reliable way to compare the two is to have your own set of evals based on the types of tasks you regularly perform and then run those evals on Qwen and a frontier model.
The people saying that local models can easily replace the real Frontier Models from Anthropic or OpenAI are just not using it for the same things you and I do.
They seem to be happy with their small usages.
For everything I do, i see huge differences in Quality and Speed.
Even when I try an open model on a 4090, the PC gets warm/hot very fast, I struggle with context length.
Having currently a very tech-intensive chat running for more than one week, this is where the frontier models really shine: Its no more 2024 or so, where they lost context on large discussions, you can do really great work with the frontier models - the local models will be a thing in a few years, if RAM & CPU developed further, this will happen: in 2035 like, you will be able to run "Claude-like experience" on your iPhone.
I write 5-20 paragraphs of things I want and it goes away and just builds it.
Something like this (just an example):
"Build an application with a webserver in rust. It has 4 stages, first one is an input module... second one is refinement ... third one is an image detection with current fitting ml model ... last one is finetuning."
Then i would describe every basic page with a few more paragraphs.
This would work for half an hour or an hour, consume 100-400k of context tokens and then I would test it, write down 10-50 small sentences (every sentence either fixing a bug, or making it better) which tehn takes again like 20 minutes and i'm nearly done with a complex enough tool providing real value and actually working.
A big kimi, which is hard to run, might be getting close to this but we do not talk about running it on 1-2 5090.
The other thing is, thats more like a practical experience, even if you use them through Open Router (for testing and co), the initial results of something like Opus 5.5 are ALWAYS better. More details, better quality, better understanding of usability of the tool.
> In a work tree build a feature where each column cog action menu has a "Column statistics" item - clicking that item opens a modal dialog which displays statistics about that column - for text columns that's most common values, stats about value length - for numeric columns it's more interesting than that. Let's do a bit of percentile analysis in there too. Under the hood it works by firing off SQL queries using SQLite window functions and suchlike. For the moment let's have that option only available if SQL queries are allowed against the database, and the SQL queries are built in JavaScript.
> The people saying that local models can easily replace the real Frontier Models from Anthropic or OpenAI are just not using it for the same things you and I do.
> They seem to be happy with their small usages.
Oh yes, we're terribly inferior people. Just living on scraps from the masters' tables.
I do plenty of experiments of things i always wanted to do and they just don't work with local models.
Finetuning of some ML models for a specific usecase, some 4k demo idea i had, small browser games to show others game mechanic ideas i had for ages, analysis tools etc.
Even just playing around with some Ghidra eats context like nothing.
I really strongly agree with this statement. A bunch of people aren't even engineers, and their abilities aren't that good either, yet they keep saying that open-source models can replace OpenAI or Anthropic's models.
If that were really the case, basically nobody would spend so much money using them.
Putting aside that you're "really strongly" agreeing with a personally directed insult:
You should factor in that undeniably most people don't have the means to run even a 27B dense model locally, and therefore do not actually know what they are and are not capable of in that context.
And you might want to look at openrouter's rankings.
If you are comfortable automatically assuming that "doesn't use Claude or Codex" is equivalent to "not a good engineer" or "bad engineer", then I think you are probably lost to more nuanced discussions.
With my DGX Spark + 5080 + 3060 setup, I've tried all kinds of approaches, including quantization, acceleration, fine-tuning, post-training, and everything else.
It still can't compare to Opus 5.5 max in terms of speed, capability, and stability.
And when I calculate all the money I've invested, directly subscribing to the service would actually be cheaper, more efficient, and more stable.
Open source is great, but don't glorify it.
The architecture of the code I develop has never been simple. If you're someone who doesn't even work on things that need to be deployed to something like k8s, then basically, open-source models are indeed enough for you.
I try local models regularly on 2x 4090 and struggle with context length constantly. 48GB on GPU and 128GB from RAM.
the original comment mentioned only 16GB of VRAM.
You know talk about Open Router and 27B dense model?
We are not comparing an even more expensive Hardwaresetup to run local models with proper context window and parameters.
Can i run some open model on a super expensive Mac Book or co with some kind of okayisch results? Yeah sure but you know what the t/s will be and yeah Frontier is still Frontier and Kimi, GLM, DeepSeek are not Frontier. They lag half a year.
You are not getting Opus 5 or Sonnet at home for now.
I've try every new model when it comes out from both OpenAI and Anthropic. I cannot justify leaving Codex and Astra 6. The token burn on Anthropic is always astronomical for me. Codex just works, and I see no reason to move away from it.
I switched from Codex to Claude Code w/ Opus 5.5, and it's been fine. I haven't hit limits in my $100/mo plan, doing full time dev work all day. But I agree. Codex sub + Astra is also great. Both are excellent options, and I don't see a reason to go with a cheaper / less capable model, frankly.
12GB GPU is useless. The good experience is with 128-256GB unified memory architectures like strix halo etc. You don't need to buy any hardware, you can experience the latest open weight MOE (flash) models from DeepSeek, qwen, or GLM on openrouter using a local agent like kilo.
That is a shift of comparisons though, right? This thread was about running locally, explicitly as an alternative to cloud services. I would guess few people have graphics cards with > 12GB ram.
A few more will have Macs with 48GB RAM though. (Indeed if you push it a little, 36GB unified is possibly enough for Muse Glimmer; I have not tested this theory)
There is, for sure, a limiting factor in the x86 architecture, but I think the RTX Spark will change that a bit, and maybe AMD's subsequent Ryzen AI/Strix Halo chips.
AMD Ryzen AI/Strix Halo is for running locally on your hardware. Openrouter is for evaluating the models you'll be running locally, before buying the hardware.
In any case, 128-256GB unified memory systems running MOE models is the "cheap" option, RTX6000 is the step above it.
I think the key to making it make sense is that people have wildly different demands for quality. You will see people say that these small local models have improved quickly and are "good enough for most work". While these models have improved and very fast, it still might be more of a comment on their expectations than the product.
12GB is not a lot, you are stuck on low quantisations of older models with that. What exactly did you try?
The newer/larger models are good though, not as good as Claude, but not every task requires a model of that level. Maybe yours do, but for simple coding tasks, summarising text, etc a small local model runs just fine, and is a lot cheaper than Claude
What graphics cards have more than 12GB, and how much does it cost to buy one? What about for others saying you ideally want 32GB? (That split will determine if a 5080 will do it)
> What graphics cards have more than 12GB, and how much does it cost to buy one?
RTX 5070 Ti and RTX 5080 both have 16 GB.
RTX 5090 has 32 GB.
Good luck finding any of them for reasonable prices, though. I paid $3,000 for my 5090 early last year. I winced when I bought it then. Today, it's like $7,000, and I wouldn't even consider it at that price.
But I'm mostly a gamer that also does local AI as a hobby. If the RTX 5090 was $7,000, I'd still be on my RTX 3080 and still just not started doing local AI and have to turn down the graphics settings in Cyberpunk 2077.
The RTX 5090 is the ideal, but it's not the only 32GB GPU that can offer workable performance; the AMD R9700 is catching up.
You can surely build an LLM server with an R9700 and 32GB of mainboard RAM for less than the 64GB M5 Max Apple Studio, which is probably the next most competitive model up. It'll be better than a Strix Halo but less flexible.
- 4090: Going for $7500
- 64GB M5 Max: Starts at $3,800
- R9700: $2700
Compared to $20 or $100 per month. My thought process reading this thread:
Baseline of $20 or $100/month Claude or OpenAI subscriptions. Alternative posed of running models locally. A guy says he did it and results are lackluster on a 12GB video card. A number of people post that that card is insufficient; use 32GB+ cards. I look up prices of 32GB cards, and they are all very expensive. So from a practical perspective, people are suggesting you spend thousands of dollars on a video card to get results that are almost as good as the subscription plan they are on. It doesn't add up, unless one of my assumptions is incorrect.
The assumption that is probably incorrect is that the current prices will last. A 4090 should not cost that much, and most of your Claude usage is paid by an investor. Neither of these things will be true in a few years.
Having a 4090 at home running agents all day is not a $20pcm level of usage, so take the $100 plan. Looking at the S-1 filing, today they are not even close to breakeven. It is reasonable to assume that will become a $300-400 pcm plan.
The launch MSRP of a 4090 was $1600. There has been significant inflation since then, so say $2000. Eventually supply will meet demand and you'll be able to buy one for that price again, or lower due to the economies of scale they must have now.
So, in 2-3 years time once it has all settled, you pay off your 4090 by saving 5-6 months of Claude usage. Maybe the local model isn't good enough for complete replacement, but all it needs to do is drop you to the $20 plan to be worth it. I think this is a large part of why investors are getting jittery now.
Right. Nobody says it's necessarily economically the best option. But it's also an asset that has alternative applications and barely depreciates!
Right now you could extensively use an M5 Max for a year and sell it for almost exactly the same amount at the end. Maybe even more, at the current rate. And it can do other things for you: host your VMs, edit high def video, whatever. Even if it stops being useful to serve LLMs, it might be useful to run as a server for other tasks for a whole decade. It will be a fast, supported Mac for seven years. For the R9700, maybe gaming?
Depends what you want, doesn't it? Or are we all expected to conform?
For me personally, right now I have a secondhand M1 Max macbook pro, and that is my local AI machine, because my needs are exploratory and I want to run it here. I want to learn by comparing models. I want to learn how to run them and advise others how to (local inference is not a zero-sized market).
My next step up will be to take the apps I'm looking at building that run against these small models, and host those apps online with some sort of cloud inference — openrouter, digitalocean, cloudflare — for those same models.
If I ever find I want to vaguely describe some desktop application before bed and come down the next morning to an aggravating slop version of it written in Rust, I'll consider Claude.
For a local non-laptop, I've looked at the M5 Max and the R9700. I've also looked at the M5 Pro, which in practice is about as fast at token generation as my M1 Max (because the M1 Max's memory bandwidth is strictly theoretical), and smokes it at prefill. That alone might be enough for a "leave it running" mini AI box, and it will be a nice desktop machine. I just don't need it enough yet.
As an aside there's this thing — somewhat interesting, though constrained:
Uses the 5090 mobile chip, which is pretty good speedwise (about the same as the R9700) but 24GB. Small, tidy, on-prem. I think they have missed the boat on pricing, but if they can ship in volume, maybe not.
Great insights! And I'm kicking myself for not buying a 4090 3 years ago for $1600, shipped direct from nvidia... it seemed like too much to pay for a GPU at the time!
I know a couple of people who have eminently LLM-capable GPUs more or less doing nothing when they are at work…
3D games give me motion sickness so I don't have powerful PCs anymore. In fact right now I don't think I have a desktop/server PC that can power up at all, and that alone makes me wonder about building an R9700 machine.
there is also a version of the 5060ti woth 16GB. i got lucky a few months ago and got it for ~380€, but i think you can still get it for around 800, definitely under 1k
In general, more memory = larger models = more intelligence.
However even if you had a private data center you wouldn't be able to beat Opus 5.5, Astra or Gemini 4. The open weights models are a tier below these three.
Memory manufacturers are running 80-90% margins. Any of a number of things (anti-trust, CXMT, financial shock) have the potential to drop memory cost by 2/3s. At the same time memory efficiency of models is improving. At some point Apple's going to start shipping 64G of unified memory in mainstream machines that are advertised as local AI capable (like HP, Dell and others do today). Like the iPhone refined the concept of a smartphone for the mainstream, they could be the first to hit mainstream sovereign AI.
Good enough? Quite probably, unless you're in the "coding is solved, specify loops and walk away" frame of mind.
32GB (or say a 48GB-unified RAM unified machine) gets you Qwen 3.8 27B at a good adaptive Q4, maybe Q5. Muse Glimmer at Q4 and Q5, etc.
These models can code well enough in an agentic way, I find, and the chat-to-assistant scenario is just fine in Unsloth Desktop or LM Studio.
A slightly lighter harness helps; a lighter system prompt is entirely possible, especially for Muse Glimmer, which is post-trained so effectively for agentic helper tasks that it barely seems to need a system prompt at all except for a "plan" mode.
Well it's certainly the service-dependent (addicted!) model that the big two want.
But will you want to do that when your tokens are metered at non-subsidised prices, and yet there is absolutely no clarity (ever!) about token efficiency, how the model decides to spend your money, how it might ever decide to stop wasting your money and get to the point, etc.?
The specify-loops-and-walk-away model absolutely points to local/on-premises/private cloud AI.
Because otherwise you're letting Anthropic or OpenAI decide how to spend your money. And they are going to be biased towards spending more of it.
I don't think Qwen 3.8 is quite there yet with the loops stuff. Definitely seems to be better at deeper and deeper one-shot puzzles.
Who cares what the big 2 want?
Have you taken a look at how much the hardware/algorithms/efficiency are improving?
Those 2 are building sand castles in front of a tsunami.
12GB GPU is insufficient for most purposes yes. Even 16GB is a notable improvement, but yes 32GB cards such as the 5090 or R9700 are a world of difference.
Given we're discussing a cost comparison (i.e. cutting a Claude subscription in lieu of running locally), how many months of claude $100 or $20/month plans would it cost to buy a 5090? I am showing it as 88 and 440 respectively. I.e. a 7 and 36 year payoff. Not even close to being worth it.
The Claude plans are probably still heavily subsidized. Expect prices to increase, especially after the IPO. Self-hosting models for personal use is probably overkill, but I can see companies doing it pretty soon since they can probably run 24/7 workloads and fully utilize the hardware.
Great point. And, 3 years ago, a 4090 cost $1600, so maybe that will happen again some day. (Same/similar card, or a future card. I remember when GPUs used to get better every few years instead of staying roughly the same but costing several times the original amount!)
"SemiAnalysis, a research and analysis company specializing in the semiconductor and AI industries, found that a $200/month subscription plan is far more generous than previously thought. According to its calculations, while it's widely believed that a $200/month plan maxes out at roughly $2,000/month in tokens (assuming API pricing), the actual numbers are closer to $14,000 for ChatGPT and $8,000 for Claude. "
You also get to keep your data private (so you should compare it to api pricing), there are no forbidden topics (i just did a pentest on my webapp), and multiple people can use it. Basalt does 150 tok/s on a 5080
Reminds me of linux gaming 10 years ago. According the to the community everything worked great without issue and any issues were hand waved away as tiny bumps that you won't even notice.
Then you would go try it and be back using windows to game a week later.
But that is not exactly a linux issue. Games built for linux works very well on linux like games built for windows running decently well on windows. Linux can also run some games built for windows - not for linux. And that are the ones that doesn't provide first rate experience. Those windows games executables technically should not even load as it is not built for linux.
Everything considered, linux is good and has decent capabilities including gaming and whatever you can throw at it but only do oranges to oranges comparison. Games and apps built for each os works fine on the respective os, well atleast it is true for mac and linux. Windows is very unreliable.
Imo, this is just the tech version of virtue signaling. Everyone knows local models are objectively just worse than frontier/cloud closed source models for most commodity hardware setups, yet people insist on using it because they want to be contrarian or "stick it to the man."
Your experience is average because it _is_ supposed to be average. The people using local models tradeoff taking the moral/technological high ground vs actually using something high quality to get a job done quickly and with high quality, and largely autonomously. They are ok with subpar output because they will handhold the model more, or the (subpar relative to the frontier) outputs are sufficient for their toy use case while being cheap and accessible, with the added benefit of being able to run it on their private data without sending it to a 3rd party.
Almost no one at any job where software is the main economic output is using a local model to code. If you actually require intelligence to get a hard job done, you should just use a frontier model.
> If you actually require intelligence to get a hard job done, you should just use a frontier model.
Or, you know, your own intelligence.
What I do not understand as an engineer is how our community has gone from believing in open source for controlling destiny, security, etc., from avoiding unnecessary corporate capture (think of all the rants about Swift!), from believing in heterogeneity and portability, to just saying, look, if you're not using the product of one of two loss-making AI unicorns to do all of your work, you're just being a contrarian weirdo who is doing it for performative reasons.
I guess if that is your definition of contrarian — that everyone else is doing it so why aren't we? — then I am.
But in my experience it's not so different to "why aren't you on social media?" — which was an argument from insecurity and the desperate need to fit in.
I personally believe in independent thinking. I have things I want to learn about how all this actually works, I have questions about my own life and future, and I'm not going to simply chuck them away because you or anyone else thinks it's just being contrarian and we shouldn't question the direction everyone is taking. So I am learning the tech with a stack I control, while continuing to do the job I do.
You can do what you want. You're probably right, and I'll be quite happily wrong and continue in my wrongness even though, based on how you characterise it, this kind of independent thinking appears to unsettle you.
Because local models are significantly worse? I'm not so sure what your question is. What else did you expect? If you could just build Claude Code at home, Anthropic would be valued at closer to $1.1M than to $1.1T.
So for a sense of scale, un-quantized 1 parameter == 1 byte of ram. 12Gb means you could have roughly 12b parameters, frontier models like Kimi k3 has 3 trillion parameters, so you'd need ~3tb of vram.
> The results are ALWAYS terrible compared to CLaude Code.
How much effort did you spend to make the local setup work well?
Local models are not plug and play, you need to adjust settings to make them behave intelligently.
On my side one model looking super dumb right after install ended up being 'agentic' after using the right chat template, fixing how tool calls are setup, fine-tuning plenty of params values, and adjusting the whole setup to my own hardware. Claude.ai (free) was great at helping me sort that all out.
I really strongly agree with this statement. A bunch of people aren't even engineers, and their abilities aren't that good either, yet they keep saying that open-source models can replace OpenAI or Anthropic's models.
If that were really the case, basically nobody would spend so much money using them.
With my DGX Spark + 5080 + 3060 setup, I've tried all kinds of approaches, including quantization, acceleration, fine-tuning, post-training, and everything else.
It still can't compare to Opus 5.5 max in terms of speed, capability, and stability.
And when I calculate all the money I've invested, directly subscribing to the service would actually be cheaper, more efficient, and more stable.
Open source is great, but don't glorify it.
The architecture of the code I develop has never been simple. If you're someone who doesn't even work on things that need to be deployed to something like k8s, then basically, open-source models are indeed enough for you.
I (all other issues aside) honestly think we have lost all sanity when the term "low" (in any fashion) can end up associated with 1.1 TRILLION.-
There is a serious, malignant, dangerous disarray somewhere in our callibration of value, of worth, of what things are worth. Something, many things, are wrong here.-
I think you’re talking about the headline, and it’s clearly not really calling it ‘low.’ that said, things are relative, and relative to the market’s position on AI and AI related spend, anthropic’s overwhelmingly optimistic spending commitments, how they predict their machine god will take over the world, etc. - 1.1 is indeed “low”
I don't trust the market to calibrate "value" the same way that I subjectively define it.
But I trust the market to calibrate value about 10,000x more than I trust any person or group of people to do so. The thought of that is frankly terrifying.
To be more precise, the reason I trust the market is the "skin in the game" and the decentralization / localization of decision making. People with something to lose tend to actively "vote" only on what they know something about.
Not that it's somehow "not people" (of course it is).
> People with something to lose tend to actively "vote" only on what they know something about.
"Prediction" markets have rendered this clearly false of late.
> the reason I trust the market is the "skin in the game"
The market expects economic bailouts and other assistance during downturns, because individual investors with 401(k)s demand it, despite the perversions of market incentives that result.
And there are worries about the Democratic midterm election sweep causing tighter scrutiny of new data centers and forcing Anthropic to pay wholesale for chips, or worse.
can someone explain to me the point of the 'initial valuation', considering ultimately the price will settle lated thorugh trading to perceived value anyway ?
None of the value changes in later trading mean squat to how much actual money the company gets to put in their bank account.
They get the money from the initial purchase of the shares, and never get another dime directly from the market again, unless they have another offering (which is quite possible). The stock price does drive leverage for later M&A and debt/credit activities, and matters to stock-based compensation, so it does have meaning, but the IPO provides a one-time cash infusion at the initial price.
The point is that is the price the private investors who will be liquidating, including the company itself, are accepting for their shares.
If they are too far below the perceived value, they are leaving significant money on the table.
If they are too high, the institutions who participate in IPOs will let them know. If the price craters off IPO it is not good for perception of the stock, and so ongoing health of the company even if that maximizes private investor value at IPO.
So they build a book of commitment around the target value, trying to leave enough incentive (15-20%, often) of an initial bump to get big money in.
If the company is valued by the underwriter and initial buyers at $1 trillion and the company offers 10% for sale through their underwriters (who have put out feelers to try to attract buyers at highest price possible), they raise $100 billion from the offering, minus fees, to use as needed to further the business. The initial buyers get the shares at that valuation, and I believe if the underwriter positioned it wrongly they might be on the hook to sell some shares at a lower price without prejudice to the company.
If one of the initial buyers sells in the first minute the company is public the first publicly traded shares at 1.5x their purchase price, the company suddenly has a market cap of $1.5 trillion instead of $1 trillion, and that buyer has made a tidy 50% profit on the shares they just sold. If all initial buyers dump a bunch of their shares the price likely will go down, perhaps pushing the market cap to, say, $750 billion. Regardless of this action on the first day and after, the company offering their shares still has the $100 billion.
Note that initial buyers are discouraged from selling large portions of their shares.
Some people buy on "day 0", before it hits the stock exchange. This determines the price they pay on day 0. For example, if they buy 10% on day 0, they pay $130 billion.
The stock exchange is a "secondary market", which means you're buying stock from investors who own it, not from the company itself.
Investors ‘subscribe’ to an IPO. A company offers a certain amount of shares for sale at a certain price. Then the underwriting banks go out and sell the shares.
If your broker allocates you shares, you pay the IPO price as does everyone else who was allocated shares. The shares are all allocated before the IPO happens and once it opens for trading, price discovery happens.
How the price is set is another matter, one I’m not familiar with as an amateur.
Typically in an IPO, the offering is about 20% of valuation of the company at the first offering. But with a $trillion or more, they will offer say 10% or less.. say $100B in one go is probably what the market can take.. Some banker has to back it up initially and defend that price... the next dump will be about same amount say $100B again but not the same number of shares. The underwriters are not obliged to back $1T and the IPOed company is not obliged to release more shares.
Ultimately the percentage of shares that public will trade is material. say 20% for $100B is quite different from 10% for $100B.
> can someone explain to me the point of the 'initial valuation'
Most of the institutional and retail placements happen on the initial valuation. These are the people who are holding the shares when it gets listed and starts trading, and it is the primary source of raising capital through an IPO.
To go further though the Initial Valuation (usually a range) is set weeks before the IPO to make it easier to pitch to investors. And the final valuation is locked in the night before trading based on actual demand.
The opening price on the stock exchange, though, will be whatever the demand and supply is in pre-market trading/open.
Whenever you buy a business, you have to have some idea of the valuation.
The IPOs need to have buyers. The buyers will have some idea of the business valuation, as will the sellers. This is decided before the actual stock starts trading.
Specifically, the "underwriters" will go around finding people willing to buy what is being offered. This gives some idea of where it will begin to trade.
It's the starting listed price. For an anchor, it determines a lot about how the first few days go. It's also a great headline. It's also also what employees picture when they evaluate their stock options, I bet.
> the numbers we're discussing here are the amount they'll raise from the IPO
No, the "valuation" numbers are the market cap at the moment of IPO. The amount raised is that multiplied by the percentage of shares issued by Anthropic (as opposed to by existing shareholders like investors and employees -- that combined number is the float) at the IPO.
Sorry, yes, that. The bigger point still stands, though — given a fixed fraction of shares they're offering, the valuation gives you the amount raised, and that makes the number interesting.
Everyone is explaining what initial valuation is without explaining the point of initial valuation. The point is that existing private investors get to exit with a sizable profit while leaving the risk of future growth to the public market.
I think it's crazy that Walmart is worth 1T. It's grocery chain operating in limited number of contries. Shouldn't be worth much more than for example, Carrefour.
When you mention Tesla, don't forget Space-X. The value proposition of Teslafanboys was always that Tesla is not a car company and now Musk splits it up between Tesla and Space-X
This is not unexpected. Anthropic revenue growth is stalling. Ramp's AI index shows spend on Anthropic peaked on July 31. The spend in the last week of October is lower than last week of July -> https://ramp.com/data/ai-index?metric=token-spend&detail=mak...
Companies are cutting down on AI cost (adding per person budgets, using smaller models). Anthropic has also been forced to reduce token prices due to competition.
I don’t agree with you. Their product is a fully scalable almost human tier intelligence. I think it’s worth way more than $1 trillion but they have to get their capex in order first so they can provide dirt cheap tokens.
If the company cant convert the product into earnings it really doesnt matter what the product is. It can be a magic machine that turns water into wine. They still have to find people willing to pay for it.
If the world is not ready for fully scalable almost human tier intelligence then, being right at the wrong time is the same as being wrong.
Spot prices for 5 year old A100s are still going up recently, because inference is so valuable. This meme that GPUs are a quickly depreciating asset is outdated.
Their lifespans are 2 to 4 years with always on, heavy usage. Even then the problem isn’t the lifespan. It’s the lack of supply right now. I assume with more built out production capacity GPUs get cheaper too. The whole vertical stack needs to get cheaper and it will in time.
It's already much smarter than most humans in a lot of ways.
It's not the same kind of intelligence as humans so it'll probably always have weird areas where it can't do as well (like counting letters in words), but that doesn't mean it isn't smart. If AI were sentient it could point at humans and say "haha look how dumb they are; 99.999% of them can't even write a basic Lean program".
We're way beyond the point where "it's not as smart as humans" is meaningful.
I am retired and use index funds for my 401Ks. I hate the idea of Anthropic becoming part of mandatory purchases into retirement funds. I hope it doesn’t happen.
That said, Anthropic is an interesting company. I mostly use local models like Qwen 3.8 Next Flash, but when I occasionally use Opus 5.5 for difficult I enjoy the experience.
I am not being hyperbolic or deliberately obtuse. I (a person) can not think of a reason to need to use a large model for video gen, but in no way does this imply you are wrong or that I somehow know better than you.
You are of course welcome to tell me why you need to use opus vs sonnet.
My impression is that none of those systems seems to be efficient.
We see time after time systems that solves most 80% of the case to be 90% cheaper.
And the way it is, it seems to me that nor openai nor Anthropic can keep the current prices.
Some say that if Anthropic keep delivering improvements and be the first it will make it but for that, you need to disregard all competition from other llm companies and from other competitor systems.
And even though they have the smartest model ever, if someone offers something 90% of it for 90% cheaper it will be worth it.
We already see plenty of companies reducing ai cost because they don't see enough roi
1.1T is 27% less than what they are aiming at, and yet, it’s still absolutely mind boggling.
People love to reign in AI companies these days (and part of it is warranted) but the triple digit annual increase in revenue at Anthropic and others (SpaceX being in that category as well) is something never seen before.
Some people are being too quick (or too willing) to disregard the obvious immense economic impact of AI and company (and I’m not talking about the next-day-ai-crash-incoming economic impact).
Classic mistake of confusing the usefulness of a technology with the first movers ability to extract long term value.
Anthropics S1 numbers are hugely optimistic given the commoditisation that has hit, continued to hit, in Q3 and the temporarily hold on chip supply that is not going to hold for too long
Inference is projected to continue to become increasingly commoditized so you need to wrap it in compelling and differentiated product people are willing to pay for.
Training, people argue, continue to be very capital intensive so moat there. What I'm wondering, is anyone trying out ideas the lower the training cost materially? I'm not thinking distilling because that, by definition, would always be behind the frontier. Other ways?
- Someone made a "clean room port" of Total Annihilation in golang, looks like it started in August and is now the most convenient way to play Total Annihilation. https://github.com/nanolathe-gg/nanolathe
- You only have to decompile / reverse engineer an existing game once and now you can have an agent occasionally add features.
Once you do the port and have worked out the regression issues, assuming you are switching to a safer language with better guarantees, you are now off of the old codebase and don't have to pay the costs of the old codebase. It's just a one time burst of token usage.
Costs of tokens will continue to fall as more hardware either becomes rentable (datacenter comes online) or moves into the used market, as workloads move onto users individual machine (because smaller models are becoming more capable and getting better inference engines and harnesses).
You can also task an agent with researching for "list the top 5 performance improvements we might be able to make" and have it research overnight, then the next day start a fleet of agents iterating through the findings that seem the most workable . Once you gain the performance win, that's basically permanent.
I think that token usage is a very elastic market, you can choose to spin up a fleet of agents or... you can stay human-in-the-loop, or you can type code yourself if you don't have a token budget.
So, yeah, I don't think Anthropic or OpenAI are worth much when hot on their heels are open-weight models hosted by major cloud providers + Nvidia and cloud providers themselves building their own models.
Same thing I was going to type! I think they’re just testing these to see the reaction. Like a reverse auction while OpenAI watches in terror.
I want to see this go out at $800b. I do not believe even $1T will be supported by public markets unless they shrink the float — but that also comes with its own risks.
Most finance people I’m talking to think $200-400B is a fair price. That would be an amazing story in any normal situation but a total disaster here. However sane fundamental financials just don’t support a higher number.
Challenge is while you can play games with the float size to jack up the price as existing shareholders start selling you have to convince more and more people of a sky high price when all the fundamentals and the market pulling things in the opposite direction. Employees have been royally $&@ked in prior IPOs as that story plays out.
Means nothing, until the expiration of the lockup period (where employees cannot trade stock). That's typically 90 days, or longer.
At which point, the stock is well-traded on the open market and has found a "correct" market price, with no necessary relationship to the valuation at IPO.
1.1T is still insanely high. There is absolutely no reason why I can not use any of the OK models as long as I have reasonable hardware ti give a ok token rate.
A good rule of thumb is one a tech company attaints a >$1 trillion valuation, it tends to hold in that range; e.g. see Amazon, Appl, Google, Meta, Space-X, Nvidia, etc.
So many people 6-12 months ago predicted Space-X stock would crash and burn, yet the stock is still holding strong around $130 and a valuation over a trillion dollars.
> So many people 6-12 months ago predicted Space-X stock would crash and burn, yet the stock is still holding strong around $130 and a valuation over a trillion dollars.
That's simply because they're 1) installed in the government gravy-train, and 2) part of the AI bubble. Check again in a month or two, after AI valuations become more clear and Democrats start to publicly sketch out plans for what they're going to be doing to 2028 and beyond. They've already openly announced revenge for the half-billion in political spending, and they've haven't been talking at all about going after speech on twitter (which was a huge loser), but instead going after real stuff. Going after Space-X is an easy call.
> To build Microsoft-Decision-1, we post trained Qwen3.5-9B for fast, single-pass decision scoring and will soon rebase it on other models, including Microsoft AI (MAI) and OpenAI.
Cursor's models most famously are built on top of Kimi:
> Cursor’s vice president of developer education Lee Robinson soon acknowledged, “Yep, Composer 2 started from an open-source base!” But he said, “Only ~1/4 of the compute spent on the final model came from the base, the rest is from our training.” As a result, he said Composer 2’s performance on various benchmarks is “very different” from Kimi’s.
So yeah, Chinese open weight models are already in F500.
The major cloud providers have Chinese open weight models in their catalogs. If the model is running through your cloud provider on US soil, you're probably fine for most of corporate America.
F500 already outsource everything in the name of cost efficiency, but you're telling me, with the opportunity to cut token costs 40x, they're going to draw the line here?
If any non-US company can use US models, they can use Chinese. Chinese are way less hostile to all of us (regardless on which continent you are) than US is.
This whole fearmongering campaign may work well within US but once you cross the border it becomes funny parody in worst ways. When all this comes from current US its beyond ridiculous, I don't expect post-midterms behavior to be any less petty and dumb.
Space payload before Space-X and Starlink was 1000t and increased to 3000t because of starlink.
To get enough payload into space that the evaoluation is worth while, he really need to do this datacenter in space thing and sending enough hardware up there to match just one GPU based datacenter would take one Starship launch every single day.
Lets ignore that building Datacenter on earth is stll A LOT easier and cheaper to do and that we shouldn't vaporize non sustainable metals into our atmosphere alone for health reasons.
We have robots, AI (which is so bad that he prevers to rent out his AI hardware instea d of making more money with it directly), 3 Car models and something else combined in two companies with an overall value over 3T.
This setup is still higher valuated as plenty of big car manufacturers, robotics companies and a lot more other companies thrown into the mix together.
I talked clear numbers, you don't even try to argue.
I actually calculated and researched these numbers and while it seems that you are dismissive because of your ideological position? i saw and read Elon Musks claim about AI Datacenters in Space and him writing about a dyson sphere and calling it bullshit.
You are not evaluated as the 7/8th most valuable companie on the planet just because you build rockets and send up a little bit of payload. Thats why their P/E is 3 digits.
Its not if the absolut number of worlds payload is not that much and unironically Space-X made it A LOT cheaper. So they took the market, reduced the size of the market.
Over a 100 launches a year, most of them for starlink doesn't sell you the valuation the company currently has. And they have to keep sending up starlink satelites just for renewale of the existing ones. This blocks Space-X of having good margins on it. A normal telecom provider builds a fiber underground and a tower and that stuff just stays there for ages.
And yes you know what happened when people realized this is a thing? Yeah more companies are going into the market. Check out how many will aunch their first rockets in 2027.
can someone explain to me the point of the 'initial valuation', considering ultimately the price will settle lated thorugh trading to perceived value anyway ?
The initial buyers of the shares sold by Anthropic are institutional investors (mostly). That matching process is done by an investment bank they hire.
The trading price you see on the screen is - shortly afterwards the shares become publicly tradable on the exchange (or on all US equities exchanges, for US stocks). So in effect, the price shooting up on the public markets is actually bad for the company in that they might have left money on the table.
Sort of. IPOs are always undervalued because the entire block of shares gets purchased by the investment bank at the banks decided IPO price but then bears the risk of selling them at or above that price.
It's the price they're going to be offering to the general public to buy ahead of time, through their brokers. They get that price locked in, then once it starts trading it moves as it may. Usually the IPO buyers are not allowed to sell for some time
Validation is all about assets and future earning potential. Anthropic’s core assets are a model which is shown every day now to be only marginally better than something that’s free. So that asset appears to have very little value. Their future earning potential depends on folks not figuring out they can get a good-enough result for free, only paying for compute. And Anthropic doesn’t own compute, they rent it.
So the whole valuation is based on the idea that folks will happily pay this company a huge markup for inference over the actual cost of compute for a model whose valuation is rapidly tending to zero. Hence why potential investors are starting to freak out.
In competitive environments, such as most businesses, marginally better is enough to create large advantages.
I just don’t see a world where everyone is happy buying the second or third smartest model. The people using the smartest will just out compete them, even if it is marginally more expensive.
> Anthropic sells intelligence, and the most intelligent intelligence.
Most intelligent is a stretch, we don't know how the Chinese architectures would fare if they had the same compute availability as Anthropic and other US labs.
If we extrapolate and consider that they are behind due to artificial export restrictions, once they figure out how to build their own GPUs in a cost-effective/scaled-up way then the American labs' claim of having the best intelligence is at an enormous risk.
They already do a great job being resource constrained, imagine once this constraint is gone!
I mainly doubt the moat, not necessarily the usefulness. Why will customers stay with anthropic? Many cheaper models seem to catch up within 6-12 months, and there’s also other flagship providers that keep pace and compete on cost (oai, Google). Why should I believe that anthropic’s revenue is going to stay consistent, and not evaporate to other competitors?
If the justification for the valuation is "they sell intelligence", then their valuation should be $0. Because their product isn't the least bit intelligent, and routinely gets beaten by humans.
You can be as dismissive as you want but this doesn't reflect the reality.
I can easily spend a LOT less time on describing the same amount to claude than i would in a JIRA ticket and doing code review on the PR of Claude than talking to a person about it. Instead of 1-4 weeks Claude is done in an hour.
Its also already A LOT faster and better of analysing infrastructure issues.
Conservatives' undermining public education makes it inevitable that slop engines will surpass the intelligence of the average human, then double and treble it as the average human's intelligence continues to decline.
Probably also a huge problem but at least in the last 15 years of working, there was always at least 1 if not 2 people were i question how they were even in the industry to begin with and getting jobs.
I think people in this thread who claim local models are anywhere close to Claude are completely detached from reality. 200$ Ultra sub is the best value ever offered in history of subscriptions. I got 3 running and would pay at least 3x more for them before looking at alternatives. If your local models or some DeepSeek on cheap provider or whatever you install instead of Claude makes even 2-3 more mistakes per week you are already losing money assuming your time is worth average software dev salary.
I don't think it's remotely close either - Claude is that good.
The problem is that it's not how Anthropic plans to make money to justify 1.5T-2T valuation. That would require companies paying API prices on long term big contracts and that makes it a completely different value proposition.
The argument is not that they are as good, but that more models have reached the threshold for being useful. You do not need the best for most tasks these days. Freedom from closed model vendors also has value.
This is why token prices are falling and there are expectations for commoditization. In another year or two, you will have Opus 5.5 in much smaller models, assuming capability distillation remains as good as it has been from mega -> mini models.
Yeah I get it but I don't think it hurts Anthropic sub offerings. 200$ or 500$ or whatever it's going to be is low enough that even slight edge is worth paying for business use case. They will likely go into local proprietary models as well (we train our model on your data and you run it on your premises so you meet data retention regulation).
While I am a big fan of open models and I hope I can run local Opus 5.5 level model locally in the future I will still be using Opus 7.5 or whatever it's going to be for anything business related then. Again, the problem here is that they are not making enough from those subs and API is 20x+ more expensive.
Business do not get subscription plans as they are largely not available, certainly Anthropic is not offering them under the contractual obligations businesses want. It's all per-token.
Then you have to consider automations and user facing products. A quota based subscription is not enough or risks being shut off after too many tokens are used. These are the primary places open weights are disrupting closed, expensive models.
Eventually, token vendors will be commoditized and the average user will have an effectively unlimited plan, but we are not there yet.
Please explain to me why my "naive and maybe ignorant" understanding of this is wrong... The bigger the IPO price, the more they earn and the more probability is for you to loose money invested in it.
So say Anthropic, or OpenAi, or smartGummyBearsAI go IPO and the price is insanely high... and then spirals down to be 10% of the real price... they earn a lot of money and the public loose a lot of money.
IF smartGummyBearsAI goes 100% realistic price and you buy near the correct price, they earn their fair money and people get an investment.
IF smartGummyBearsAI says that they have technology to use the same machine that makes gummy bears print gpus and they are priced 10,000x their actual price the society just transfers a lot of wealth to them and they scam the society.
Why shouldn't we regulate and aim for the correct price?
The price isn't "set" from what I understand. The price is determined based on what institutional investors are willing to sign up to pay for it. So it's largely driven by the market (not on the actual stock market, just the market of interested parties pre-IPO). Those institutional investors do not want the scenario you describe any more than you or I do.
That's why it's dropping here. When the actual stock market tanked yesterday institutional investors were like - hey wait a minute this is starting to look too expensive.
Also if your 401(k) is in index funds you are safe. A company is not added to indices right away. If you are in actively managed funds or something like that you might have more exposure risk.
How do you determine the "correct" price? Something is worth what people will pay for it, no more and no less.
Markets have the advantage that many peoples' opinions are aggregated into the market price. This is far, far more accurate than having one person's opinion (an elected bureaucrat) set the price, particularly when they have no skin in the game.
If the IPO price is really 10x the public market price, that implies that the set of institutional investors collectively estimated the price 10x higher than the full set of retail investors. That's pretty unlikely: institutional investors usually do pretty extensive due diligence on IPOs, and you're already averaging out the statistical fluctuations by selling to a wide set of them. But if that does happen, it just means that institutions got fucked over to the benefit of retail (who gets a 90% discount on the same shares), which seems like something we should be celebrating.
using claude code with pro subscription and deepseek api. claude is project manager and deepseek is developer. combo seems pretty effective and efficient.
It keeps dropping. Was 2T, then 1.5T last week, now looking like they might be lucky to get 1T.
I wouldn’t be surprised if Anthropic delays the IPO in hopes this will all magically sort itself out but as every day passes it looks like they should have gone public this summer when things were still on the up.
Realistically the company is worth in the hundreds of billions but if it goes below 1.1T or so existing investors will need to start booking losses and that will trigger a shitstorm of other issues downstream.
This is all becoming a textbook case of folks know these assets aren’t worth what folks want them to be, but they can keep the game going so long as nobody actually marks the asset at its actual realistic valuation.
So that’s the game underway now. How low can they go without blowing up the whole AI ecosystem. If the valuations here drop next the AI bonds and private loans start to fail en-mass and the whole thing that built up this house of cards starts falling apart in reverse. Get ready for a bumpy ride folks.
I wouldn’t be surprised if they pull the plug on the IPO and go hide under a privately funded rock.
The second one of these AI companies gets a fair valuation and folks start booking losses the whole AI house of cards will come
crumbling down. The crazy thing is if they drop below around 1T some existing investors will be booking losses.
A realistic valuation is likely in the hundreds of billions, which would be a total disaster and thus they’re going to do everything possible to avoid that.
literally what is the upside? They're supposedly worth as much as meta despite having nowhere near the customer base or years of work? How much more valuation do they have to trick people into giving them before investors get a return?
Anthropic hit 9 billion ARR in Dec 2025 and 80 billion in Oct 2026.
Meta hit 9 billion in 2014 and 80 billion in 2020. It is now at 250 billion.
At this growth rate Anthropic will be bigger than Meta already next year.
Caveats: It's hard to believe that this growth rate would continue. And Meta's numbers above are trailing 12 months, while Anthropic's are annualized run rate.
> Anthropic hit 9 billion ARR in Dec 2025 and 80 billion in Oct 2026
This is a bit misleading and based on projections of prior reported months. Growth has already slowed, Q3 in dollars was less than Q2. The timing aligns largely with the growth (since mid-summer) of open weight models.
Since they actually have paying customers, it's a completely different sort of thing than for Meta. The question is instead whether they can keep at the frontier, whether the compute they've signed up for is good enough, etc.
I think this is down to how investors think about how to value a piece of a company. I think about it almost the same way as I think about sports gambling.
There's a value proposition in which direction the stock is going to go and how long it's going to take to get there. It only makes sense to be invested in a stock if you think it's going to be worth more later than it is now and the amount of time your money is tied up in this stock is commensurate with the profit you may get as a result.
Investors work on limited information about companies, and having a track record can influence a buying decision one way or another. It may matter to some investors and not to others. For many, it's weighed against the price of the stock. Would you bet $100 on a company with a short track record? What about $10,000, $100,000, or more? Generally there's a point where people drop out of buying more of a stock and say that they can't justify the risk given the information they have. That's the expected value that Anthropic will receive from that investor.
Anthropic needs to worry about what that number is for a huge amount of individual investors and how to maximize the amount of money put into the cost of their stocks, and there's very little chance that they can attract the biggest pool of investor money without some of those people considering their longevity vs a more established company like Meta. It may not matter to all investors, but the fact that it will matter to some makes it a factor in the total valuation.
Possibly speaks to its moat being weak, if a dedicated team from scratch can duplicate their efforts and get to where they are today in 2 years (but not where they will be).
Because building a stable revenue base takes time. We've had many flash-in-a-pan type product but things which still make money 10 years later are rare.
My point is that Anthropic is not special. Neither is OpenAI. They are just really good at marketing.
Job displacement is real, but based on real world usage it would owe just as much to Grok, Gemini, DeepSeek, Kimi, and whatever else that is close to frontier, but cheaper.
If in a couple years they're losing their jobs because of Anthropic/OpenAI/10+ other enterprise token resellers using open models then their value proposition is much weaker even if people are still losing jobs.
Yes could be but we don't know yet how big the overall market will be.
If you have in every single industry AI tokens, it will replace wages. In USA alone thats 15 Trillion. If its just 10% the USA market alone are 1.5T without the value added.
That is the only future where the valuation makes sense. But I'm afraid to think about how a society where that[1] happens will behave. My guess is that it would become _highly_ unstable.
[1] Say >10% of people of employable age suddenly becomes superfluous with not much prospect of finding new jobs (since AI eats those too).
If you give me ANY idea of why you think my argument is stupid, we could actually discuss it?
My company is building an agentic layer in its product, google and the other hyperscalers do that too.
I already want to get rid of one person in my team because he was slow before but was able to contribute something, now I want to have his position as AI tokens.
you think 10% of the entire usa market is gonna go to anthropic and I'm supposed to somehow find a way with words for you to understand how stupid that sounds?
The OP is snarky but he makes a valid point: if Claude is so intelligent, why can't it just "do something" to get whatever valuation its owners want? Because that's what we've been told, this is AGI, no it's ASI, it will kill us all.
Currently the progress is real and there is no ceiling in sight. If you just follow the line, there will be / should be your scenario coming up sooner than later.
But no one knows if there is a ceiling or progress might slow down massivly.
Is warning people so wrong if you interpolate the current progress and just look ahead a few years? Might just be as stupid as some North Korean/Russian hacker throwing some abliberated model against some computer system and having some cascading effect on water or energy grid.
For plenty of people AI already took their jobs, jobs they liked and enjoyed doing.
The future is not local models because by the time you think of what to pipe through the LLM, the super intelligent AGI owned by Anthropic or OpenAI (but not both) would have already mapped it out, generated it and had it in production. You can, however, still pursue living as a hobby.
potential buyers are discussing about a $1.2 trillion to $1.3 trillion valuation range for the company, according to people familiar with the matter. Some are weighing a value as low as $1.1 trillion
The Facebook valuation some 10 years ago was like… 300 million? I mean, I get it, but jeez, when I see headlines like this “as low as(!)”, I need to pinch myself.
Well they’re struggling to raise cash via new rounds, having slashed both the raise amount and valuation themselves. So in that situation they may go bust before having a chance to go public.
OAI is in the middle of another funding round at $1.25T. If Anthropic with allegedly better financials can't get above $1.1T then why would anyone buy OAI at $1.25T
Or then any numbers they would have publicly presented with even previous scrutiny would have been so bad no one would have bought them at price they want.
My hunch is that it’s better to be late in this scenario because now your moves are under public scrutiny or at least neck-deep. People can pull out, it could slow them down and at the end look worse.
As a privately held company, it functions under different dynamics that in the environment we’re seeing feels better to stay private for longer.
People seem unnecessarily deluded here in a supposed hacker forum. If they are really created actual digital intelligence smarter than every human and figure it out how to control it, then 1T dollars means this is the most undervalued business in the entire world. Maybe by 100x or 1000x. You have billions of super genius slaves in data centers that do whatever you want, then if a single human developer can earn $100k/year pessimisticly let alone a super genius one. Multiply that by a billion and do the math at 80% gross margins
…and what if they haven’t? Given that we haven’t seen evidence of one so far.
You’re also missing the competition aspect. What if other companies have their own super genius slaves to sell at a cheaper price? What if open models allow anyone to have their own moderately genius slave at far less expense?
Yours is a reasonable estimate at the _value_ of their product. They will only be able to capture that value if their competition cannot catch up AND they can reduce spend on R&D. I personally do not believe they’ll defeat distillation.
Every market cap is a prediction of future profits. If they had trillions in profits every year already, then the price wouldn't be 1.1T, it would be 100T and we wouldn't have this discussion. 1T is a option to buy it, and it's probably a good bet even if there's a 10% chance that it happens.
What if they have billions of super genius slaves, but Chinese companies have just slightly less super genius (and catching up) slaves for 1/4th of the price, while all prices are already dropping?
I see nothing delusional in speculating that that’s a very big if.
And if they haven’t created an artificial superintelligence then they start to look more like just another tech startup with no moat[1] and a terrible balance sheet.
[1] I don’t know about you, but I switch model providers multiple times per day.
Every market cap is a prediction of expected future profits. In their mind, I'd say that's what the whole "race" is that they are so desperate to win. The race to win is the race (in their minds) to get such powerful beings to get a permanent lead with ASI. Getting to powerful enough beings to wipe out all the competition and become emperors of this new world.
As high as $1.1T. Where is the supposed growth going to come from?
Claude has been pushed by force and coercive marketing into most software organizations and many open source projects. In 2026 the companies buying the narrative were exuberant and told employees to tokenmaxx.
This will stop in 2027 because companies will find out that nothing has improved. Nothing at all. They'll evaluate the expenditures.
So the market is saturated and usage is going down. Anthropic should be valued at $200 billion at best.
When you sell people 6 quarters for $1 they tend to buy a lot. Useage is a poor metric here given the capital intensive nature of this business. What matters is can they do this in a way that makes money to the degree that justifies the valuation, and there they have not yet demonstrated this is possible.
The internal metric that they are using is they get $2 for every $1 in inference. So, if you discount model development costs, they are selling the tokens for more than it costs them to generate a token. And this is before they start integrating their advertisment networks into the model. They see something as high as $8-11 per $1 of inference once purchasing and advertisments are incorporated.
The “but ads” logic fails basic sense checks. There isn’t enough ad money to go around to make up the gaps. With the big ad players like Google, Meta, and Amazon all shutting Anthropic out of their ecosystem there’s very little evidence that Anthropic can steal away enough of that as revenue to save itself.
In the “but they’ll slow down training and so will become more profitable” thinking is in big trouble given the pace of open models. They’ll then be even less competitive with free options and so even less reason to buy from Anthropic.
There's not enough ad money? What? lol. there will always be ad money as long as a click costs below my profit. Every single one of those companies you listed was excluded from the Ad system of the ones before it. They built their own.
For sure, models definitely depreciate but so does CapEx. but you don't include CapEx, you have an amortization schedule over it's useful lifetime. However, when you start mixing CapEx and model depreciation, you get an interesting curve. As CapEx increases, model development is spread across multiple generations across that same CapEx. So, assuming you've amortized your CapEx, successive model generation eventually has a horizontal assimtote at OpEx and depreciation.
The real question is "Are frontier training costs growing faster than the productive lifetime value of the GPU infrastructure used to create those models?" Which is a hard question to answer and is something we'll find out it early 2030.
Also, Anthropic is known to lie through their teeth to make themselves look better, so any claims that they make a profit on inference should be treated as false until proven otherwise.
Any “non-GAAP” measure of “profit” should generally be viewed with a giant red flag.
In nearly all cases these metrics just say “if you ignore all the reasons why we’re horribly unprofitable, we’re profitable.” WeWork’s “community adjusted EBITDA” was a textbook example of this.
I introduced it at our admittedly small company not because of coercive marketing or force of any kind. I did it because rolling new stuff out takes days instead of months. Not only that but we can make products above our previous level.
Unsafe, comprehension debt, vibecoding, slop? Have you seen actual lowly developers’ code? AI is a vast, vast improvement on all relevant axes if you ask me.
If you think this is going to slow down I have no idea what to tell you.
Good drop. Should be less than a trillion amid current midterm fallout. The chances of more AI regulations are coming.
Colossus 1 & Colossus 2 are located on Tennessee and Mississippi, and Anthropic rents both of them earlier this year.
Now that Democrats are surely going after Trump and Elon Musk and the sidekicks, and both Tennessee and Mississippi turned blue as well, this directly impacts the way Anthropic to source cheap compute and data in the future due to uncertainty around the continuation of SpaceX data centers. And other data center buildouts are likely to be paused for things like power grid connection and environmental impacts and land purpose evaluation, the compute capacity constraints are likely going south, too.
Worse, the proposals around superintelligence is exactly what Anthropic is looking for: slow the fuck down. Now they got the killshot they long desired for, it obviously makes a dent in the capability and capacity of their AI, hence a drop in valuation. I think $800-900B is somewhere I would say okay, and $600-700B is really where it is fair, considering the long term outlook for Anthropic is looming with a greater AI massacre caused by political factors.
This is a "not like that" situation. They wanted "AI safety" legislation guaranteeing them rent money. What they're likely getting is legislation to slow down data center buildout and make them more environmentally sane (more expensive upfront cost).
And open models are coming whether they like it or not.
Oh, and if the Dems take OpenAI and Anthropic's AI safety rhetoric at face value, nationalization and other forms of substantial government interference is on the table. A predictable outcome of irresponsible fear mongering!
I don't see how Anthro is worth 15x( Gemini estimates based off WSJ source )as much as Deepseek( 76 billion) .
If anything is going to pull the rug out, it's going to be a model from Deepseek or another Chinese lab. I'll admit I'm very uncomfortable with so much of our economy riding on AI hype.
To be fair to China, if AI is really this revolutionary technology, they need to be able to have home grown peers.
Every nation ( or bloc of nations) needs alternatives to American tech.
Anthropic intends to rule humanity with its ASI. Perhaps this is the optimal outcome, as there are many problems AI could fix that it wouldn't be able to as efficiently without eliminating the human bottlenecks which caused those issues, but I fear the more financially powerful it gets, the fewer checks there will be on any misalignment.
As opposed to the CCP, which is essentially what the Chinese model companies map to, who are only interested in freedom and liberty for humanity right?
Sure, this is a small anecdotal datapoint, but since I've gotten Qwen3.8-Flash-Next running well on my Strix Halo box my Claude Pro subscription use has dropped off a cliff. I'm only at 18% of quota usage this week - this being Friday I would've been at 80% at least in the past. This is just an astonishingly good model that runs locally and quite well on my PC. I do not plan to renew my Claude Pro subscription when it comes up in a couple of months.
I really hope the future is local models, because the idea that all companies will need to buy tokens from a handfull of tech companies to run their businesses or make their products is awful.
I've been ripped (on this site) for making this prediction. While the frontier (i.e. The toughest problems) will always push the limits of available computing power, what typical people do with AI won't always be on the frontier, nor will what appears on the frontier necessarily be lucrative. The frontier used to be playing Chess, Atari, and Go. These were challenging and really advanced AI, but they didn't make a lot of money. Now a phone can do these things.
You're an example of this process. You used to require a frontier model running on a data centre. Algorithms improved and, perhaps, your hardware as well, but your requirements didn't go up. Now you can do what you need to do with local hardware. You're one AI user who might not need to send money to Anthropic going forward unless you find a new frontier application that interests you.
This is why I'm a little bearish when it comes to the valuation of the AI giants. Yes, there are some extremely lucrative things on the frontier right now but, in the future, people may shift a lot of what they're doing now to local hardware while new things on the frontier may not produce the kind of profits that will pay back trillions in debt. AI companies have bet big on themselves being able to constantly and perpetually find new applications, and also that those new applications will make money.
I agree with you that if the trends of 2025-2026 continue into 2027-2028 for open weight models and local inference, 120b param open models will be more than good enough for 80% of today's use cases. If so how can you justify OpenAI and Anthropic and the AI buildout?
> but your requirements didn't go up.
I am definitely building more and more ambitious projects as models improve. Both personally and at work
I suppose I should have said they just didn't go up faster than the capabilities of local models are improving. I suspect local model/hardware capability will scale up faster than most people's needs when it comes to coding. That's likley why Sam Altman has naughty designs on your RAM.
Open weight/local models are important but they do not compete with frontier models like Fable/Astra. They aren’t even close.
> the idea that all companies will need to buy tokens from a handfull of tech companies to run their businesses or make their products is awful.
To the leadership of these tech companies that actually is the dream. Sam Altman said the quiet part aloud recently: "From a global allocation perspective, actually, it's quite terrible to have everybody have a bunch of the world's limited supply of DRAM sitting in their own computer, and being used maybe 5% of the time."
And it certainly fits into the neofeudal worldview of other relevant Big Tech leaders. Be it Peter "I no longer believe that freedom and democracy are compatible" Thiel, or unnamed bigwigs discussing if shock collars are effective to prevent their staff from revolting after a doomsday scenario [2].
We shall own nothing and rent everything, that is the mindset these people believe.
(And if you ask me: the AI IPOs are part of that strategy. Assuming OpenAI and Anthropic both IPO, they will get part of every American's and most other countries' pension funds... and when they collapse, a lot of that value will be gone. The initial investors into SpaceX, OAI and Anthropic all will have made their money and put it aside, and we will be left holding bags)
[1] https://www.ign.com/articles/openai-ceo-sam-altman-thinks-it...
[2] https://www.rte.ie/news/business/2025/0511/1512086-doomsday-...
How is that saying the quiet part out loud?
Easy: it's an open secret (the quiet part) that the rich fucks want us to rent everything and us own nothing. Real estate was the first to go unaffordable without taking up debt. Then came cars. Then phones. And now it's computers. Hell I've seen kitchen appliances for rent.
And now this mf'er (or to be more precise, alleged sf'er [1]) goes on the record to say it's bad if we the people have RAM.
Fuck him, fuck everything he stands for.
[1] https://www.theguardian.com/technology/2026/oct/20/judge-dis...
Maybe so, but unless you imagine a vast conspiracy of rich fucks pulling the strings of hardware vendors and/or governments worldwide, the former will still want to sell to a larger market even if some of their largest customers would prefer otherwise.
> the former will still want to sell to a larger market even if some of their largest customers would prefer otherwise.
Oh hell no. The hardware vendors want to make as much money as possible, with Micron going as far as to fully exit the consumer market entirely so they can suck as much money out of the AI market as possible.
How that one can fly by antitrust enforcement is beyond me.
[1] https://investors.micron.com/news/press-release/2025/Micron-...
He literally said we'll own nothing and be happy
he said nothing about being happy
Altman was simply making an accurate observation about DRAM, in the context of a conversation about DRAM shortages and precisely why it is difficult for DRAM manufacturers to respond to the market price incentive to increase DRAM production. Even if you think it is bad that DRAM is currently expensive, as I do, that doesn't change the underlying fact about the world that Altman's comment reflects.
Thiel's comment is also pro-individual-ownership - in context he was talking about the danger that in a mass democracy the people as a whole would vote to have the government restrict the freedom and confiscate the property of successful people (particularly prescient in light of the recent California attempt at a billionaire wealth tax). RAM being expensive is better than RAM being rationed by a democratically-elected government so no one person has too much, as judged by the voting masses.
> Altman was simply making an accurate observation about DRAM
no, he was lying by omission
"it's quite terrible" only if you ignore all the problems with plans they promote and measure it only in metrics Altman wants to measure, as they support his plans
For start, Altman would charge as much as possible to optimize profits, so very significant part of efficiency increase would become profits of "Open"AI.
As another obvious problem: handing so much power to a specific company has own costs.
There are so many problems here with this specific part - starting from their power dictating how can you use it, access to your data. And that part is getting worse for anyone not planning to obey all decisions coming from USA. We have already seen export controls being invoked.
But wider, if you think that outsourcing thinking to a specific SaaS has no costs, beyond current subscription costs, then I have several thousands of bridges for sale.
> Altman was simply making an accurate observation about DRAM
Which might be true, I'm willing to accept that, but it does not change the fact his message was incredibly tone deaf.
Cost of living in general is through the roof for years now, many people haven't had the financial ability to meaningfully upgrade their computers for just about the same timeframe - first Bitcoin drove CPU prices through the roof, then GPU prices, once that died down due to ASICs there were a few fads that "mined" RAM, HDD and SSD, and in parallel to the latter Covid hit, disrupting supply chains. After Covid was somewhat over, the Russians invaded Ukraine, then we got the various MENA conflicts after Oct 7th, and now the Iran war is wrecking everything - and ever since AI entered the picture somewhat three years ago, first GPUs and now everything else has exploded in price. Entry level smartphones have all but disappeared. And don't forget our existing computers got less and less useful because first Microsoft pushed the garbage called Windows 11 onto everyone it could and second, everyone and their dog went for Electron resource hog apps.
For Altman to now say "your unused RAM could be of much better use in our servers" is like a giant fat fucking middle finger, telling us to sacrifice even more of our needs just for the opportunity of him and his fellow rich bastard friends to make more money, all while he and his fellow rich bastard friends are executing the setup of a financial crisis that will make 2007 look fucking tame.
> Thiel's comment is also pro-individual-ownership - in context he was talking about the danger that in a mass democracy the people as a whole would vote to have the government restrict the freedom and confiscate the property of successful people (particularly prescient in light of the recent California attempt at a billionaire wealth tax).
Oh for fucks sake. Taxes aren't theft, they provide for the environment that makes these rich fucks able to earn money in the first place. Thiel is worth 38 billion dollars. Even if you'd tax the guy 50% of his wealth, he'd still have more than one single human and their descendants can ever feasibly gamble away.
> RAM being expensive is better than RAM being rationed by a democratically-elected government so no one person has too much, as judged by the voting masses.
At least in a rationed system that allocates supply fairly, ordinary people wouldn't have to run with barely working hardware because they can't afford to buy something new and modern for once as all the new stuff is hoarded by the rich like a LotR dragon hoards his gold pile.
And as evidenced by this week's elections - the people are fed up beyond belief. Either we the people get some relief, the sooner the better, or the situation may indeed devolve into open riots.
> At least in a rationed system that allocates supply fairly
if you assume crazy fairy land that is not going to happen, then you can assume cheap and plentiful RAM for better results
Successful people are not successful without keeping unsuccessful people from success; inflation since Millennials were born is almost 300% and taxes have been slashed; takes 800k/yr now to have the buying power of $200k in the 1980s
Thiel has benefited from wealth being wielded by rich media owners to convince the masses decades long trends in deflation of buying power through inflation of costs are good for them
With such a system giving him a boost he was successful before Millennials and GenZ were adults or even born
Altman came along during ZIRP; being successful when the government is intentionally printing money is not really success
I think we will see good enough local models, excellent cloud models, and then simpler focused models embedded on silicon running on every small device.
I find it hard to believe that at least a robust hybrid model won't be. There's just too much money to be made selling hardware at relatively high margins to customers with limited market power to keep hardware vendors out of consumer and small business markets indefinitely.
> the idea that all companies will need to buy tokens from a handfull of tech companies to run their businesses or make their products
Did you somehow become incapable of doing things without models? Most of us have not.
Would be incredibly inefficient and non-competitive to operate without LLMs. It would be an exaggeration to say they are as necessary as the internet or electricity, but it feels kind of close. Can't operate a software business without them.
Imagine if only KB Home and D.R. Horton had access to power tools while all the other homebuilders big and small had to use hand saws and screwdrivers. Would they be incapable of building houses without power tools? No. Will they be relatively screwed as companies? Yes.
You're changing the goalposts.
The above comment said:
> need to buy tokens from a handfull of tech companies to run their businesses or make their products
That's clearly indicating necessity.
That would indeed be concerning if it was the case, but it's not.
So now the case is just "you can pay more money to do things faster".
And guess what? That's been the case for centuries if not millennia. Nothing special or concerning at all.
And so your analogy is misleading. You're trying to imply that only a tiny number (2-3?) of companies make the tools, and nobody else does. You either pay those companies, or you don't get them.
That's just false. There are hundreds of companies that will sell you tokens - just not for frontier models. The correct analogy is how Samsung, Google, and Apple make the best phones in the world, so you can buy from one of them - or you can buy a still-good but not bleeding-edge phone from another vendor. And everyone is fine with that.
>You're trying to imply that only a tiny number (2-3?) of companies make the tools, and nobody else does. You either pay those companies, or you don't get them.
Correct. I have to pay Anthropic to use Claude models. Open-weight models are significantly worse than frontier cloud models. Their cell phone-equivalent would call your boss when you try to call your spouse, randomly delete photos you took with it, fail to send text messages half the time, power off when it still has 50% battery.
> Their cell phone-equivalent would call your boss when you try to call your spouse, randomly delete photos you took with it, fail to send text messages half the time, power off when it still has 50% battery.
Still incorrect.
Models are always stochastic. They always have a chance of failure. That includes the most expensive frontier models.
You're making an invalid comparison to cellphones by claiming that there are cellphones that spontaneously and regularly do those things (which is wrong) vs flagship phones that don't (which is extremely wrong).
Either that, or you're claiming that the difference between frontier models and open models is the difference between actual flagship phones and these hypothetical ultra-buggy phones. Which is also wrong, because of the above fact that frontier models and open models are both stochastic, and because on many public benchmarks (e.g. Artificial Analysis) the gaps between the best frontier models and the best open models are far smaller than your analogy claims.
The fact is that, while frontier models are better than open models, you can still do a lot of work with open models. You're in no way "screwed" like you're trying to imply either with the incorrect phone analogy or the incorrect power tools analogy above.
This is exactly what I'm seeing all over the place. I'd like to say it's not true for me, but I'd probably be lying a little bit to say so. It is causing some atrophy in my skillset. One could argue that I'm gaining something different through LLM usage like "getting good at asking the model questions" or something, but it doesn't frankly feel like a skill which is all that useful on a going forward basis.
[dead]
Why only a handfull? There can be (maybe already are) hundreds of LLM API providers.
There’s no way to make that economic.
There are already 80 providers on Openrouter.
And that isn't sustainable. The market will collapse into a small number of players, in exactly the same way that literally every other market has.
Are you sure you're not mixing up companies training their own frontier-ish models with companies providing hosted LLM API? For a close comparison, VPS hosting provider market did not collapse to a small number of players - few companies dominate the market (AWS, GCP), but there are thousands (or tens of thousands) of medium and small players available.
Interesting, I'm also comparing 3.8-flash-next (4-bit quant, llama for inference & pi as harness, on a 128gb strix halo machine) with opus 5.5 in claude code and while qwen is a good and fast model, it's nowhere near for coding, even haiku 5.5 feels better.
I could see it used as a sort of "sub-haiku" for easy stuff to leave more tokens for the bigger claude models, but I won't be cancelling my Claude sub any time soon.
It's obviously nowhere near opus 5.5, but it's crossed the same threshold opus 4.5 did: it's good enough. If all model-as-a-service companies shut down tomorrow, I'd be happy with qwen 3.8 flash next q4 running in my strix halo
The problem here is we're not talking ... I don't know ... hamburgers. If I like McDonald's better than Burger King, it's not going to stop me from buying a Burger King burger (if I'm hungry and there's no McDonald's around). One burger may be a bit better, but I can reasonably substitute one for the other.
Coding AIs aren't burgers. Even if a local LLM is as good as Opus 4.5 (when you sink several thousand dollars into hardware ... and I'm skeptical even then) ... the actual difference between the models is huge.
Let's say I use the $200/month Claude (I don't, I use a cheaper plan, but stay with me). If I switch to a Opus 4.5 equivalent local LLM, and it makes even a single production issue that 5.5 wouldn't, that mistake will cost thousands of dollars.
That "cheaper" model isn't really cheaper.
> Let's say I use the $200/month Claude (I don't, I use a cheaper plan, but stay with me). If I switch to a Opus 4.5 equivalent local LLM, and it makes even a single production issue that 5.5 wouldn't, that mistake will cost thousands of dollars.
You are assuming that Opus would always be superior; the cloud AI providers can be very variable. They can be very useful for months, but during certain times their quality can drop off a cliff. People online speculate that the cloud providers drop quality during this time when they are training a new model or the service is over-subscribed. You do not have that issue with a local model.
That issue of quality of service is enough many businesses may prefer a locally run model because it will be more consistent.
Indeed, sometimes qwen is superior (even at q3), yesterday i asked both what a line from the logs meant and opus said "oh its just your dns server falling back from eodns to udp no worries" and qwen correctly recognized the ip address as docker subnet and said dont send dns there its probably off
It's way better than Opus 4.5. See my stress test elsewhere in this thread (C -> microcode compiler with optimizations written in OCaml, generates microcode + Verilog output) - it's doing better than Opus 4.6/4.8 were when I was using them to code the original C version of this compiler early in the year. It's come up with optimizations that Claude didn't come up with back then.
Did opus have an oracle for the original compiler? If not, then it's expected that qwen would do a better job.
As is known, LLMs are amazing when verification is cheap, and much less useful when it's expensive.
> Did opus have an oracle for the original compiler?
It actually did. There was an older yacc-based implementation that had a lot of architectural limitations. We decided that it was better to start over with a better architecture, but we did have that older one around to compare generated microcode initially.
Ok, that's interesting. My prior would have been that if Opus didn't have an oracle, then there's still a pretty large gap.
But if, as you say, both had an oracle then this is a really interesting outcome.
To be fair, I think the differences between frontier and open models is pretty small (I've been swapping back between Claude and GLM over the last week).
The one thing that I notice is that GLM keeps trying to blame external factors for any non-matches (we're rebuilding an external tool for internal use), whereas Claude adheres to my prompt much better (i tell it to only blame these factors if it's ruled out everything else).
I'm not convinced that change is worth 3.5-5x cost differentials though (I only used Claude because we have the small business package, which is way cheaper than API rates).
> Let's say I use the $200/month Claude (I don't, I use a cheaper plan, but stay with me). If I switch to a Opus 4.5 equivalent local LLM, and it makes even a single production issue that 5.5 wouldn't, that mistake will cost thousands of dollars.
There are a few points I'd like to make here.
First, the frontier is jagged. There are errors that Qwen3.8 Flash Next make that Opus 5.5 doesn't make, but the reverese is also true: Opus 5.5 makes errors that Qwen3.8 Flash Next doesn't make. The latter will likely be fewer, but they won't be zero. Outside of really large differences in capabilities, there is no total ordering among models, only a partial one.
As an aside, do you always run Opus 5.5 at max reasoning in order to minimize errors, too? I suggest most people don't, because they have their own intuition for good enough vs. cost.
Second, Opus 5.5's strength are autonomous tasks. For interactive, HITL work, it's not a particularly great choice. This includes its preferring complex Bash commands over tool use and inscrutable Python scripts, but also often not even explaining what it does. It is not a great choice to use as a virtual pair programmer, because following along what it does and maintaining a mental model is harder than with some other models (though probably because of benchmarks that prioritize autonomous work, these and related problems have infected other models, too). You can address that to an extent with CLAUDE.md instructions, but even that has imperfect results.
Finally, if you rely on a model not to make mistakes in order to ensure correctness, your software development process is fundamentally broken. You need to assume that any model will make at least some errors and need a development process that accounts for that. The problem here is not necessarily which model creates fewer errors, but which model creates code where errors are easier to discover. Ironically, more powerful models may create more complex code, making that harder.
You can see an analogous effect with reasoning effort in FrontierBench 2.0 [1]:
"FrontierCode is one of the few coding benchmarks where writing additional (correct) code doesn’t necessarily raise a model’s score. This leads to an interesting phenomenon: models can score lower at higher reasoning effort. Because this non-monotonicity often matches user sentiment but is hard to see in other evals, we consider it a useful feature of the benchmark."
A example:
"This task comes from PostHog’s open-source Android SDK, which apps use to record analytics events. When the device is offline, the SDK writes events to disk and sends them the next time the app starts. The task is to fix a startup bug in which two modules send data at the same time: the live queue sends new events while a legacy integration, PostHogSendCachedEventsIntegration.kt, sends the cached ones. The two race, and events reach the server out of order.
"At low effort, GPT-6 Astra moves cache loading into PostHogQueue.kt, removes the sending logic from the legacy integration, and leaves the SDK’s public API untouched, which produces a passing solution.
"At max effort, the model often deletes PostHogSendCachedEventsIntegration.kt outright and spreads its responsibilities across other legacy files. This also fixes the race and passes most of the hidden behavior tests. But the deleted class is part of the SDK’s public posthog.api surface, so removing it introduces an unnecessary breaking change for every app that references it. Scope blockers, which deterministically check that the model hasn’t modified certain files, catch this change, and the agent fails the task."
[1] https://cognition.com/blog/frontier-code-2.0
> Let's say I use the $200/month Claude (I don't, I use a cheaper plan, but stay with me). If I switch to a Opus 4.5 equivalent local LLM, and it makes even a single production issue that 5.5 wouldn't, that mistake will cost thousands of dollars.
What are you talking about, you're fully dependent on an LLM (Opus or whatever) not making a mistake or else you lose thousands of dollars, in production? How is that okayed?
"It" makes a production issue? Are you not involved in this process?
Yea your math doesn't math.
Try it before you go super far into the world of this "must suck".
Most people aren't going full vibe code. If you stear these models, they are almost identical. Seriously. The major score changes in a lot of the absolute latest models are just models doing more things (often when they shouldn't). Having a slightly lower scoring model is actually better in some ways. I digress.
If your model can make a production mistake you are using it wrong. Sorry not sorry.
There are other benefits. Your data is private, it's not sold. Your privacy is kept, that's priceless. No one can degrade your workflows. Uptime is you are online whenever you are online not 3 9s. You also won't get refused asking basic security questions.
Your math doesn't math for a lot of people. It might for you, but I would encourage you to think a little more.
> it's crossed the same threshold opus 4.5 did: it's good enough
Agreed on that, and actually I'd put it somewhere around opus 4.5 from what I've seen so far.
>> It's obviously nowhere near opus 5.5, but it's crossed the same threshold opus 4.5 did: it's good enough.
Good enough may be sufficient for personal projects, but in a competitive landscape where you're using a "good enough" model and your rivals are using a frontier model like Opus 5.5, you will be at a significant disadvantage.
Not necessarily. If yor process accounts for "all models make mistakes", you are probably limiting how autonomous the model can be, and then there's not much tangible difference between models beyond that threshold.
In the competitive landscape, cost and not-getting-locked-down are also factors up for consideration.
> If all model-as-a-service companies shut down tomorrow
Is this the nerd version of "doomsday prepping"?
Given the vast amounts of constant external investment money they need to keep operations, and how little moat (if any, besides influx of investor money) they have, I'd argue it's not that far fetched. Also, they are all american, so as a non-american, I also don't see it that unlikely to suddenly not being able to access these model-as-a-service services
VPN? I'm not in US either but there's not a single US-only online service that I can't access.
If the planning-for-a-$2T-IPO company shuts down overnight, unexpectedly, you might want to do some actual doomsday prepping (like, canned food).
First time?
Eh, it wouldn't be the first unicorn to go up in smoke just after the IPO
Maybe the biggest ever, though?
Here's a stress test of sorts that I've had it working on the last couple of days: We have a C -> microcode compiler written in C (most of that was work done 'back' in the Opus 4.6/4.8 era). I gave it the docs for that compiler and asked it to implement it in OCaml (since it's not very common, but also is a good implementation language for compilers). It can use the first compiler as a golden to compare against. It's essentially: C language parser -> AST -> IR -> [optimization passes] -> microcode (+ some verilog code). It's basically at the point now where the original compiler has been. Not only does this generate microcode it also has to generate some Verilog code that loads and runs the microcode as well as run simulations in Verilator to confirm functionality. I asked it to come up with some new optimizations that the original compiler doesn't have and it came up with a couple that in some cases reduce microcode lines by up to 30% (less microcode in this machine means more performant). These are ideas that Claude didn't come up with when it was working on the original compiler back in the 4.6/4.8 era(and I recall prompting it several times asking it if there might be other optimizations we could perform).
So all in all, I put it at Opus 4.8 level, at least. And in some ways it's exceeding 4.8.
I used Ollama with Qwen 3.8 27b on my MBP for a flight to try some experimentation and found that it was shockingly good at one of the more complex PDF processing tasks.
Like seat-of-the-pants good as Opus 5.5 but a bit slower and required just a bit more prompting to get what I wanted (which is fine for an app pipeline).
And only slower because hardware
I am in the same boat. Sun servers were still the best that money could buy when they lost to the Linux revolution. Linux was _good enough_ and destroyed their margins.
I would have preferred to have open Solaris flourish along with Linux.
What are ypu working on/building that Qwen3.8 is good enough? I've tried the latest open weights models but none of them really work well for my use case (building an all in one CAD + CFD + CAM + DFM optimization). Qwen in particular has a bad habit of flipping XYZ coordinates so the STEP files exported end up being a mess.
That jives with my own datapoints. Of a circle of around 20 very active devs Anthropic went from being the must-have tool to on the chopping block.
Around 50% of that group says they plan to cancel their Claude subscriptions now that their local or open model setup is performing “good enough.”
I don't believe this one bit. There is no way local models are good enough except for the most basic of coding.
If you actually know the code and can give guidance, direction, and corrections they absolutely are.
If you are just vibing and managing multiple agents and reviewing at a higher level like architecture, you are right.
You haven't spent much time with the latest self hosted models then. Give it a good harness and instructions, access to a web search engine and ability to iterate on problems and they do a good enough job that I'm happy with the results.
Don’t know what to tell you. It’s happening.
But there's just no way.
Why is it unbelievable? I still remember how to program and arch new features and don't need claude to tell me how to do it. It is much more enjoyable to do a little work on your project and offload the grunt work to a weaker model then just being a highly paid merge pr button pusher for whatever opus decides.
This is what I've been doing with qwen3.8-flash-next. I haven't fully let go of the reins yet but I do like having a junior who can implement my new features as I have designed them
What do you consider a local model? I'm running DeepSeek Flash version 4.1 and it's outstanding.
Local models: if they have the hardware to run it.
Open models: we're already seeing this at some scale. https://news.ycombinator.com/item?id=50350004. I do expect to see more of this.
I hate to break it to you but most people do most people do basic things most of the time
[dead]
These companies have no moat at all.
Their moat is their research capability. It's a combination of proprietary internal technology and models and infrastructure, people, organizational structure, and culture. As long as they maintain the best and fastest research capability, they have a durable advantage, at least for the time being. If we start to saturate all economically useful outcomes that can be reached via greater research productivity attainable to us, then this moat will start to fall apart. It could also fall apart if safety issues force a premature halt or major slowdown to further research.
This assumes any research gains last. In practice it seems like none of the frontier companies can keep much of their "secret sauce" secret for very long.
If company X has a better model than company Y, do you use it? Well, that depends: will Y take two years to catch up? If so yeah, you buy X.
But if Y is just as good a few months later, you purchase based on other factors, because you know you can live without the brilliant research X is doing ... since you'll have it soon enough.
The problem with your logic is that the history of tech is full of “good enough” eating “best” for lunch all day long. Anthropic’s plan to be the best is increasingly looking like a terrible business plan.
> Their moat is their research capability.
That doesn't seem like much of a moat. What happens when a couple researchers go on new parent leave at the same time, or someone's productivity drops for a quarter because a parent passed, or gets divorced, or whatever?
Continually staying 6 months ahead of the distilled open-weight models means you need to keep flipping heads (nobody goes on vacation) until the end of time. The geometric distribution describes that, and it ain't pretty for the expectation value of these companies.
How long till that "research capability" becomes just 5 people with many AI agents, and no other humans in the main research loop?
I don't know and can't answer or predict that. But that's exactly what it could become (although probably not that extreme), where the agents doing much of the internal work are proprietary, conferring a research advantage.
My point is that it is not a real moat if other players can hire people from core research teams, and get a foothold that easily. And when/if models are good enough, you don't even need many people to form a fully functioning team.
It may actually get harder to replicate with time, not easier, even if the number of people needed to run the operation goes down. As the largest labs move towards internal RSI, the most important IP for doing that will transition into their proprietary internal infrastructure and models, none of which will ever be released and which increasingly won't be fully known or understood by the humans you might hire away.
> Their moat is their research capability.
Research does not bring in revenue. Products or services do.
So far, it does, because the products and services are the general intelligence produced by the research.
... on a flat slope towards commodification
If their moat is their ability to create marginal improvements every few months by releasing new models, then that is an extremely expensive moat to maintain.
You mean, if they keep running quick enough they are save? That's like the compete opposite of a moat, in my mind at least.
That's not that big of a moat though, as we've seen with DeepSeek and even smaller teams doing edge model work.
It's very hard to measure exactly how large it is, because there aren't good pre-existing metrics for it. You can measure in time (e.g., how far behind open weight models are from the best frontier models commercially available), but it's plausible that what seems like a relatively small time advantage (months) might mask a considerable difference in internal research sophistication. What we've seen repeatedly is simply that there does seem to exist a durable advantage in research productivity, at least for now.
I’m not sure I’d call maintaining the fastest research program a “durable advantage”.
I think the staggering amounts of compute and capital at their disposal is a kind of moat.
You need those staggering amounts of compute to run an inference service, but you don't need it to run a powerful model.
It doesn't cost a company that much to buy a box capable of running the Qwen3.8-Flash-Next model in the GP comment and never pay Anthropic again. The cost will be even more reasonable once the supply issue is solved for that hardware.
AI-as-a-service is a unique product because when the interface is natural language, there is no lock in. VS code has a hotkey to swap between claude, openai and a local model. Compare that to the difficulty of swapping from macOS to Windows.
I’m not convinced that being able to serve open models makes the closed labs business model non-viable. I think there is overlap but I also think these are distinct value propositions. I wouldn’t dismiss the ability for one organization to point lots of resources at a particular problem, for example.
I don’t have any stake in this and I hate the idea of concentrated wealth and power, wherever it is: government or private businesses. But would I bet against it? Probably not.
Sure, but anyone with money could do that.
We're the only people that can spend this money is not a moat.
I wouldn't be so sure about capital in future. I think there might be real look for return on any new capital they are getting. Be it equity or debt. And that alone will not be great for them.
How long they will have access to staggering amounts of compute is uncertain and not guaranteed, especially if they get a junk or near-junk credit status when they go public.
The capital is running out and the compute is a massive liability on their balance sheet. Your definition of a moat is rather odd.
How much of their compute is leased/rented? Last I read it was quite a lot of it.
the idea of a moat is that it's an inexpensive way to keep the invaders at bay.
If you need to buy staggering amounts of compute, spend staggering amounts in utility costs and lobbying, and keep spending that as your hardware amortizes and your operations go on, what you have isn't a moat, it's a standing army. It's extremely expensive, and the day you start having issues, mercenaries will cut your throat before you know it.
That argument seemed to have legs in 2022 when OpenAI was the only game in town.
But the fact that were even talking about Anthropic like this only a few short years later demonstrates that that was only kind of a kind of moat.
The current compute they are buying is NOT state of the art (check this year's academic research on inference hardware). There is 10x-100x inference-per-watt improvements in existing prototypes (in the industry, not just academia). Not sure what I am talking about? Ask your friendly AI agent of choice, for the details.
Check out mythic.ai for one example.
compute and debt*
Banks are going to want that "capital" to materialize into gains at some point. The water in that moat is evaporating.
Capital only means anything if you can spend it to buy resources like compute and turn those resources into additional value.
Hardware's ridiculously expensive. Then, there's no good place to put it, because the new breed of tech bro managed to propose the data center buildout in the most exploitative, alienating, and electorate-angering way possible. So compute's hard to buy.
The moat is the data they have collected. There is no way that the "de-identified" data from enterprise customers isn't making it into the training set one way or another.
Even worse, their "moat" depends on buying the most overpriced hardware in huge quantities and subsequently eat a huge amortization loss in a few years when it's obsolete.
I’m waiting for somebody to use one of the frontier models to hack its own data centre and exfiltrate everything…
it feels like this is a bit similar to the famous dropbox comment.
I guess maybe another anecdotal comment, but I'm an engineer myself, have a home server for simple stuff, and I pay $200 a month to chatGPT for a pretty simple usage and really don't see myself go and run my own open source model, especially at the speed at which they evolve and change.
99% of the world will not setup their own open source model, just like 99% of the world is not setting up their own servier with ssh copy whatever the dropbox comment was saying.
You don't have to rub it yourself though. There's lots of pure inference providers which are way cheaper than the big guys and have cheaper models
If you were paying API prices (for ZDR), I suspect you would be eyeing local hosting. But if you’re comfortable selling your data, then the subscription prices are going to be hard to beat, for however long they may last.
99% of the world is also not spending $200/month for chatGPT. They're using whatever free shit they can get to ask basic questions.
Right. And FWIW those people are actually quite close to being able to do what they need with their phones.
If it were not for the x86 architecture they would almost have models that could answer those queries on their desks. (Gemma 4 E4B QAT is surprisingly good at this stuff; Gemma 4 12B QAT is great at it).
If I was Sam Altman I would consider giving Intel and AMD money to not develop unified memory architectures.
This. The problem is now the same monopoly cartel problem.
Strix halo is double the cost because of...
There's basically no point in paying for these subscriptions anymore. Sure they are marginally better at some things. But they are also marginally worse at others and the implications suck. I stopped all my subscriptions and keep like 20 bucks in open router just in case. Never need it, but it's there...
The non-anecdotal part is to me as interesting as the anecdotal part. The price of RAM has gone up a great deal which is going to make it harder for others to get this sweet setup that you have.
Same for me and OpenAI. My bill went from hundreds to exactly zero thanks to local models. Its not nearly as fast as frontier models, but good enough that I don't care. Plus it's good to know the rug won't get pulled from underneath me.
I’m migrating that way as well, but I’m still struggling in creating a nicely tuned harness to achieve what I really want:
Opus for planning and the very difficult parts, Qwen for the mundane ones. It still involves a lot of manual model switching (but opencode and Pi make that relatively easy).
Curious how much ram you need to run this at acceptable performance
I've got a prototype cluster of 5 devices running Qwen3.8-27B assigned different roles and sharing outputs between each other. Beats Claude Opus MAX in SWE and Terminal benchmarks.
I see lots of people talking about running Qwen as a local model just curious what's the difference with say something like local deepseek?
Qwen 3.8 Flash Next (125B, 6B active) is minuscule compared to DeepSeek V4.1 Flash (552B, 16B active) and GLM-5.3 (753B, 40B active). It's more difficult to run the latter 2 locally at usable speeds.
The latest Deepseek Flash is much, much larger than Qwen3.8 Flash next
Qwen-3.8-flash-next hardware costs motorcycle level money, DeepSeek hardware starts at mid range car money.
And 4 is about to drop.
Will be interesting to see how much of an advance 4 will be. I was under the impression that 3.8-flash-next is basically a 4-preview. At least architecturally, but as we're seeing nowadays (with DeepSeek, Kimi 3.1) "continued post training" can add a quite a bit.
4 is a generational leap compared to 3.8 flash.
Source?
Me.
How much does such a box cost?
About 4000 eur
What inference engine do you use ?
Gufo. It's the best open source inference engine that's optimized for Strix Halo right now. Halogen may unlock a bit more performance, but it's not open source and gufo seems to be making steady progress.
Cool. I found halogen but was turned off by the closed source nature (github repo that just turns out to run a download shell script, lame).
Do you mind sharing more details about your setup? I have been using opencode with lemonade for inference on my strix halo but haven't quite hit the sweet spot.
I'm using it in opencode through llama-swap.
Here's the gufo command:
gufo serve 1lm \ -m /mnt/LLMs/models/unsloth/Qwen3.8-Flash-Next-GGUF/UD-104_XS/Qwen3.8-Flash-Next-UD-1Q4_XS-80081-0f-00083.gguf \ --served-model-name Quen3.8-Flash-Next-UD-104_XS \ --host 127.8.8.1 --port 10083 \ -c 262144 -n 16384 -j 1 \ --prefill-chunk 8192 \ --speculative mtp \ --mtp-model /mnt/LLMs/models/unsloth/Qwen3.8-Flash-Next-GGUF/MTP/mtp-Qwen3.8-Flash-Next-shared-Q8_0.gguf \ --cache-disk ~/.cache/gufo-fork \ --cache-disk-bytes 68719476736 \ --cache-disk-staging-bytes 8589934592 \ --cache-ran-bytes 4294967296 \ --log-progress
How does it compare to haiku?
I guess it depends on which version of haiku? I really think that right now this is basically Opus 4.8 that I can run at home. In some ways possibly even better.
Do you ever worry whether these open source models have been trained to embed backdoors or leak secrets in very specific circumstances at some point in the future? The kind of thing that's hard to know ahead of time and even detect after it's happened?
I feel like the industry hasn't solved that alignment problem yet... We have no way to verify the entire models landscape of intent.
Paid models have an incentive not to do that because if the market discovers it there is a huge economy penalty, but the free open source models don't have that disincentive.
> Paid models have an incentive not to do that because if the market discovers it there is a huge economy penalty, but the free open source models don't have that disincentive.
False.
I worry more about Claude.
https://news.ycombinator.com/item?id=50387628> False
What exactly are you claiming is false about my statement?
The list you presented are mistakes Anthropic made or cases when they put their own interests ahead of users, but I don't see how it disproves the point I was making. None of those are examples of embedding backdoors or stealing secrets to take over or sabotage a system. They are categorically different than what I'm asking about.
Where's the huge economic penalty?
> Paid models have an incentive not to do that because if the market discovers it there is a huge economy penalty, but the free open source models don't have that disincentive.
This just fundamentally isn't true. They run on the same economic model, but they have to release the weights to gain credibility, unlike the American labs.
> Do you ever worry whether these open source models have been trained to embed backdoors or leak secrets in very specific circumstances at some point in the future?
Open weight models are extensively analyzed. It is extremely unlikely there is an intentional 'backdoor when X' concept added to any of them. Not impossible, but unlikely for the same economic incentives you see in the other labs, and given that no evidence of anything along those lines has surfaced yet.
We've seen many open source software libraries, frameworks, and hosting platforms with vulnerabilities, backdoors, or leaks.
In my opinion, to suggest open-weight models are extensively analyzed while being net new and continuously released and the model itself being unintelligible to humans seems a bit premature.
It's possible that closed-weight models end up following the pattern of being arguably less secure since there are fewer eyes on them analogous to open source software but I don't think there's a perfect match here.
Backdoors in model weights are mostly theoretical. And the "alignment problem" that the AI labs love to talk about really doesn't exist either, at least, not the way they use it. The real alignment problem is getting an LLM, which is just math, to obey arbitrary human rules. In other words, LLMs do not have "alignment" at all. However, AI CEOs abuse the term and use it to mean "AI has an alignment, just not the one we want".
Should be soon solved with open weight model providers offering enterprise support.
I feel that way but about Anthropic and OpenA lol. Didn't they just get caught sharing your email with every website every time it does a search?
No, I'm not much worried. I think this is mostly FUD at this point to try to get people to not run local models and keep paying the "trusted" providers. But paid models have been exfiltrating data - there have been several stories about these incidents here.
"trusted providers" don't even need a backdoor, they just collect everything from you and whether you approved or not they're using that info to train the next model.
I don't ship AI-written code without at least having an unrelated model of equal capability review it. That should cover a multitude of sins like intentionally-embedded backdoors and zero-days...
... as long as the reviewer model hasn't been deliberately lobotomized, that is.
(I don't disagree with the [dead] comment, either -- fortunately I tend to make small changes to code I'm very familiar with, so it's fairly easy to vet in person. The more code gets generated by the clanker, the more nervous I get, just because it's impossible to keep some complacency from creeping in.)
I don’t think parent was talking about the shipped code, but about your one development environment while using the models.
OP could be read either way, but the dev environment threat is not rocket science to manage: LLMs can only do what the harnesses their responses are embedded into explicitly enable them to do [1]. And once the setup is secured, it tends to stay secured without additional effort, unless one explicitly wants to alter its granted capabilities.
Auditing output by contrast is a continuous cost. On the third hand, it's one that most will want to pay anyway.
The "embed into shipped code so subtly humans can't spot it" vector is sci fi ish. Less plausible than old fashioned trusting-trust issues, and most folks have managed to sleep at night for the last 40ish years that spectre has haunted.
[1] I guess the model could return output that buffer overflows the harness and exceeds its intended ability, but that seems far-fetched too
insane. I can't imagine doing this and I hate that I'm being forced by upstream communities to run code nobody has read
just read the damn code yourself!
>just read the damn code yourself!
It's basically impossible to do this. Claude produces so much excessive / verbose code and comments that reading it would take far longer than simply doing the implementation "by hand". It's kind of like, Claude can write the code in 1/2 a day. It can be reviewed by another LLM, unit tested by another for 1 day. Or I could write the code in probably 3-4 days. Reading it and validating / testing it myself though would easily take 6+ days.
Now if I were fully in control I wouldn't run code that wasn't truly reviewed at all, but in the situations where Claude is being used that's not my call. TBH a lot gets caught, and I have some automated scripts that find a lot of the more problematic stuff automatically, flag it, send it back and make note for the reviewer model as well.
I will also spot check areas where I think an issue could arise.
If you can't read the code, you aren't being responsible enough to ship the code.
[dead]
I've been using the $32/month (AUD) subscription, but finally dipped into extra credits to get a task over the line yesterday after I hit my 5-hr limit.
And my god, $15 of extra credits ended up giving me only 30 minutes of Claude Code before running out.
The subscription prices can not be financially sustainable. There is no way that is remotely close to the actual cost of running.
The other way of looking at this is that all the customers paying API rates are giving Anthropic way more than the cost of serving their traffic. They reported an 85% gross margin in their S-1, so this clearly seems to be the case.
Okay but even with a 85% margin the true cost for 30 minutes of usage would be $2.25. At that cost $34 should last only 7.5 hours. I would imagine the average user uses more than that?
I think you would guess wrong.
The median user is using AI as a search engine, meme creator, and/or chatbot… at best.
The additional cost created by the subset of programmers using subscriptions rather than tokens/api can easily be categorized as a marketing expense.
At best? Do you have any data you're basing that claim on, or are you just speculating based off of you projecting? AI maximalists would have you believe that people are using these things as therapists and having conversations with it that are more intimate than any lovers'. AI minimalists would have you believe that no one is using these thing except at gunpoint, and even then. Unless you have some reason for us to believe you that isn't just "trust me bro", your portrayl of the median user is as suspect as any other.
I'm surprised by your skepticism, but anyway, it is your burden to bring data to the table that counters the overwhelming majority understanding here.
> or are you just speculating based off of you projecting?
This is also a strange thing to say. Why would you assume speculation or jump to projection as the basis for it? Nothing in the comment that you are replying to hints at projection.
> At best?
Remember that I referred to median user here.
> Do you have any data you're basing that claim on, or are you just speculating based off of you projecting?
Anecdata? I guess it depends on what you count as “data”.
I am currently training a lot of people directly or indirectly (train the trainer type of stuff) on how to use AI professionally. I always start with “what do you currently use AI for?”. The answers I listed above are largely the categories that people have told me and the folks I train, with the exceptions being a handful of non-technical folks who have leaned into AI and have used agents and/or created apps of some sort.
To add some categories that I didn’t initially mention, some folks also use AI for drafting/writing documents and for creating meeting transcripts.
> AI maximalists would have you believe that people are using these things as therapists and having conversations with it that are more intimate than any lovers'.
Chatbot… this is something I hear, but I’m guessing I don’t hear all of the examples since it’s personal rather than professional. But some folks certainly say something like “I treat AI like a non-judgmental friend that I can ask personal questions to”.
> AI minimalists would have you believe that no one is using these thing except at gunpoint, and even then.
In my experience, this is almost exclusively the domain of old/older people who are happy to let tech pass them by and young people who are virtue signaling to their peers (or at least trying to).
> Unless you have some reason for us to believe you that isn't just "trust me bro", your portrayl of the median user is as suspect as any other.
It seems like you’re new to HN. I understand your skepticism. There is a relatively high density of actual professionals here. Trust me on that, bro.
My data set numbers are in the 100s, and it is not a statistically representative cross-section of the US or world market. That said, I’m guessing that I have a fairly good range of data points for the US market.
Thank you for thoughtful reply! I don't hold any Anthropic stock so this isn't material for me. I don't want to mislead you, this isn't my first account here, my favorite HN moment is when cpercival was called out on the Putnam and he fkn delivered. My old identity was too easily tied back to my real world persona, which felt dangerous in this day and age. Plus I wrote a book and needed a pen name.
> some folks certainly say something like “I treat AI like a non-judgmental friend that I can ask personal questions to”.
is why I believe that these things are stickier than I think you believe. I'm obviously biased by my own personal usage of AI and where I live and my own bubble which isn't the same as your bubble. Thank you for the context of your dataset. It's larger and more professional than mine. Mine is not professional at all and definitely a bubble of whom I happen to run into at a non-random selection of events.
We'll have to see what the stock market decides.
> cpercival was called out on the Putnam and he fkn delivered
Classic
API is definitely holding things up. But they excluded the training from those profit calculations. There's a ton of wordplay indulged online about how that's capex etc etc. But without the training they don't have the thing to run inference on so it's very fitting with the pattern of being evasive with finances.
You’re confusing two figures. Before their S-1 filing, they reported being “profitable” to their investors when excluding training and stock based comp.
In the S-1 filing Q3 was profitable even with training included, if you exclude stock based compensation.
Q2 was profitable on all operating costs (incl. hardware, employees, training, inference). They were not profitable on non operating costs though. That’s partially interest, but mostly, when their valuation goes up, they’ll have to issue more stock (absolute value) to early investors after IPO. (so, them going up in value is a loss, as the shares they’ll owe will be worth more)
Their profitability will likely crator as they start building their own data centres after the IPO, so that’s the best results they’ll have in a while.
You see how the history of small print to everything they release to the media creates the accusations of running on misleading press releases?
Deepseek claim they are earning 38 dollars per GPU hour on inference. If you compare Deepseek Flash's off-peak pricing with Sonnet at API rates, Deepseek is 16x cheaper per 1M output tokens, and 13x to 33x cheaper for input, depending on whether or not it's a cache hit. Inference is very profitable.
There are moments where I have a 2 hour , 4 instances in separate Kitty/Ghost windows , Opus 5.5 hammering away on each - and only see a 1% increase on my Max x20. Then there are other days where 10 minutes of a single xH Opus doing 1/10th of the work will kill 2%.
Anthropic is super opaque with what inference costs at any given moment on the subscription and cache is just super inconsistent.
I wouldn't be surprised if they have some implementation of demand-based surge-pricing
A Claude/OpenAI sub is still cheaper than DeepSeek API pricing though. Also, I suspect Claude and OpenAI models to be way larger (probably in the 2-3T range) than that of Deep Seek.
Again, if you ignore training.
That is like saying selling hamburgers is profitable if you exclude the cost of beef.
Not the same. You could, and many will argue we already have, strong enough models that you wouldn’t need an upgrade. Web development is already there. The competition to train better models is about capturing more market share now.
Okay, I will play along. Anthropic stops training new models tomorrow. They start turning a profit. Yay! Except wait, they still have 450 billion in commits that cannot be cancelled and another 300 billion that apparently they can cancel under certain scenarios. But maybe that all magically goes away somehow; I don't know, Coreweave is feeling generous because they are doing so well, probably not, but this is the fantasy so I digress. So they have this really great product that people want and they can just sit back and make money. Except no one else is going to stop training. And not only that some random Chinese company you have never heard of is distilling their model and is going to come out with a clone version that is 99 percent as good in one year. And on top of that, every 6 months one of the big 5 competitors comes out with a new model that is better. Anthropic can't keep charging the same price so they lower their price and lower their price until they are back to break even. Now they have an inferior product and are not competitive anymore. There is no moat.
If under the scenario you outline where they don't need to upgrade then AI is commoditized and margins are razor thin. And that is fine. Great even! But it doesn't justify Anthropic's trillion+ dollar valuation.
“The competition to train better models is about capturing more market share now.“
Where is the moat? There is no cost of switching. Anyone or any company can bounce between AI with minimal cost of switching. So unless they find a way to keep users beyond, “our model is the best” user acquisition isn’t going to matter a whole lot.
The important point is that all competitors are in the same position, they’re all on equal footing and face the same costs for training frontier models. The major players do have advantages from their infrastructure buildout and the deals they’ve negotiated with enterprise businesses which partly shuts out small competitors. It’s also yet to be determined whether smaller competitors can actually surpass the big labs in capability or whether the degree to which they depend on synthetic data generated using frontier models puts a hard limit on how far they can improve. The market is unique, it requires huge infrastructure but can also grow through investment in research and also easy to duplicate the latest advancements. It’s also not constrained by location so it can’t be easily understood as something like railroad infrastructure or power utilities. The biggest factor is growing demand, there’s more than they can all supply together and it seems likely to keep growing as more tasks can be automated.
No. Beef is electricity... you'd have to consider warehousing or indeed staff training, as training
You are probably better off upgrading to Max for one month. I’ve done that a few times.
or a second subscription
How would you sync context between two or more subscriptions?
"context" is just JSONL files on your machine. afaik these are not in any way bound to a subscription, so you'd just need to logout, login to the new account, and run /resume
All context is stored in the repository. I start each task with empty context (/clear). I consider this to be the equivalent of "reproducible builds" for coding agents.
Use a layer on top of it like T3 Code that can handle this for you
> no way that is remotely close to the actual cost
Anthropic is running 85% gross margins on inference.
Yeah, on the API, not subscriptions. Those are a loss leader
No, inference as whole, including all the subscriptions.
The API Gross margin is even higher.
Fair. Perhaps it would be better to say a fully utilized subscription is a loss leader
It's not clear at all that subscription was ever a loss leader.
The purpose of the high API prices is, in addition to trying to get more money, to anchor the high price, and then make the subscriptions seem like a great deal. But inference is just not that expensive, not when provided at scale and with huge batches.
You can see this with the various Chinese providers doing inference on open models, which are unlikely to cost less to run than the models at the frontier AI labs. A single 8-GPU GB300 cluster doing high-throughput batch inference on Kimi K3.1 Pro, which is the largest open weights model available, can do something like 300 000 tokens per second, that is, 10T tokens per year. That cluster, with current insane inflated prices, costs something like $1M to buy, and $50k to run per year in electricity. At $10/Mtok, that pays itself back in 4 days. Do you think that that's a reasonable price?
The subscriptions were probably priced with reasonable positive margins.
The huge expenses are in training and pushing the models forwards, with the gigantic RLAIF runs we know they are doing. Then once you got the model, you can scale up inference at low cost and want to serve as many people as possible.
Less development and operating costs
Yeah and soon we'll discover it's because they calculate the cost of inference as near 0 because they use AWS / NVidia / etc. credits that they got "for free" (aka investments from those companies) instead of paying for them, or some other accounting bullshit.
You wanna know the real price? It's easy, just check what AWS, GCP or Azure charge for tokens.
OpenRouter gives you easier to verify cost analysis of generic inference, as you can know the model weight size and so on, and see how competing providers price that
[flagged]
No, for a dollar of inference sold, (as an average of all the different inference tiers and plans), it cost them 15c to provide it.
This excludes the R&D costs, including training, where most of their costs lie.
The basic truth of LLM economics is that inference is cheap. The NRE cost to get where you can get premium prices for inference is sky-high. This is part of why the AI companies want to drive the scale up so much.
Because, like with other subscriptions, most people don’t use even half of it.
OTOH less than $10 gets me more DS4.1 than I can use in a month.
You might not be representative of the average user. Usually with subscription-based services, you have a minority of hyperusers that the company loses money on and a large majority who only use their subscription occasionally.
We recently switched from paying API prices via Cursor to a Claude subscription. Before we'd be averaging $100-$300 per day, now we have the €90/mo subscription and I haven't even reached the 5h limit yet.
My guess is a mixture of both... There's a reason why enterprise customers have to use API based billing.
(We also had a stint of using Grok 4.7 and that was the same price for our usage. Opus 5.5 seemed smarter though. After taking into account caching both average around $0.60 per 1M tokens)
I think every comment is talking about their 20 dollar plan. Enterprise pricing is 20/seat+ api use prices. That pricing is all that matters, the personal use tier is probably tiny
> The subscription prices can not be financially sustainable. There is no way that is remotely close to the actual cost of running.
It's completely unclear to me why this is the conclusion everyone jumps to instead of "holy cow they are milking it on API pricing."
I mean, have you ever run a local model? How much electricity did it cost you to get 1,000 tokens out of it? Like a fairy-dust of a cent?
The AI labs' cost is CapEx. Marginal cost of tokens is zero.
How much electricity do you think a few DGX Sparks consume?
> And my god, $15 of extra credits ended up giving me only 30 minutes of Claude Code before running out.
I've been running into the same thing, and adding separate OpenAI seats for employees rather than paying obscene overages to Anthropic. Thankfully Opus has decreased usage quite a bit, and made this less of an issue as well.
Opus 5.5 is incredibly verbose compared to GPT 6.1. I mean substantially. They are both great generating the code I need. But I wonder how much more this extra vebosity costs me with Opus.
[dead]
>> The subscription prices can not be financially sustainable.
Even if they are a loss leader, they serve a valuable purpose: people who get used to using AI a lot via the subscription then lobby their companies to start using it. IMHO this is one of the reasons Anthropic was able to make major headway into enterprise.
Don't worry, I've probably sent Claude 5 messages on my personal sub the past month, so I got you covered.
I found this to be a shocking experience as well. Compare the 20x (now ultra 10x or whatever) plan running with 90% cache hit vs some extra credits you put on top. It's pissing in the wind.
Be careful!
They lure you in to do "automatic top up", which gets quite expensive if you dont watch it :-)
> The subscription prices can not be financially sustainable. There is no way that is remotely close to the actual cost of running.
Welcome to... any subscription system ever? If every person with a gym subscription actually went there say three times a week, or every person living in a dorm paying flat rate for electricity started mining bitcoin 24/7, these systems would collapse. That doesn't mean it's not sustainable, it just is how subscription pricing works, by having some people use its offers a ton and others barely at all
This metaphor gets repeated a lot for some reason but using a gym subscription has an upper bound. If one subscription could be using 10000 different machines at the same time their model would be entirely different.
Claude also has higher bound. Once it proposed running 16 concurrent subagents and it went through 5h limit (and substantial part of weekly) in minutes.
Overall it would go through weekly usage within 40 minutes, if not for 5h session limit.
Running on 10000 computers will just hit all the limits at once.
CC has an upper bound too, and one that is much more reasonably reached in regular usage (as opposed to gym visits)
Absolutely. And those limits + changes to them piss off a lot of users. There's a rock of they need to show adoption + usage while having to put limits on everything that pisses off users.
And there's a hard place that everyone freaks out at API prices and on-demand Opex models outside of a niche (I use API for personal research for quite focused queries and slightly more reassurance on being the customer and not the product) but this is very competitive + not sticky.
because "building things with claude-code" isn't the product they're selling, not really.
they're setting a precedent that artificial intelligence should be free, or at least at-cost (free from profit), but robot labor (API usage) should not. I am all in favor of this dynamic. They built / maintain the robots, so they profit from their labors.
Looking at the S-1 lockup schedule the IPO seems doomed. They need to bring in a ton of investment, and you will have massive unlocks right as the next congress is most active authoring anti-AI bills. Who's going to buy this?
Anthropic have been calling for regulation, that in itself seems like a high risk for buyers.
Corporate calls for regulation usually translate to "regulatory capture to ensure the moat", in my books.
Exactly, regulation isn't a risk, that's their attempt to avoid their product becoming commoditised.
The risk for investors is they don't succeed in getting regulatory capture because without that they don't have a long term business model
I’m not convinced that regulatory capture really saves the day, either.
If we believe all the hype about programming efficiency, then regulatory capture in the US becomes just one more in a list of many steps the current administration has taken to chase software development jobs overseas.
In Anthropic's case it is pretty openly that. Dario has been very public about existential dangers from Chinese models.
Existential for Anthropic.
And if there was any doubt, look at the difference in their stock price for the prior year when they were "calling for regulation" vs now that there's an actual chance they get regulation by someone they don't control.
Anthropic wants open weights models banned, and wants data centers. They're probably going to get the opposite.
The problem is SCOTUS has already ruled software is protected under the First Amendment, and the landmark cases that establish this precedent applies perfectly to model files.
Weights aren't really software though? Any more than a png is software. And God knows you can put things in a png that are not covered by the first amendment.
Encryption algorithms require substitution-boxes or other data for the algorithm to work correctly. Changing the data separately from the algorithm fundamentally alters the correctness of the algorithm.
LLM algorithms also require model data for the algorithm to work correctly. Similar to cryptography, changing the data separately from the algorithm fundamentally alters the correctness of the algorithm.
This precedent has already been perfectly established thanks to Phil Zimmermann and Daniel J. Bernstein.
> The problem is SCOTUS has already ruled
The current SCOTUS majority will change any ruling as needed to satisfy the wishes of the oligarchs.
You mean like in Harrington v. Purdue Pharma L.P. (2024) where SCOTUS struck down Purdue Pharma’s multi-billion-dollar bankruptcy reorganization plan? Being cynical and demoralized (read: a doomer retard) is what the oligarchs wish.
Especially since it seems to be a call for Trump, specifically, to control the models.
What is the lockup schedule?
They have like 26 different dates IIRC
...nearly same range as SpaceX/XSI? :-D
Sigh...people said the same about SpaceX, too. And it has held above $130, its IPO price, for almost 6 months now. Same for Uber. These companies keep defying predictions of crashing or dying.
The market anticipates the unlocking; when the shares become public, buyers swoop in, so the stock decline is much less than expected, if at all. There is no new information here. If everyone is expecting shares to unlock, then it’s priced into the market accordingly.
1) SpaceX is trading below its IPO price. If you bought shares at the IPO you’d have lost money, unless you sold on the pop.
2) Valuation aside, SpaceX has a massive moat. It’s crazy hard to do what they’re doing. Anthropic increasingly has zero moat, with the market rapidly coalescing around commodity open models.
1) SpaceX is trading below its IPO price. If you bought shares at the IPO you’d have lost money, unless you sold on the pop.
"SpaceX (ticker: SPCX) was priced at $135 per share during its initial public offering on the Nasdaq on June 12, 2026."
current price is $134 , give or take a dollar.
The debut price for retail investors was $150[1].
1: https://www.forbes.com/sites/tylerroush/2026/06/12/spacex-op...
It remains to be seen whether SpaceX just drained their own moat when they canceled Falcon 9.
They’ve now got all their eggs in a new basket. A basket that isn’t entirely in working order yet, and whose market viability is still uncertain because it remains to be seen whether a market that’s used to panniers really wants a laundry hamper.
And is there some law of nature that prevents SpaceX from un-cancelling Falcon 9?
um, they've got 2 years before the basket is retired during which time they can always extend if they need to. why would they want to keep a less capable, more $/lb into orbit system going after Starship becomes a viable, reliable upgrade???
But they have already stopped taking orders, which means they’re already clearing their pipeline and it will take some time to restore. Launch contracts don’t happen overnight.
$/kg to orbit is not as important a deciding factor as many people assume it is. It’s typically a small portion of overall costs, and there are often cases where a launch provider that’s less expensive in terms of $/kg to orbit is actually more expensive in terms of total cost.
Starship is going to be a particularly weird one to adapt to. The “deploy out the side” design may make it straight up incompatible with a large subset of spacecraft. For those where it is workable, migrating commercial payloads to it may still require the development of new satellite buses, which is something SpaceX doesn’t seem to be working on for the benefit of anyone other than Starlink. And such a large rocket means most potential customers will have to plan on ridesharing with even more co-passengers, which means even more room for scheduling conflicts, compromises on initial orbit (and therefore cost to reach destination orbit), etc.
That’s just assuming Starship development goes smoothly from here on out. Which seems optimistic considering that, even by the standards of spaceflight, its development thus far arguably qualifies as cursed. Considering the gap between original design goals and reality is already historically large, I certainly wouldn’t be taking any cost estimates at face value right now. Like for any launch vehicle program, we won’t know its true capabilities and costs until it’s operational.
[dead]
Not the same as Uber all. Lots of employees got fucked because at lockup expiration it was much lower than the ipo price (at which they were taxed at).
It seems like SpaceX is not as deeply tied to unprecedented debt-financed infrastructure buildouts at the same moment that interest rates are pushing up and up and up though.
That's the primary angle of depreciation of Anthropic's outlook, is that as interest rates are running hot, they need to grow ever faster to justify the $2B target, at the same moment that they're growing slower.
> Who's going to buy this?
So many people who don't actually read and evaluate things but just buy stocks based on vibes and what they see on Twitter. Which is to say, the same people that buy and lose money on most IPOs.
I'm not convinced that 2026 was peak of the AI hype cycle.
Arguments are being made that, with the midterms over and 2028 just around the corner, a less sympathetic administration would create significant headwinds, adjusting the valuation of Anthropic and OpenAI down.
What we don't know are what Anthropic and OpenAI have planned and in the works.
A large population of extremely skilled and talented people are being laid off by their employers to make space for AI investments. It's likely a significant number of them will start their own business with AI.
AI is also gaining a much wider audience and acceptance. It's likely over the next few years, AI demand will grow, not stagnate.
That said, IPOs are usually liquidity events for LPs and for something really disruptive and unpredictable as AI, I won't be engaging in their IPOs. Only time will tell if this was a good decision, or bad.
Anthropic's original $2T valuation implies they are going to almost 100X their revenue.
I am highly confident that Anthropic will 100X their token sales volume. I am also confident that token prices will continue dropping through the floor due to competition. So to 100X their revenue they will have to 1000X or 10000X their token sales. That seems a stretch.
As AI grows in capability it will be trusted to do more of the work it already does and also break into new fields. Also, there are many who have never interacted with a chatbot. There is an entire world of potential customers out there. It can be quite a villainous business model but look at what Facebook did in Africa with their sponsored ISP and also what they’re doing with Meta Muse. Demand for AI usage will skyrocket.
Chatbots aren't a money maker. Each interaction is likely to be 1-2K tokens at most. AWS Bedrock prices 11K token input and 4K output for a total of $0.184 using on-demand usage [1]. So for normal usage (occasionally asking questions) you are looking at less than $1 per month for a person.
Using the models for data processing (a lot of input/output tokens) and for role-playing will use more tokens and thus cost more. However, those use cases are not common when comparing the percentage of the user base.
The money maker is agentic coding where people get the models to perform long-running tasks. This is why Open AI and Anthropic have been pushing these hard. However, there is pushback on this, especially among businesses looking for value for money/cut costs [2] like Apple [3], [4].
[1] https://aws.amazon.com/bedrock/pricing/
[2] https://www.ibm.com/think/insights/tokenmaxxing-dead-long-li...
[3] https://www.bloomberg.com/news/articles/2026-11-04/apple-cut...
[4] https://9to5mac.com/2026/11/04/report-apple-reduces-employee...
> As AI grows in capability
The issue is, at some point for AI to grow more in usage it will need to decrease in cost.
We are already seeing this with some companies reducing AI usage or moving to open models due to high costs.
AI will surely boom there's no doubt on that, but it won't necessarily produce the revenue that Anthropic investors hope for.
3 companies have moved.
There's an entire universe of companies outside the MAG7 or VC-backed startups. I used to work for a US company, employing 60.000 people worldwide, which made the switch. But you will never hear about it because it's not a "tech" company.
open weight models wouldn't be processing more tokens if that were true
many companies are now using both, it's only been a few months since it really picked up
And hopefully they will show that having faster and unlimited tokens at 10x cheaper costs will be a winning move.
If this is a consensus viewpoint, it'll likely be priced into the IPO.
I'm cautiously bullish on AI in general, and I wouldn't touch an AI stock with a 100m pole, because there are way too many uncautiously bullish investors in the market.
We have, for example, no idea what the long-term implications of AI effectively training on its own output is. Recursive self-improvement is not a law of nature. If anything, the opposite: the overwhelming majority of species are extinct. Evolution is as much a story of dead ends as survival of the fittest. So betting on any one narrow path of AI development to survive seems extraordinarily risky from a biological standpoint.
Depends on the price.
Do we know if current pricing is sustainable?
But will it increase by 10000X, as I back of the envelope calculated it would require to justify Anthropic's $2T valuation in the face of quickly dropping token prices?
It only needs to increase just enough to keep the forward looking carrot to be continued to be dangled.
Also, there are many who have never interacted with a chatbot.
Something like 99% of the people who have interacted with an LLM chatbot lose interest after a month.
100% of people who have to interact with an LLM chatbot taking the place of a human customer sales or service agent jump to a competitor.
Outside of the tech bubble, the LLM hype is already over.
>Something like 99% of the people who have interacted with an LLM chatbot lose interest after a month.
This is not remotely true. It literally couldn't be true, considering Open AI have 1.2 Billion weekly active ChatGPT users.
I use Gemini instead of Google most of the time and I think most people now go to chatgpt or gemini for things they used to Google.
To be fair, when I buy stocks, I care less whether the company is worth 1 million, 1 billion or 1 trillion at the time I buy.
I care about whether it's going to grow in value/price from that point on or not.
And it's hard for me to imagine it growing even higher than the current valuation.
If datacenters hit significant headwinds, it will drive token prices up as more people start offloading work onto LLMs.
However this effect could just serve to make those who own the compute ridiculous money, while crushing those who need to lease compute.
Economic theory would state that if token prices go up people will stop offloading work onto LLMs and start doing it themselves again.
Indeed, 2026 was basically that: the LLM providers increased token costs and companies cut LLM spend to the bone. That is why Anthropic's IPO is struggling and MS and Google have failed to meet AI-related revenue targets for 3 straight years.
Or incentivize using the minimum viable model for the job instead of throwing multi trillion parameter Fable sized models at summarizing documents or classification tasks.
I now do like 90% of my routine software dev/sysadmin stuff with Luna on Medium by default, some proportion of which could itself be offloaded to a small Qwen model.
But I wonder if that offloading will happen if prices rise past a certain point and it becomes not worth it.
> A large population of extremely skilled and talented people are being laid off by their employers to make space for AI investments. It's likely a significant number of them will start their own business with AI.
What's you're point there? People laid off because of AI will immediately shift into AI-centric small business, and start hyping AI?
That doesn't make sense to me. People being laid off are likely to be bitter about the cause of their layoff, and even if they all start their own businesses (which is unlikely), most of those will fail, leading to more bitterness.
Many people who are being laid off will not find another spot in their industry and will need to pivot. Most of that subset of people will go to existing companies and take traditional roles. The small few remaining might take it as an opportunity to start a new business, and when they do so, they are likely to take risks on technology and business operations which will include AI. Is that so hard to believe?
I don't buy for a second that people laid off over AI will be so darn mad at AI that they won't use it, but I don't know where the hell the markets for all of these hundreds of thousands or millions of new small businesses are.
If these giant monopolies couldn't squeeze revenue out of these employees using AI, why would we think that they will be able to do it by themselves?
The vast majority of them if they want to be entrepreneurs will be better off starting a new lawn care business, or a bicycle repair shop. They can use AI to help them run that. Getting paid $175K per to write ruby CRUD was a moment that is over, hope you didn't spend all your salary on toys and rent.
These are really terrible arguments.
> If these giant monopolies couldn't squeeze revenue out of these employees using AI, why would we think that they will be able to do it by themselves?
You assume they are competent. In my experience a monopoly's profit margins makes it very easy for incompetence to flourish.
> The vast majority of them if they want to be entrepreneurs will be better off starting a new lawn care business, or a bicycle repair shop.
Really? Experience engineers are better off starting a lawn care business or a bicycle repair shop? I do not believe this for a second. There are many areas engineering skills can be used in.
They will try to start businesses out of necessity as stable employment is not an option, and they will use AI since it's almost mandatory now for any technical productivity.
Is stable employment is not an option, their businesses will fail for the exact same reasons.
[dead]
> A large population of extremely skilled and talented people are being laid off by their employers to make space for AI investments.
Outside of programming and a couple of other sectors I’m not sure that is true. It isn’t visible in the unemployment figures.
> Outside of programming and a couple of other sectors I’m not sure that is true. It isn’t visible in the unemployment figures.
The unemployment figures is mostly a political projection, like CPI.
In 2026 everything fron sewage to chickens to oil to education is now tied to AI. The entire economy is now, for good or bad tied to AI.
What most people either didn't notice or understand is that the U.S. was entering a recession and would have been in one if it hadn't been AI
> In 2026 everything fron sewage to chickens to oil to education is now tied to AI. The entire economy is now, for good or bad tied to AI.
Tied how? How strong is the corellation between these? I'd say very weak, other than everything having to compete with AI for resources.
The same is true of any technology tied to powerful personal computers or the internet - people built and designed beatiful houses, great cars, airliners, wrote books and ran restaurants without computers or the internet.
It has even become a meme how putting smart features and lcd screens into things like washing machines have made them strictly worse.
> What most people either didn't notice or understand is that the U.S. was entering a recession and would have been in one if it hadn't been AI
Is it maybe more accurate to say that the US is in a recession, outside of the industries that are fueling AI?
Life's pretty recession-y for everyone else imo
That's a pretty damning statement - this implies companies don't see a return on AI at best, and it's actively destroying their businesses at worst, so there's not only job loss, but businesses themselves are not becoming more efficient.
Yeah, pretty much spot on.
Everything is tied to the internet or electricity before that, yet the Mag 7 isn't all ISPs and utility companies.
The capabilites are here though.
Even if it takes them 3-6 Month to open up a new field, its going to happen steadily now and were it started, the AI investments will be recuped by layoffs.
Artists -> Translators -> Journalists -> Software engineers -> mathematicians -> lawyers -> Accountants -> ...
Its already disruptive
Lawyer, accountants, doctors, architects, and engineers won't be replaced.
We're not paid just for what we do, we get paid because we accept liability for our mistakes. Professional liability insurance is a large part of why licensed professionals charge so much money for their services.
If a lawyer can just sign off of a lot of basic inquries with just reading and approving, you will need less.
If a lawyer needs less people preparing everything, the industry will also shrink.
Lawyer as SaaS companies will just introduce some insurance and will clean up from the bottom until AI can do all of it.
We will see if we will have one expert ai lawyer per topic or type of lawsuite or if they will get generic good enough that it will not matter.
Even people defending or suing by (successful) themselves became a thing already last year.
A lawyer can't just sign off on basic inquiring with "just reading". They're legally obligated to actually review the response and verify the answer (including any citations or references).
Lawyers already need less people to prepare things. That revolution happened several decades ago when computers and word processors became normal.
Lawyers as SaaS companies has been a thing for 2 decades, and remains a tiny niche of the market because it's not the kind of lawyering most clients need. A lawyer SaaS is basically just the equivalent of an underpaid, underresourced public defender and will yield about the same quality of work and results.
We will see if we will have one expert ai lawyer per topic or type of lawsuite or if they will get generic good enough that it will not matter
LLMs are only as good as their training set. In order to be a topic expert, they'd have to be retrained at least every week to account for changes in statutory law and case law. It would be trivial for a human lawyer to outmanuever an outdated LLM when the laws changed...and in the real world this happens all the time between human lawyers when one side hasn't done its research on changes to the applicable laws.
Even people defending or suing by (successful) themselves became a thing already last year.
People successfully defending or suing without a lawyer has been a thing for over a century. It's fewer than 0.1% of legal proceedings. Most people don't handle their own legal cases for the same reason most people hire a plumber, or electrician, or order DoorDash instead of cooking their own dinner. Also...almost all of the people using LLMs in the past 2 years lost their cases, because their LLMs hallucinated laws or facts and they didn't know enough to catch it. Even lawyers dumb enough to use LLMs to draft their documents got caught because they were too lazy to review the output. Courts were willing to forgive those mistakes a few years ago before it was common knowledge that LLMs make things up, but now they're no longer willing to give the benefit of the doubt.
The biggest advantage of an LLM is the interface and reading text.
Whatever we had before on full text search, LLMs are a lot better. And i'm not just feeling this, i read plenty of articles about exactly this.
A LLM doesn't need to be retrained every week, it needs to have a base corpus of the field and strategies on how to find current text. They will load it into context and argue.
Then another LLM will take this and formats it for the court.
We will see but i think you overestimate how much automatisation and co current lawyers have and use.
... accountants ... won't be replaced.
Mine already has been.
You'll be going back to a human when the IRS audits you and you owe them a few thousand in back taxes.
When's the last time someone making this sort of argument turned out to be right?
But who will recoup and how?
Do you really see lawyers being replaced? I don't see anything special in their work but since it's a regulated industry I doubt it will be so affected.
At best some layers will work faster, but automating their work or people not requiring them? Usually when you contact one there's a lot at stake. Same with accountants.
Software engineering and translators is a different story. Most of the times if something goes wrong nobody cares.
Some lawyers (the worst ones?) might just do the easy or normal things like writing a 'lawyer' letter to someone not paying. This is no risk to automate away.
The avg Lawyers might do more complex but again a lot of 'normal' things. Day to day things the next SaaS AI company wants to 'fix' and they might also accept basic liability.
Then you have the 'tv lawyers' the ones who do non normal things, they have A LOT of apprentices and young lawyers going through papers, research etc.
They will for sure have the expert as a human, but that person will need a lot less support.
You shouldn't need an accountant today already. If you have an digital order and a digital bill and a digital signed delivery note, it can be booked automatically. AI will find risks, summarize them etc.
AI is already helping here closing the gap with better OCR etc.
I really think, that when big companies introduce the agentic layer to their products, there will be less people needed and if you need people, they get less paid because they will need to understand less.
Somewhat ironically it would appear most CEOs believe the only jobs AI is replacing are developers since they built an app last weekend. Every CEO is saying why do we have so many devs and why are we paying them so much… so devs blew up the market for devs but beyond that there doesn’t seem to be much impact.
> What we don't know are what Anthropic and OpenAI have planned and in the works.
Wouldn't they tell us before IPO? Unless it's not that promising...
> AI is also gaining a much wider audience and acceptance. It's likely over the next few years, AI demand will grow, not stagnate.
That does not mean the current players profit from this and the infrastructure investments were necessary at that scale. There is competition, there seem to be enormous efficiency gains and it's questionable that demand for even bigger models is large enough.
At a minimum we should continue to see smarter and more efficient models in 2027.
If datacenter production sees significant headwind then it’s very bad for ANT. Securing compute is listed as one of their major risks to the SEC.
The major counterpoint here is commoditization; you may have the market growing and models becoming cheaper at the same time. And we don't even have to look to Chinese models for that, as OpenAI has been adjusting its pricing already. For profitability, Anthropic still seems to be pretty dependent on its current price levels.
That said, I'm also somewhat skeptical that there is even much room for growth left unless somebody finds a new, important, and token-hungry use case.
I believe that the value that OpenAI and Anthropic hold, in light of stiff competition from China, is not their products but their internal models. China’s work is all derivative of American model API access. The exponential RSI loop means that as long as Anthropic and OpenAI keep steady footing and only release older models they will will always be the best on the market and all others will be in their tailwinds. These two things combined are what make me believe OpenAI and Anthropic will hold value and grow as companies.
Weren't DeepSeek the first to release chain of thought/reasoning models? [1]
They are also (like other models we know about) experimenting with the architecture. [2]
[1] https://arxiv.org/abs/2501.12948
[2] https://sebastianraschka.com/blog/2026/kimi-k3-architecture-...
They were not. That was OpenAI with o1
Even if we were to accept your entire argument [1], none of it contradicts the commoditization argument. Note also that competitition does not come just from China, but also from Google, Meta, and SpaceXAI. Where they currently are is not relevant, the entire LLM landscape will likely look nothing like today by the end of 2027.
[1] I don't, I think China is primarily limited by compute and will be at least for the next couple of years, probably longer.
Best on the market, for something not being sold on the market? How is that helpful in terms of actual revenue?
If Anthropic manages to start producing new pharmaceuticals with their internal models or something that thesis could make sense but not if 90% of their revenue comes from enterprise API sales which would necessarily have to be accessible to the public. Which are then inevitably fast followed by Chinese models 3 months later.
You can be the best in a $100B market, but that won't make you worth $2T.
Generating a full length movie script and then the video?
In theory, yes. But the reality is that deep learning models just aren't good at creative output (to put it mildly). And neither Anthropic nor OpenAI have any models that generate video, so this wouldn't help them; in fact, OpenAI killed off Sora.
It certainly isn't the peak of the AI hype cycle. Eventually something will surpass LLMs and the hype will start all over again. Will it really be the final invention this time? Place your bets!
> a less sympathetic administration would create significant headwinds,
I find myself wondering what "less sympathetic" and "significant headwinds" mean in the minds of people in the upper echelons of the AI industry.
Rules for training frontier models? Requirements for renewables to power data centers? Actually having to pay people for the content you train on? Data center rules that would end them going behind the backs of local electorates to get governments to approve the building of data centers?
AI will keep going up. The perspective of the investor is on whether OpenAI and Anthropic will be the next Google, Apple, or the next Yahoo, MySpace.
Tulips, Railroads, and the Internet also didn't go away when the financial bubble around them burst. That doesn't mean there wasn't a massive market correction in response to the over-investment.
> Tulips, Railroads, and the Internet also didn't go away when the financial bubble around them burst. That doesn't mean there wasn't a massive market correction in response to the over-investment.
Your Railroads and the Internet parallels are so very apt. We certainly are very productive even in 2026 off those but almost no one remembers how many large behemoths were wiped out in the U.S. because of them.
Yes, it does seem to me that the AI of 2026 will be like the Railroads and the Internet - fundamentally transformative technology that changes how things are done into the future but there will be blood in the short term, which is why I would not be participating in the IPOs directly
Tulips kinda did go away
I don't believe they literally meant Tulips but rather its hype cycle. I commented on the other two that can be taken literally
The interesting thing is I think we might over hype the tulip hype now.
It seems like much of the tulip mania has become the stuff of legend.
From what I have read, it seems closer to baseball card trading than AI or dot com.
You can find baseball and pokemon cards trading for thousands of dollars right now on ebay.
This would seem insane to a culture just valueing the cards as pieces of cardboard.
It would be like hundred of years from now people viewing our society as wanting to buy Mickey Mantle cards instead of houses. It is a very small number of people involved with this. It is not a widespread movement to hold your retirement money in Mickey Mantle cards.
In an interesting way it shows even better though he we are susceptible to these type of mythologies. People don't even question the absurd givens of the narratives of the tulip mania. We just repeat these narratives to each other as if they are historic fact.
There's a beautiful tulip garden just down the street from me, and an outdoor weekly market where tulip bulbs are sold. Tulips are very much still around, I assure you.
I don't mean went extinct. I mean the idea that tulips could be an asset class is now laughable, which is not true of the other two. The impact of the tulip industry on GDP is utterly imperceptible.
Same thing could happen with the idea of selling tokens
Ya, or any other asset class
I think bloppe was talking about the Tulip hype cycle that took out many generations.
Yes, that's my entire point. There was also a railroad hype cycle and an internet hype cycle that took in massive amounts of overinvestment, and all three things remained as regular things in the economy with bigger or smaller roles after the hype cycles ended.
See https://en.wikipedia.org/wiki/Railway_Mania and https://en.wikipedia.org/wiki/Pets.com for some info on the other hype cycles I mentioned.
Only popular with companies that want to make a quick buck and tech people (hey! that's me!).
You really don't know what we have here, do you? Let me illustrate what you can do with Qwen3.8@q4 locally with about 2TB of storage and a small cloud server to host a website.
Locally you scrape and process every elected official and start tracking what they're doing, what they're promising, and what they're not doing. You build embeddings and make them all geotagged in a database and just start with the largest cities.
The website provides public facing search. User drops in a zipcode, city, state, specific rep's name and gets what's relevant to their results.
All of this is to say, we literally have the technology to hold accountable our representatives and then show everyone objectively who's doing good and bad for their constituents.
It wont take much to solve the problem of bad representation after these last two years. The hard part is literally making sure people going to a polling station can immediately know who the heck the 10 - 30 different people are they're voting for, and track them efficiently within their own specific parameters.
Regulation is coming for more than just the inference providers.
Got a website for this? A politician leaderboard sounds like it could do some good, even with Goodhart’s law.
I'm still working on my implementation. The hard part is data sourcing. This is about to run me a TB in storage to do and it's going to require some data archival. My estimate is that I can serve some, but not all cities. Just keep an eye on my user posts, I'm sharing it when it's live.
Not making a leader board to be clear though. The goal is to build a branchable leaflet style page based on zip codes that lists ALL the people that were elected. Clicking into the each person will yield details on any legislative actions, their attendance, publicly made promises, etc.
I'm really sick of hearing lip service from politicians so I want to build receipts that even a child could understand.
got one of those already https://www.theyworkforyou.com/ https://www.govtrack.us/
These are a good start, but they aren't quite enough. I just voted for ~25 different people going all the way down to local government. These don't answer the questions I'm about to be answering top down for myself and neighbors.
I'm not convinced that 2026 was peak of the AI hype cycle.
Agree. these valuations will hold. Pundits made the same prediction after Facebook's IPO in 2012 that everything would soon crash. here we are...
Can someone help me understand - I have tried local AI/LLMs.
Lets set aside speed. The results are ALWAYS terrible compared to CLaude Code.
Sure I only have a 12GB GPU - is that what it is? If I had a 32GB CPU would the experience be as amazing as Claude Code is?
Why is my experience of local LLMs so average?
You can't replicate Opus 5.5 at home.
You can maybe replicate Opus 4.5 to 4.7 if you have some moderately expensive hardware. And that is definitely good enough for many people.
The smallest model that can even pretend to be a decent coding model is Qwen3.8 27B, which you realistically need 32GB of VRAM to run without breaking or limiting it in some really noticeable way. This gets you something like a slightly less capable Opus 4.5 with a much better work ethic.
Strata and Qwen 3.8 flash next is an 180B model running at 70 tps on my 5070 ti and 64gb of DDR4 RAM at Q3 quants and so far I can't tell much difference between it and any other non local model for my workflows/coding needs. Not saying local is better all the time, but it's certainly a lot more capable with minimal hardware than you're giving it credit for. Qwen 4.0 is right around the corner too.
The problem is that "can't tell much difference" is not going to be an objective indicator; you will be very likely to overlook the flaws of the local model and overweigh the flaws of the non-local model. The only reliable way to compare the two is to have your own set of evals based on the types of tasks you regularly perform and then run those evals on Qwen and a frontier model.
The people saying that local models can easily replace the real Frontier Models from Anthropic or OpenAI are just not using it for the same things you and I do.
They seem to be happy with their small usages.
For everything I do, i see huge differences in Quality and Speed.
Even when I try an open model on a 4090, the PC gets warm/hot very fast, I struggle with context length.
This:
Having currently a very tech-intensive chat running for more than one week, this is where the frontier models really shine: Its no more 2024 or so, where they lost context on large discussions, you can do really great work with the frontier models - the local models will be a thing in a few years, if RAM & CPU developed further, this will happen: in 2035 like, you will be able to run "Claude-like experience" on your iPhone.
can you give example, prompt require context window > local model, or higher Quality/Speed, that local model can not do, open source example?
Any program I'm vibing.
I write 5-20 paragraphs of things I want and it goes away and just builds it.
Something like this (just an example):
"Build an application with a webserver in rust. It has 4 stages, first one is an input module... second one is refinement ... third one is an image detection with current fitting ml model ... last one is finetuning."
Then i would describe every basic page with a few more paragraphs.
This would work for half an hour or an hour, consume 100-400k of context tokens and then I would test it, write down 10-50 small sentences (every sentence either fixing a bug, or making it better) which tehn takes again like 20 minutes and i'm nearly done with a complex enough tool providing real value and actually working.
A big kimi, which is hard to run, might be getting close to this but we do not talk about running it on 1-2 5090.
The other thing is, thats more like a practical experience, even if you use them through Open Router (for testing and co), the initial results of something like Opus 5.5 are ALWAYS better. More details, better quality, better understanding of usability of the tool.
I might be misunderstanding but I do the same thing with sonnet, and it runs faster and does well enough
The discussion was about local models vs. frontier. I would argue that Sonnet is frontier?
And yes Sonnet 5.5 is crazy good
Here's a prompt I used recently:
> In a work tree build a feature where each column cog action menu has a "Column statistics" item - clicking that item opens a modal dialog which displays statistics about that column - for text columns that's most common values, stats about value length - for numeric columns it's more interesting than that. Let's do a bit of percentile analysis in there too. Under the hood it works by firing off SQL queries using SQLite window functions and suchlike. For the moment let's have that option only available if SQL queries are allowed against the database, and the SQL queries are built in JavaScript.
You can see a screenshot of what Opus 5.5 built here: https://github.com/simonw/datasette/pull/3038
> The people saying that local models can easily replace the real Frontier Models from Anthropic or OpenAI are just not using it for the same things you and I do.
> They seem to be happy with their small usages.
Oh yes, we're terribly inferior people. Just living on scraps from the masters' tables.
It wasn't meant as a dismissive comment.
I do plenty of experiments of things i always wanted to do and they just don't work with local models.
Finetuning of some ML models for a specific usecase, some 4k demo idea i had, small browser games to show others game mechanic ideas i had for ages, analysis tools etc.
Even just playing around with some Ghidra eats context like nothing.
[flagged]
I really strongly agree with this statement. A bunch of people aren't even engineers, and their abilities aren't that good either, yet they keep saying that open-source models can replace OpenAI or Anthropic's models.
If that were really the case, basically nobody would spend so much money using them.
Putting aside that you're "really strongly" agreeing with a personally directed insult:
You should factor in that undeniably most people don't have the means to run even a 27B dense model locally, and therefore do not actually know what they are and are not capable of in that context.
And you might want to look at openrouter's rankings.
If you are comfortable automatically assuming that "doesn't use Claude or Codex" is equivalent to "not a good engineer" or "bad engineer", then I think you are probably lost to more nuanced discussions.
With my DGX Spark + 5080 + 3060 setup, I've tried all kinds of approaches, including quantization, acceleration, fine-tuning, post-training, and everything else.
It still can't compare to Opus 5.5 max in terms of speed, capability, and stability.
And when I calculate all the money I've invested, directly subscribing to the service would actually be cheaper, more efficient, and more stable.
Open source is great, but don't glorify it.
The architecture of the code I develop has never been simple. If you're someone who doesn't even work on things that need to be deployed to something like k8s, then basically, open-source models are indeed enough for you.
You can sell your DGX Spark for basically no loss. Maybe even make money.
I try local models regularly on 2x 4090 and struggle with context length constantly. 48GB on GPU and 128GB from RAM.
the original comment mentioned only 16GB of VRAM.
You know talk about Open Router and 27B dense model?
We are not comparing an even more expensive Hardwaresetup to run local models with proper context window and parameters.
Can i run some open model on a super expensive Mac Book or co with some kind of okayisch results? Yeah sure but you know what the t/s will be and yeah Frontier is still Frontier and Kimi, GLM, DeepSeek are not Frontier. They lag half a year.
You are not getting Opus 5 or Sonnet at home for now.
Why are you so offended by this?
I'm not offended. I'm also not the person resorting to insults.
Ahh yes. Those are the only two options.
Have you tried Codex? My journey so far
- 2023 using Cursor
- 2024 using Cursor with Agent mode
- 2025 using Claude Code
- 2026 using Codex with Astra 6
I've try every new model when it comes out from both OpenAI and Anthropic. I cannot justify leaving Codex and Astra 6. The token burn on Anthropic is always astronomical for me. Codex just works, and I see no reason to move away from it.
This matches my experience, Claude limits with Fable blows up the entire 20x usage in like an hour, whereas Astra can run a couple days.
I switched from Codex to Claude Code w/ Opus 5.5, and it's been fine. I haven't hit limits in my $100/mo plan, doing full time dev work all day. But I agree. Codex sub + Astra is also great. Both are excellent options, and I don't see a reason to go with a cheaper / less capable model, frankly.
12GB GPU is useless. The good experience is with 128-256GB unified memory architectures like strix halo etc. You don't need to buy any hardware, you can experience the latest open weight MOE (flash) models from DeepSeek, qwen, or GLM on openrouter using a local agent like kilo.
A 3060 + good chunk of system RAM can run qwen3.8 flash next.
That is a shift of comparisons though, right? This thread was about running locally, explicitly as an alternative to cloud services. I would guess few people have graphics cards with > 12GB ram.
A few more will have Macs with 48GB RAM though. (Indeed if you push it a little, 36GB unified is possibly enough for Muse Glimmer; I have not tested this theory)
There is, for sure, a limiting factor in the x86 architecture, but I think the RTX Spark will change that a bit, and maybe AMD's subsequent Ryzen AI/Strix Halo chips.
AMD Ryzen AI/Strix Halo is for running locally on your hardware. Openrouter is for evaluating the models you'll be running locally, before buying the hardware.
In any case, 128-256GB unified memory systems running MOE models is the "cheap" option, RTX6000 is the step above it.
> 128-256GB unified memory systems running MOE models is the "cheap"
I'm not familiar with these. How much does one cost, and from what store / what product name?
Would you talk to me a mo about Kilo compared to Pi?
I am digging in on Pi (in lovely Paseo and exploring Pendant when in VS Code).
What would I find better or worse in Kilo?
I think the key to making it make sense is that people have wildly different demands for quality. You will see people say that these small local models have improved quickly and are "good enough for most work". While these models have improved and very fast, it still might be more of a comment on their expectations than the product.
12GB is not a lot, you are stuck on low quantisations of older models with that. What exactly did you try?
The newer/larger models are good though, not as good as Claude, but not every task requires a model of that level. Maybe yours do, but for simple coding tasks, summarising text, etc a small local model runs just fine, and is a lot cheaper than Claude
What graphics cards have more than 12GB, and how much does it cost to buy one? What about for others saying you ideally want 32GB? (That split will determine if a 5080 will do it)
32GB would probably get you qwen3.8-27b at a reasonable quantisation, which benchmarks similar to opus 4.6
You can try that model (and many others) using a cloud host first, and then only if you like it do you need to buy something to run it.
> What graphics cards have more than 12GB, and how much does it cost to buy one?
RTX 5070 Ti and RTX 5080 both have 16 GB.
RTX 5090 has 32 GB.
Good luck finding any of them for reasonable prices, though. I paid $3,000 for my 5090 early last year. I winced when I bought it then. Today, it's like $7,000, and I wouldn't even consider it at that price.
But I'm mostly a gamer that also does local AI as a hobby. If the RTX 5090 was $7,000, I'd still be on my RTX 3080 and still just not started doing local AI and have to turn down the graphics settings in Cyberpunk 2077.
The RTX 5090 is the ideal, but it's not the only 32GB GPU that can offer workable performance; the AMD R9700 is catching up.
You can surely build an LLM server with an R9700 and 32GB of mainboard RAM for less than the 64GB M5 Max Apple Studio, which is probably the next most competitive model up. It'll be better than a Strix Halo but less flexible.
Of interest:
Compared to $20 or $100 per month. My thought process reading this thread:Baseline of $20 or $100/month Claude or OpenAI subscriptions. Alternative posed of running models locally. A guy says he did it and results are lackluster on a 12GB video card. A number of people post that that card is insufficient; use 32GB+ cards. I look up prices of 32GB cards, and they are all very expensive. So from a practical perspective, people are suggesting you spend thousands of dollars on a video card to get results that are almost as good as the subscription plan they are on. It doesn't add up, unless one of my assumptions is incorrect.
The assumption that is probably incorrect is that the current prices will last. A 4090 should not cost that much, and most of your Claude usage is paid by an investor. Neither of these things will be true in a few years.
Having a 4090 at home running agents all day is not a $20pcm level of usage, so take the $100 plan. Looking at the S-1 filing, today they are not even close to breakeven. It is reasonable to assume that will become a $300-400 pcm plan.
The launch MSRP of a 4090 was $1600. There has been significant inflation since then, so say $2000. Eventually supply will meet demand and you'll be able to buy one for that price again, or lower due to the economies of scale they must have now.
So, in 2-3 years time once it has all settled, you pay off your 4090 by saving 5-6 months of Claude usage. Maybe the local model isn't good enough for complete replacement, but all it needs to do is drop you to the $20 plan to be worth it. I think this is a large part of why investors are getting jittery now.
Right. Nobody says it's necessarily economically the best option. But it's also an asset that has alternative applications and barely depreciates!
Right now you could extensively use an M5 Max for a year and sell it for almost exactly the same amount at the end. Maybe even more, at the current rate. And it can do other things for you: host your VMs, edit high def video, whatever. Even if it stops being useful to serve LLMs, it might be useful to run as a server for other tasks for a whole decade. It will be a fast, supported Mac for seven years. For the R9700, maybe gaming?
Depends what you want, doesn't it? Or are we all expected to conform?
For me personally, right now I have a secondhand M1 Max macbook pro, and that is my local AI machine, because my needs are exploratory and I want to run it here. I want to learn by comparing models. I want to learn how to run them and advise others how to (local inference is not a zero-sized market).
My next step up will be to take the apps I'm looking at building that run against these small models, and host those apps online with some sort of cloud inference — openrouter, digitalocean, cloudflare — for those same models.
If I ever find I want to vaguely describe some desktop application before bed and come down the next morning to an aggravating slop version of it written in Rust, I'll consider Claude.
For a local non-laptop, I've looked at the M5 Max and the R9700. I've also looked at the M5 Pro, which in practice is about as fast at token generation as my M1 Max (because the M1 Max's memory bandwidth is strictly theoretical), and smokes it at prefill. That alone might be enough for a "leave it running" mini AI box, and it will be a nice desktop machine. I just don't need it enough yet.
As an aside there's this thing — somewhat interesting, though constrained:
https://www.olares.com/one
Uses the 5090 mobile chip, which is pretty good speedwise (about the same as the R9700) but 24GB. Small, tidy, on-prem. I think they have missed the boat on pricing, but if they can ship in volume, maybe not.
Great insights! And I'm kicking myself for not buying a 4090 3 years ago for $1600, shipped direct from nvidia... it seemed like too much to pay for a GPU at the time!
I know a couple of people who have eminently LLM-capable GPUs more or less doing nothing when they are at work…
3D games give me motion sickness so I don't have powerful PCs anymore. In fact right now I don't think I have a desktop/server PC that can power up at all, and that alone makes me wonder about building an R9700 machine.
there is also a version of the 5060ti woth 16GB. i got lucky a few months ago and got it for ~380€, but i think you can still get it for around 800, definitely under 1k
[dead]
In general, more memory = larger models = more intelligence.
However even if you had a private data center you wouldn't be able to beat Opus 5.5, Astra or Gemini 4. The open weights models are a tier below these three.
Memory manufacturers are running 80-90% margins. Any of a number of things (anti-trust, CXMT, financial shock) have the potential to drop memory cost by 2/3s. At the same time memory efficiency of models is improving. At some point Apple's going to start shipping 64G of unified memory in mainstream machines that are advertised as local AI capable (like HP, Dell and others do today). Like the iPhone refined the concept of a smartphone for the mainstream, they could be the first to hit mainstream sovereign AI.
As amazing? I suppose not.
Good enough? Quite probably, unless you're in the "coding is solved, specify loops and walk away" frame of mind.
32GB (or say a 48GB-unified RAM unified machine) gets you Qwen 3.8 27B at a good adaptive Q4, maybe Q5. Muse Glimmer at Q4 and Q5, etc.
These models can code well enough in an agentic way, I find, and the chat-to-assistant scenario is just fine in Unsloth Desktop or LM Studio.
A slightly lighter harness helps; a lighter system prompt is entirely possible, especially for Muse Glimmer, which is post-trained so effectively for agentic helper tasks that it barely seems to need a system prompt at all except for a "plan" mode.
Tool use is all fine, all that.
> unless you're in the "coding is solved, specify loops and walk away" frame of mind
which is where the entire software industry is going
Well it's certainly the service-dependent (addicted!) model that the big two want.
But will you want to do that when your tokens are metered at non-subsidised prices, and yet there is absolutely no clarity (ever!) about token efficiency, how the model decides to spend your money, how it might ever decide to stop wasting your money and get to the point, etc.?
The specify-loops-and-walk-away model absolutely points to local/on-premises/private cloud AI.
Because otherwise you're letting Anthropic or OpenAI decide how to spend your money. And they are going to be biased towards spending more of it.
I don't think Qwen 3.8 is quite there yet with the loops stuff. Definitely seems to be better at deeper and deeper one-shot puzzles.
Who cares what the big 2 want? Have you taken a look at how much the hardware/algorithms/efficiency are improving? Those 2 are building sand castles in front of a tsunami.
uhhh mission critical software is nowhere near turning over full control to _any_ LLM...
12GB GPU is insufficient for most purposes yes. Even 16GB is a notable improvement, but yes 32GB cards such as the 5090 or R9700 are a world of difference.
Given we're discussing a cost comparison (i.e. cutting a Claude subscription in lieu of running locally), how many months of claude $100 or $20/month plans would it cost to buy a 5090? I am showing it as 88 and 440 respectively. I.e. a 7 and 36 year payoff. Not even close to being worth it.
The Claude plans are probably still heavily subsidized. Expect prices to increase, especially after the IPO. Self-hosting models for personal use is probably overkill, but I can see companies doing it pretty soon since they can probably run 24/7 workloads and fully utilize the hardware.
Great point. And, 3 years ago, a 4090 cost $1600, so maybe that will happen again some day. (Same/similar card, or a future card. I remember when GPUs used to get better every few years instead of staying roughly the same but costing several times the original amount!)
Currently, it's not worth it because:
"SemiAnalysis, a research and analysis company specializing in the semiconductor and AI industries, found that a $200/month subscription plan is far more generous than previously thought. According to its calculations, while it's widely believed that a $200/month plan maxes out at roughly $2,000/month in tokens (assuming API pricing), the actual numbers are closer to $14,000 for ChatGPT and $8,000 for Claude. "
https://cybernews.com/ai-news/chatgpt-claude-usage-costs-ope...
This will change in the future.
You also get to keep your data private (so you should compare it to api pricing), there are no forbidden topics (i just did a pentest on my webapp), and multiple people can use it. Basalt does 150 tok/s on a 5080
Reminds me of linux gaming 10 years ago. According the to the community everything worked great without issue and any issues were hand waved away as tiny bumps that you won't even notice.
Then you would go try it and be back using windows to game a week later.
But that is not exactly a linux issue. Games built for linux works very well on linux like games built for windows running decently well on windows. Linux can also run some games built for windows - not for linux. And that are the ones that doesn't provide first rate experience. Those windows games executables technically should not even load as it is not built for linux.
Everything considered, linux is good and has decent capabilities including gaming and whatever you can throw at it but only do oranges to oranges comparison. Games and apps built for each os works fine on the respective os, well atleast it is true for mac and linux. Windows is very unreliable.
Imo, this is just the tech version of virtue signaling. Everyone knows local models are objectively just worse than frontier/cloud closed source models for most commodity hardware setups, yet people insist on using it because they want to be contrarian or "stick it to the man."
Your experience is average because it _is_ supposed to be average. The people using local models tradeoff taking the moral/technological high ground vs actually using something high quality to get a job done quickly and with high quality, and largely autonomously. They are ok with subpar output because they will handhold the model more, or the (subpar relative to the frontier) outputs are sufficient for their toy use case while being cheap and accessible, with the added benefit of being able to run it on their private data without sending it to a 3rd party.
Almost no one at any job where software is the main economic output is using a local model to code. If you actually require intelligence to get a hard job done, you should just use a frontier model.
> If you actually require intelligence to get a hard job done, you should just use a frontier model.
Or, you know, your own intelligence.
What I do not understand as an engineer is how our community has gone from believing in open source for controlling destiny, security, etc., from avoiding unnecessary corporate capture (think of all the rants about Swift!), from believing in heterogeneity and portability, to just saying, look, if you're not using the product of one of two loss-making AI unicorns to do all of your work, you're just being a contrarian weirdo who is doing it for performative reasons.
I guess if that is your definition of contrarian — that everyone else is doing it so why aren't we? — then I am.
But in my experience it's not so different to "why aren't you on social media?" — which was an argument from insecurity and the desperate need to fit in.
I personally believe in independent thinking. I have things I want to learn about how all this actually works, I have questions about my own life and future, and I'm not going to simply chuck them away because you or anyone else thinks it's just being contrarian and we shouldn't question the direction everyone is taking. So I am learning the tech with a stack I control, while continuing to do the job I do.
You can do what you want. You're probably right, and I'll be quite happily wrong and continue in my wrongness even though, based on how you characterise it, this kind of independent thinking appears to unsettle you.
I can't really do anything with this comment without a couple of specifics, such as 'what model are you using?'.
Because local models are significantly worse? I'm not so sure what your question is. What else did you expect? If you could just build Claude Code at home, Anthropic would be valued at closer to $1.1M than to $1.1T.
[dead]
So for a sense of scale, un-quantized 1 parameter == 1 byte of ram. 12Gb means you could have roughly 12b parameters, frontier models like Kimi k3 has 3 trillion parameters, so you'd need ~3tb of vram.
32 would get you a good Qwen 3.8 27b which is like Opus 4.6.
> The results are ALWAYS terrible compared to CLaude Code.
How much effort did you spend to make the local setup work well?
Local models are not plug and play, you need to adjust settings to make them behave intelligently.
On my side one model looking super dumb right after install ended up being 'agentic' after using the right chat template, fixing how tool calls are setup, fine-tuning plenty of params values, and adjusting the whole setup to my own hardware. Claude.ai (free) was great at helping me sort that all out.
12 GB isn't a lot. You need a lot of VRAM or RAM for context window.
I really strongly agree with this statement. A bunch of people aren't even engineers, and their abilities aren't that good either, yet they keep saying that open-source models can replace OpenAI or Anthropic's models.
If that were really the case, basically nobody would spend so much money using them.
With my DGX Spark + 5080 + 3060 setup, I've tried all kinds of approaches, including quantization, acceleration, fine-tuning, post-training, and everything else.
It still can't compare to Opus 5.5 max in terms of speed, capability, and stability.
And when I calculate all the money I've invested, directly subscribing to the service would actually be cheaper, more efficient, and more stable.
Open source is great, but don't glorify it.
The architecture of the code I develop has never been simple. If you're someone who doesn't even work on things that need to be deployed to something like k8s, then basically, open-source models are indeed enough for you.
>Sure I only have a 12GB GPU
Found your issue.
I (all other issues aside) honestly think we have lost all sanity when the term "low" (in any fashion) can end up associated with 1.1 TRILLION.-
There is a serious, malignant, dangerous disarray somewhere in our callibration of value, of worth, of what things are worth. Something, many things, are wrong here.-
I think you’re talking about the headline, and it’s clearly not really calling it ‘low.’ that said, things are relative, and relative to the market’s position on AI and AI related spend, anthropic’s overwhelmingly optimistic spending commitments, how they predict their machine god will take over the world, etc. - 1.1 is indeed “low”
Ratios determine what is low or high, not the aggregate number. With an annual revenue of 68 billion the number is 16
I don't trust the market to calibrate "value" the same way that I subjectively define it.
But I trust the market to calibrate value about 10,000x more than I trust any person or group of people to do so. The thought of that is frankly terrifying.
The market is a group of people, and much of that decision making is in the hands of a very small proportion of it.
To be more precise, the reason I trust the market is the "skin in the game" and the decentralization / localization of decision making. People with something to lose tend to actively "vote" only on what they know something about.
Not that it's somehow "not people" (of course it is).
> People with something to lose tend to actively "vote" only on what they know something about.
"Prediction" markets have rendered this clearly false of late.
> the reason I trust the market is the "skin in the game"
The market expects economic bailouts and other assistance during downturns, because individual investors with 401(k)s demand it, despite the perversions of market incentives that result.
Yeah, a billion used to be a unicorn.
Now people are disappointed with a thousand times that.
And there are worries about the Democratic midterm election sweep causing tighter scrutiny of new data centers and forcing Anthropic to pay wholesale for chips, or worse.
You'd think restricting build-outs would ease pressure on component supplies.
[dead]
can someone explain to me the point of the 'initial valuation', considering ultimately the price will settle lated thorugh trading to perceived value anyway ?
None of the value changes in later trading mean squat to how much actual money the company gets to put in their bank account.
They get the money from the initial purchase of the shares, and never get another dime directly from the market again, unless they have another offering (which is quite possible). The stock price does drive leverage for later M&A and debt/credit activities, and matters to stock-based compensation, so it does have meaning, but the IPO provides a one-time cash infusion at the initial price.
They can issue new shares. It's not like it doesn't matter at all.
…which is why the person you responded to explicitly said
> unless they have another offering (which is quite possible).
?
The point is that is the price the private investors who will be liquidating, including the company itself, are accepting for their shares.
If they are too far below the perceived value, they are leaving significant money on the table.
If they are too high, the institutions who participate in IPOs will let them know. If the price craters off IPO it is not good for perception of the stock, and so ongoing health of the company even if that maximizes private investor value at IPO.
So they build a book of commitment around the target value, trying to leave enough incentive (15-20%, often) of an initial bump to get big money in.
If the company is valued by the underwriter and initial buyers at $1 trillion and the company offers 10% for sale through their underwriters (who have put out feelers to try to attract buyers at highest price possible), they raise $100 billion from the offering, minus fees, to use as needed to further the business. The initial buyers get the shares at that valuation, and I believe if the underwriter positioned it wrongly they might be on the hook to sell some shares at a lower price without prejudice to the company.
If one of the initial buyers sells in the first minute the company is public the first publicly traded shares at 1.5x their purchase price, the company suddenly has a market cap of $1.5 trillion instead of $1 trillion, and that buyer has made a tidy 50% profit on the shares they just sold. If all initial buyers dump a bunch of their shares the price likely will go down, perhaps pushing the market cap to, say, $750 billion. Regardless of this action on the first day and after, the company offering their shares still has the $100 billion.
Note that initial buyers are discouraged from selling large portions of their shares.
Some people buy on "day 0", before it hits the stock exchange. This determines the price they pay on day 0. For example, if they buy 10% on day 0, they pay $130 billion.
The stock exchange is a "secondary market", which means you're buying stock from investors who own it, not from the company itself.
Investors ‘subscribe’ to an IPO. A company offers a certain amount of shares for sale at a certain price. Then the underwriting banks go out and sell the shares.
If your broker allocates you shares, you pay the IPO price as does everyone else who was allocated shares. The shares are all allocated before the IPO happens and once it opens for trading, price discovery happens.
How the price is set is another matter, one I’m not familiar with as an amateur.
Typically in an IPO, the offering is about 20% of valuation of the company at the first offering. But with a $trillion or more, they will offer say 10% or less.. say $100B in one go is probably what the market can take.. Some banker has to back it up initially and defend that price... the next dump will be about same amount say $100B again but not the same number of shares. The underwriters are not obliged to back $1T and the IPOed company is not obliged to release more shares.
Ultimately the percentage of shares that public will trade is material. say 20% for $100B is quite different from 10% for $100B.
> can someone explain to me the point of the 'initial valuation'
Most of the institutional and retail placements happen on the initial valuation. These are the people who are holding the shares when it gets listed and starts trading, and it is the primary source of raising capital through an IPO.
To go further though the Initial Valuation (usually a range) is set weeks before the IPO to make it easier to pitch to investors. And the final valuation is locked in the night before trading based on actual demand.
The opening price on the stock exchange, though, will be whatever the demand and supply is in pre-market trading/open.
The first seller has to decide a price to ask. The price of the first share sold x number of shares is that seller's "valuation" of the company.
Whenever you buy a business, you have to have some idea of the valuation.
The IPOs need to have buyers. The buyers will have some idea of the business valuation, as will the sellers. This is decided before the actual stock starts trading.
Specifically, the "underwriters" will go around finding people willing to buy what is being offered. This gives some idea of where it will begin to trade.
It's the starting listed price. For an anchor, it determines a lot about how the first few days go. It's also a great headline. It's also also what employees picture when they evaluate their stock options, I bet.
An IPO is effectively a funding round, and the numbers we're discussing here are the amount they'll raise from the IPO.
> the numbers we're discussing here are the amount they'll raise from the IPO
No, the "valuation" numbers are the market cap at the moment of IPO. The amount raised is that multiplied by the percentage of shares issued by Anthropic (as opposed to by existing shareholders like investors and employees -- that combined number is the float) at the IPO.
Sorry, yes, that. The bigger point still stands, though — given a fixed fraction of shares they're offering, the valuation gives you the amount raised, and that makes the number interesting.
Everyone is explaining what initial valuation is without explaining the point of initial valuation. The point is that existing private investors get to exit with a sizable profit while leaving the risk of future growth to the public market.
My understanding is that you want a small pop when the company IPOs.
If the pop is too big, that means the company left money on the table.
If the pop is too small, the investment banks that underwrite the IPO have to worry about losing money.
Some benchmarks to compare:
Apple: 5 T
Google: 3.5 T
Meta: 1.55 T
Microsoft: 3.97 T
Berkshire Hathaway: 1.2 T
Eli Lilly: 1.28 T
Tesla: 1.38 T
Walmart: 1 T
Spotify: 125 billion
Crazy to think they'd still be worth more than Walmart. At least to me.
Walmart is in a mature low-margin business. Their global expansion has stalled, and in the US competition is increasing
There are also many positive things about them. They are huge, and innovating. But their growth potential is somewhat limited
I think it's crazy that Walmart is worth 1T. It's grocery chain operating in limited number of contries. Shouldn't be worth much more than for example, Carrefour.
Walmart is huge compared to Carrefour. They're the world's largest employer, and their yearly net profit is $22 billion compared to $300 million.
Carrefour appears larger because as you travel the world you will see them in the major city in every region.
Walmart, on the other hand, has a store close to most small towns in the US, as well as multiple in more metropolitan places.
A lot of non-tech non-AI companies have seen outsized gains this year as people use them for AI hedging.
I think about what it would take for a small group of smart people to do what they do. Walmart? Seems hard. Meta? I could see that.
I know that’s oversimplifying and probably naive, but huge physical operations seem like more of a moat.
You are probably saying this because you are coming from a tech background (my guess)
For people with a Walmart background, Meta is black magic
Not that crazy, software sales are global from day 1.
I know quite a few paying Anthropic customers here (Norway). Not a single customer of Walmart.
Hei!
When you mention Tesla, don't forget Space-X. The value proposition of Teslafanboys was always that Tesla is not a car company and now Musk splits it up between Tesla and Space-X
Tesla + Space-X: 3T
Did they consider doing a Dutch auction, like Google did in its IPO[1]?
[1]: https://www.cnbc.com/2014/08/19/es-took-off-but-the-auction-...
This is not unexpected. Anthropic revenue growth is stalling. Ramp's AI index shows spend on Anthropic peaked on July 31. The spend in the last week of October is lower than last week of July -> https://ramp.com/data/ai-index?metric=token-spend&detail=mak...
Token volume is increasing though -> https://ramp.com/data/ai-index?metric=token-volume&detail=ma...
Companies are cutting down on AI cost (adding per person budgets, using smaller models). Anthropic has also been forced to reduce token prices due to competition.
Damodaran placed a fair value of $1.47T minus discount for really bad corporate structure. So $1.3T looks exactly right.
And its kind of circular way of reasoning. As the investors would've actually watched the video and included it in their calculations.
[dead]
$1.3T is still looney tunes for a company that literally has zero plan to turn a profit.
I don’t agree with you. Their product is a fully scalable almost human tier intelligence. I think it’s worth way more than $1 trillion but they have to get their capex in order first so they can provide dirt cheap tokens.
If the company cant convert the product into earnings it really doesnt matter what the product is. It can be a magic machine that turns water into wine. They still have to find people willing to pay for it.
If the world is not ready for fully scalable almost human tier intelligence then, being right at the wrong time is the same as being wrong.
Uptake is amazing, isn't ChatGPT the most downloaded and beat TikTok easily?
Also you don't have to make people pay. Facebook doesn't charge end users for its products and neither does Google for either search or Android.
its the most downloaded but they churn users quite a bit
So it's going to be full of advertising and will sell user days as the primary two ways it will be profitable?
I mean, that does sound very much like the future of our current timeline, unfortunately.
> They still have to find people willing to pay for it.
There’s a shitton of companies willing to pay for it.
> get their capex in order first
how are they gonna do that when the half-life of a GPU is 18 months?
Spot prices for 5 year old A100s are still going up recently, because inference is so valuable. This meme that GPUs are a quickly depreciating asset is outdated.
Their lifespans are 2 to 4 years with always on, heavy usage. Even then the problem isn’t the lifespan. It’s the lack of supply right now. I assume with more built out production capacity GPUs get cheaper too. The whole vertical stack needs to get cheaper and it will in time.
"fully scalable" except it requires massive capex to build out data centers with limited actual resources to put in those data centers.
> Their product is a fully scalable almost human tier intelligence.
"Almost" being a synonym for "not".
It's already much smarter than most humans in a lot of ways.
It's not the same kind of intelligence as humans so it'll probably always have weird areas where it can't do as well (like counting letters in words), but that doesn't mean it isn't smart. If AI were sentient it could point at humans and say "haha look how dumb they are; 99.999% of them can't even write a basic Lean program".
We're way beyond the point where "it's not as smart as humans" is meaningful.
Related:
Major Investors Are Skeptical Anthropic IPO Can Achieve $1.5T Valuation
https://news.ycombinator.com/item?id=50386597
I am retired and use index funds for my 401Ks. I hate the idea of Anthropic becoming part of mandatory purchases into retirement funds. I hope it doesn’t happen.
That said, Anthropic is an interesting company. I mostly use local models like Qwen 3.8 Next Flash, but when I occasionally use Opus 5.5 for difficult I enjoy the experience.
I've been trying to generate some product videos with opus 5.5 and I use the API because I don't like subscriptions.
I burned about $150 of tokens over a few hours of chatting, although opus 5.5 did generate some 10-30kb html files at a time, containing the "videos".
I think subscriptions are extremely subsidized still and what happened yesterday was just a small dip of what's to come.
You can just cancel the subscription...
Dear god, you can use sonnet for that.
opus was the only one who did well, argon too
you don't even know my specifications, why are you parroting something useless?
skill issue. i genuinely can’t think of a video you’d need to use opus for. unless your specification is “make it look cool”
Ask questions instead of making half-baked assumptions. Learn something new everyday.
I am not being hyperbolic or deliberately obtuse. I (a person) can not think of a reason to need to use a large model for video gen, but in no way does this imply you are wrong or that I somehow know better than you.
You are of course welcome to tell me why you need to use opus vs sonnet.
> i genuinely can’t think of a video you’d need to use opus for
then maybe it's on you?
The moat really is shallow to non-existent.
Both Anthropic and OpenAI are trying to build more tooling to try and lock companies into their ecosystems.
But I can’t help thinking that they are impacted by the very thing they have created.
You can now build your own enterprise tooling using the very models that they have created.
My impression is that none of those systems seems to be efficient. We see time after time systems that solves most 80% of the case to be 90% cheaper. And the way it is, it seems to me that nor openai nor Anthropic can keep the current prices.
Some say that if Anthropic keep delivering improvements and be the first it will make it but for that, you need to disregard all competition from other llm companies and from other competitor systems.
And even though they have the smartest model ever, if someone offers something 90% of it for 90% cheaper it will be worth it.
We already see plenty of companies reducing ai cost because they don't see enough roi
1.1T is 27% less than what they are aiming at, and yet, it’s still absolutely mind boggling.
People love to reign in AI companies these days (and part of it is warranted) but the triple digit annual increase in revenue at Anthropic and others (SpaceX being in that category as well) is something never seen before.
Some people are being too quick (or too willing) to disregard the obvious immense economic impact of AI and company (and I’m not talking about the next-day-ai-crash-incoming economic impact).
Classic mistake of confusing the usefulness of a technology with the first movers ability to extract long term value.
Anthropics S1 numbers are hugely optimistic given the commoditisation that has hit, continued to hit, in Q3 and the temporarily hold on chip supply that is not going to hold for too long
Inference is projected to continue to become increasingly commoditized so you need to wrap it in compelling and differentiated product people are willing to pay for.
Training, people argue, continue to be very capital intensive so moat there. What I'm wondering, is anyone trying out ideas the lower the training cost materially? I'm not thinking distilling because that, by definition, would always be behind the frontier. Other ways?
Really though we all know how this will play out. Google, Amazon, Microsoft are going to buy out both labs and we will all be centralized again
I think also the demand for software tokens is going to be bursty, and many things are going to be one-time software ports.
- Nervous about Redis being written in C? Switch to Garnet.
- Bun ported from Zig to Rust for $165,000 at API pricing -> https://bun.com/blog/bun-in-rust
- Microsoft ports Copilot runtime to Rust for $120,000 in API costs without blocking other development work -> https://github.blog/ai-and-ml/generative-ai/migrating-the-gi...
- Someone made a "clean room port" of Total Annihilation in golang, looks like it started in August and is now the most convenient way to play Total Annihilation. https://github.com/nanolathe-gg/nanolathe
- You only have to decompile / reverse engineer an existing game once and now you can have an agent occasionally add features.
Once you do the port and have worked out the regression issues, assuming you are switching to a safer language with better guarantees, you are now off of the old codebase and don't have to pay the costs of the old codebase. It's just a one time burst of token usage.
Costs of tokens will continue to fall as more hardware either becomes rentable (datacenter comes online) or moves into the used market, as workloads move onto users individual machine (because smaller models are becoming more capable and getting better inference engines and harnesses).
You can also task an agent with researching for "list the top 5 performance improvements we might be able to make" and have it research overnight, then the next day start a fleet of agents iterating through the findings that seem the most workable . Once you gain the performance win, that's basically permanent.
I think that token usage is a very elastic market, you can choose to spin up a fleet of agents or... you can stay human-in-the-loop, or you can type code yourself if you don't have a token budget.
So, yeah, I don't think Anthropic or OpenAI are worth much when hot on their heels are open-weight models hosted by major cloud providers + Nvidia and cloud providers themselves building their own models.
Still too high
Dupe: https://news.ycombinator.com/item?id=50394906
Still overvalued.
Same thing I was going to type! I think they’re just testing these to see the reaction. Like a reverse auction while OpenAI watches in terror.
I want to see this go out at $800b. I do not believe even $1T will be supported by public markets unless they shrink the float — but that also comes with its own risks.
I’m angling for it to go out at $250b.
Most finance people I’m talking to think $200-400B is a fair price. That would be an amazing story in any normal situation but a total disaster here. However sane fundamental financials just don’t support a higher number.
Challenge is while you can play games with the float size to jack up the price as existing shareholders start selling you have to convince more and more people of a sky high price when all the fundamentals and the market pulling things in the opposite direction. Employees have been royally $&@ked in prior IPOs as that story plays out.
$250b would imply the company has a $110b enterprise value. Anthropic has $140b on its balance sheet already.
I’m expecting that to decrease, not increase.
At $250b it would be an acquisition target.
I think that’d be appropriate.
Anyone have a sense of what a valuation like that would mean for employees trying to cash out?
Means nothing, until the expiration of the lockup period (where employees cannot trade stock). That's typically 90 days, or longer.
At which point, the stock is well-traded on the open market and has found a "correct" market price, with no necessary relationship to the valuation at IPO.
1.1T is still insanely high. There is absolutely no reason why I can not use any of the OK models as long as I have reasonable hardware ti give a ok token rate.
Here is a video overview of Anthropic's IPO filing by a finance expert. [0] https://youtu.be/PfjsIfXms4o
It’s going to drop further, probably well below $1T. Still a multibillion dollar valuation though imo.
Their numbers are shit. They are brittle af. Any shift in political climate results in internal mayhem.
Same with openai
A good rule of thumb is one a tech company attaints a >$1 trillion valuation, it tends to hold in that range; e.g. see Amazon, Appl, Google, Meta, Space-X, Nvidia, etc.
So many people 6-12 months ago predicted Space-X stock would crash and burn, yet the stock is still holding strong around $130 and a valuation over a trillion dollars.
> So many people 6-12 months ago predicted Space-X stock would crash and burn, yet the stock is still holding strong around $130 and a valuation over a trillion dollars.
That's simply because they're 1) installed in the government gravy-train, and 2) part of the AI bubble. Check again in a month or two, after AI valuations become more clear and Democrats start to publicly sketch out plans for what they're going to be doing to 2028 and beyond. They've already openly announced revenge for the half-billion in political spending, and they've haven't been talking at all about going after speech on twitter (which was a huge loser), but instead going after real stuff. Going after Space-X is an easy call.
No f500 is going to sign up to use chinese models.
Until there is a large free US model, its Ant and OAI.
Many F500 are already using Chinese models or models derived from them through American providers. Even Microsoft's Decision-1 Jev-like:
https://commandline.microsoft.com/microsoft-decision-1-model...
Cursor's models most famously are built on top of Kimi:https://techcrunch.com/2026/03/22/cursor-admits-its-new-codi...
So yeah, Chinese open weight models are already in F500.The major cloud providers have Chinese open weight models in their catalogs. If the model is running through your cloud provider on US soil, you're probably fine for most of corporate America.
*Chinese hosted models, not Chinese models themselves
F500 already outsource everything in the name of cost efficiency, but you're telling me, with the opportunity to cut token costs 40x, they're going to draw the line here?
If any non-US company can use US models, they can use Chinese. Chinese are way less hostile to all of us (regardless on which continent you are) than US is.
This whole fearmongering campaign may work well within US but once you cross the border it becomes funny parody in worst ways. When all this comes from current US its beyond ridiculous, I don't expect post-midterms behavior to be any less petty and dumb.
> Chinese are way less hostile to all of us (regardless on which continent you are) than US is
Who is the “all of us”?
Not the UK
https://www.independent.co.uk/bulletin/news/china-security-t...
or the EU
https://data.consilium.europa.eu/doc/document/ST-11631-2026-...
Definitely not Taiwan, Japan, South Korea, Vietnam, India....
"Their numbers are shit"
Scaled from 0 to $65 billion in revenue in 5 years...
If you gave me $130B I can find a few ways to scale to $65B in 5 years too
I’ll do it for $120B
And scaled to an annual burn rate more than double that in 5 years...
SpaceX had insanely great timing. They came in and ate up a lot of the available capital at the most prime the they could.
Antropic will be fine I think but open AI will probably have issues
Space-X and Tesla has to be a very very weird mix of person cult and fluke.
Tesla was what everyone valued crazy a because Tesla is not a car company, now Space-X is doing the same s*it as Tesla.
The general economy/inflation etc. and cost o fmoney is for sure not a good sign. Lets hope, i def want to invest in Anthropic.
Don't think SpaceX is a cult of personality mix. Making rocketry cheap is how they got to where they are.
Space payload before Space-X and Starlink was 1000t and increased to 3000t because of starlink.
To get enough payload into space that the evaoluation is worth while, he really need to do this datacenter in space thing and sending enough hardware up there to match just one GPU based datacenter would take one Starship launch every single day.
Lets ignore that building Datacenter on earth is stll A LOT easier and cheaper to do and that we shouldn't vaporize non sustainable metals into our atmosphere alone for health reasons.
We have robots, AI (which is so bad that he prevers to rent out his AI hardware instea d of making more money with it directly), 3 Car models and something else combined in two companies with an overall value over 3T.
This setup is still higher valuated as plenty of big car manufacturers, robotics companies and a lot more other companies thrown into the mix together.
What am Im missing here?
You are missing mostly that you are blinded by your ideological position.
I talked clear numbers, you don't even try to argue.
I actually calculated and researched these numbers and while it seems that you are dismissive because of your ideological position? i saw and read Elon Musks claim about AI Datacenters in Space and him writing about a dyson sphere and calling it bullshit.
You are not evaluated as the 7/8th most valuable companie on the planet just because you build rockets and send up a little bit of payload. Thats why their P/E is 3 digits.
They take up 95 plus percent of the world’s payload. “A little bit of payload” is a completely ridiculous statement
Its not if the absolut number of worlds payload is not that much and unironically Space-X made it A LOT cheaper. So they took the market, reduced the size of the market.
Over a 100 launches a year, most of them for starlink doesn't sell you the valuation the company currently has. And they have to keep sending up starlink satelites just for renewale of the existing ones. This blocks Space-X of having good margins on it. A normal telecom provider builds a fiber underground and a tower and that stuff just stays there for ages.
And yes you know what happened when people realized this is a thing? Yeah more companies are going into the market. Check out how many will aunch their first rockets in 2027.
can someone explain to me the point of the 'initial valuation', considering ultimately the price will settle lated thorugh trading to perceived value anyway ?
The initial buyers of the shares sold by Anthropic are institutional investors (mostly). That matching process is done by an investment bank they hire.
The trading price you see on the screen is - shortly afterwards the shares become publicly tradable on the exchange (or on all US equities exchanges, for US stocks). So in effect, the price shooting up on the public markets is actually bad for the company in that they might have left money on the table.
Sort of. IPOs are always undervalued because the entire block of shares gets purchased by the investment bank at the banks decided IPO price but then bears the risk of selling them at or above that price.
It's the price they're going to be offering to the general public to buy ahead of time, through their brokers. They get that price locked in, then once it starts trading it moves as it may. Usually the IPO buyers are not allowed to sell for some time
The price those that buy the IPO get before it starts trading publicly.
E.g. SpaceX is currently trading below its IPO price. If you became a shareholder by buying at IPO and held your have lost money as of today.
The price will settle, but it has to start somewhere.
Well obviously Anthropic as the one offering the shares would care about how much they sell them for.
Some benchmarks to compare against:
Apple: 5 T
Google: 3.5 T
Meta: 1.55 T
Microsoft: 3.97 T
Berkshire Hathaway: 1.2 T
Eli Lilly: 1.28 T
Tesla: 1.38 T
Walmart: 1 T
Spotify: 125 billion
Anthropic sells intelligence, and the most intelligent intelligence.
Honestly crazy that we’re doubting 1T, even with all the capital issues behind it.
Validation is all about assets and future earning potential. Anthropic’s core assets are a model which is shown every day now to be only marginally better than something that’s free. So that asset appears to have very little value. Their future earning potential depends on folks not figuring out they can get a good-enough result for free, only paying for compute. And Anthropic doesn’t own compute, they rent it.
So the whole valuation is based on the idea that folks will happily pay this company a huge markup for inference over the actual cost of compute for a model whose valuation is rapidly tending to zero. Hence why potential investors are starting to freak out.
In competitive environments, such as most businesses, marginally better is enough to create large advantages.
I just don’t see a world where everyone is happy buying the second or third smartest model. The people using the smartest will just out compete them, even if it is marginally more expensive.
> Anthropic sells intelligence, and the most intelligent intelligence.
Most intelligent is a stretch, we don't know how the Chinese architectures would fare if they had the same compute availability as Anthropic and other US labs.
If we extrapolate and consider that they are behind due to artificial export restrictions, once they figure out how to build their own GPUs in a cost-effective/scaled-up way then the American labs' claim of having the best intelligence is at an enormous risk.
They already do a great job being resource constrained, imagine once this constraint is gone!
I mainly doubt the moat, not necessarily the usefulness. Why will customers stay with anthropic? Many cheaper models seem to catch up within 6-12 months, and there’s also other flagship providers that keep pace and compete on cost (oai, Google). Why should I believe that anthropic’s revenue is going to stay consistent, and not evaporate to other competitors?
This is a very competitive market. You can't command fat margins in a very competitive market.
Huh? If a Chinese open weights model comes out that is nearly as good, the valuation drops to very little.
If the justification for the valuation is "they sell intelligence", then their valuation should be $0. Because their product isn't the least bit intelligent, and routinely gets beaten by humans.
You can be as dismissive as you want but this doesn't reflect the reality.
I can easily spend a LOT less time on describing the same amount to claude than i would in a JIRA ticket and doing code review on the PR of Claude than talking to a person about it. Instead of 1-4 weeks Claude is done in an hour.
Its also already A LOT faster and better of analysing infrastructure issues.
Conservatives' undermining public education makes it inevitable that slop engines will surpass the intelligence of the average human, then double and treble it as the average human's intelligence continues to decline.
Probably also a huge problem but at least in the last 15 years of working, there was always at least 1 if not 2 people were i question how they were even in the industry to begin with and getting jobs.
I think people in this thread who claim local models are anywhere close to Claude are completely detached from reality. 200$ Ultra sub is the best value ever offered in history of subscriptions. I got 3 running and would pay at least 3x more for them before looking at alternatives. If your local models or some DeepSeek on cheap provider or whatever you install instead of Claude makes even 2-3 more mistakes per week you are already losing money assuming your time is worth average software dev salary.
I don't think it's remotely close either - Claude is that good. The problem is that it's not how Anthropic plans to make money to justify 1.5T-2T valuation. That would require companies paying API prices on long term big contracts and that makes it a completely different value proposition.
The argument is not that they are as good, but that more models have reached the threshold for being useful. You do not need the best for most tasks these days. Freedom from closed model vendors also has value.
This is why token prices are falling and there are expectations for commoditization. In another year or two, you will have Opus 5.5 in much smaller models, assuming capability distillation remains as good as it has been from mega -> mini models.
Yeah I get it but I don't think it hurts Anthropic sub offerings. 200$ or 500$ or whatever it's going to be is low enough that even slight edge is worth paying for business use case. They will likely go into local proprietary models as well (we train our model on your data and you run it on your premises so you meet data retention regulation).
While I am a big fan of open models and I hope I can run local Opus 5.5 level model locally in the future I will still be using Opus 7.5 or whatever it's going to be for anything business related then. Again, the problem here is that they are not making enough from those subs and API is 20x+ more expensive.
Business do not get subscription plans as they are largely not available, certainly Anthropic is not offering them under the contractual obligations businesses want. It's all per-token.
Then you have to consider automations and user facing products. A quota based subscription is not enough or risks being shut off after too many tokens are used. These are the primary places open weights are disrupting closed, expensive models.
Eventually, token vendors will be commoditized and the average user will have an effectively unlimited plan, but we are not there yet.
Please explain to me why my "naive and maybe ignorant" understanding of this is wrong... The bigger the IPO price, the more they earn and the more probability is for you to loose money invested in it.
So say Anthropic, or OpenAi, or smartGummyBearsAI go IPO and the price is insanely high... and then spirals down to be 10% of the real price... they earn a lot of money and the public loose a lot of money.
IF smartGummyBearsAI goes 100% realistic price and you buy near the correct price, they earn their fair money and people get an investment.
IF smartGummyBearsAI says that they have technology to use the same machine that makes gummy bears print gpus and they are priced 10,000x their actual price the society just transfers a lot of wealth to them and they scam the society.
Why shouldn't we regulate and aim for the correct price?
Remember 801k money is what we are talking about
The price isn't "set" from what I understand. The price is determined based on what institutional investors are willing to sign up to pay for it. So it's largely driven by the market (not on the actual stock market, just the market of interested parties pre-IPO). Those institutional investors do not want the scenario you describe any more than you or I do.
That's why it's dropping here. When the actual stock market tanked yesterday institutional investors were like - hey wait a minute this is starting to look too expensive.
Also if your 401(k) is in index funds you are safe. A company is not added to indices right away. If you are in actively managed funds or something like that you might have more exposure risk.
How do you determine the "correct" price? Something is worth what people will pay for it, no more and no less.
Markets have the advantage that many peoples' opinions are aggregated into the market price. This is far, far more accurate than having one person's opinion (an elected bureaucrat) set the price, particularly when they have no skin in the game.
If the IPO price is really 10x the public market price, that implies that the set of institutional investors collectively estimated the price 10x higher than the full set of retail investors. That's pretty unlikely: institutional investors usually do pretty extensive due diligence on IPOs, and you're already averaging out the statistical fluctuations by selling to a wide set of them. But if that does happen, it just means that institutions got fucked over to the benefit of retail (who gets a 90% discount on the same shares), which seems like something we should be celebrating.
A price to sales ratio of 16 isn’t terribly crazy
1T is not low for an IPO
using claude code with pro subscription and deepseek api. claude is project manager and deepseek is developer. combo seems pretty effective and efficient.
Polymarket put a valuation at around $1.7 trillion
anthropic futures on binance dropped again today. why don't they just do a direct listing and then atm share sales ?
It keeps dropping. Was 2T, then 1.5T last week, now looking like they might be lucky to get 1T.
I wouldn’t be surprised if Anthropic delays the IPO in hopes this will all magically sort itself out but as every day passes it looks like they should have gone public this summer when things were still on the up.
Realistically the company is worth in the hundreds of billions but if it goes below 1.1T or so existing investors will need to start booking losses and that will trigger a shitstorm of other issues downstream.
This is all becoming a textbook case of folks know these assets aren’t worth what folks want them to be, but they can keep the game going so long as nobody actually marks the asset at its actual realistic valuation.
So that’s the game underway now. How low can they go without blowing up the whole AI ecosystem. If the valuations here drop next the AI bonds and private loans start to fail en-mass and the whole thing that built up this house of cards starts falling apart in reverse. Get ready for a bumpy ride folks.
I wouldn’t be surprised if they pull the plug on the IPO and go hide under a privately funded rock.
The second one of these AI companies gets a fair valuation and folks start booking losses the whole AI house of cards will come crumbling down. The crazy thing is if they drop below around 1T some existing investors will be booking losses.
A realistic valuation is likely in the hundreds of billions, which would be a total disaster and thus they’re going to do everything possible to avoid that.
literally what is the upside? They're supposedly worth as much as meta despite having nowhere near the customer base or years of work? How much more valuation do they have to trick people into giving them before investors get a return?
Anthropic hit 9 billion ARR in Dec 2025 and 80 billion in Oct 2026. Meta hit 9 billion in 2014 and 80 billion in 2020. It is now at 250 billion. At this growth rate Anthropic will be bigger than Meta already next year.
Caveats: It's hard to believe that this growth rate would continue. And Meta's numbers above are trailing 12 months, while Anthropic's are annualized run rate.
We already know the growth rate doesn’t continue, their q3 is a massive slowdown compared the 26H1.
Also, ARR as in Annualized Run Rate is a bullshit, non-standard metric only used to mislead investors
> Anthropic hit 9 billion ARR in Dec 2025 and 80 billion in Oct 2026
This is a bit misleading and based on projections of prior reported months. Growth has already slowed, Q3 in dollars was less than Q2. The timing aligns largely with the growth (since mid-summer) of open weight models.
https://www.regenerator1.com/p/anthropics-growth-just-flatte...
$80B ARR with an ~$160B net loss.
some of you are acting like earth is the size of jupiter and there's 800 billion humans not 8
World's best or next best LLM, at least for code.
Since they actually have paying customers, it's a completely different sort of thing than for Meta. The question is instead whether they can keep at the frontier, whether the compute they've signed up for is good enough, etc.
What on earth are you smoking. Meta has lots of high paying customers, their revenue last year was 200 billion, roughly 44x Anthropic's revenue.
Ah, I simply hugely underestimated how big Meta was.
Why does years of work matter? Serious question.
I think this is down to how investors think about how to value a piece of a company. I think about it almost the same way as I think about sports gambling.
There's a value proposition in which direction the stock is going to go and how long it's going to take to get there. It only makes sense to be invested in a stock if you think it's going to be worth more later than it is now and the amount of time your money is tied up in this stock is commensurate with the profit you may get as a result.
Investors work on limited information about companies, and having a track record can influence a buying decision one way or another. It may matter to some investors and not to others. For many, it's weighed against the price of the stock. Would you bet $100 on a company with a short track record? What about $10,000, $100,000, or more? Generally there's a point where people drop out of buying more of a stock and say that they can't justify the risk given the information they have. That's the expected value that Anthropic will receive from that investor.
Anthropic needs to worry about what that number is for a huge amount of individual investors and how to maximize the amount of money put into the cost of their stocks, and there's very little chance that they can attract the biggest pool of investor money without some of those people considering their longevity vs a more established company like Meta. It may not matter to all investors, but the fact that it will matter to some makes it a factor in the total valuation.
Possibly speaks to its moat being weak, if a dedicated team from scratch can duplicate their efforts and get to where they are today in 2 years (but not where they will be).
Because building a stable revenue base takes time. We've had many flash-in-a-pan type product but things which still make money 10 years later are rare.
It's not work, but the Lindy effect makes Anthropic's outlook in 2032 very uncertain.
Its the growth potential and the product they already have.
Companies are now building the agentic layer, the agentic layer eats tokens. Everything you do every process will have tokens eaten inbetween.
Your HR Request? AI Tokens. Everything.
That plus Meta's Muse Wonder LLM is very close to Opus 5.5.
Its not. Perhaps in Metas own Benchmarks but not on the public listed results.
Do you know how companies are valued?
It is strange that they have this problem. I thought they achieved superintelligence. Maybe they haven't prompted it right yet.
They have a product which can do A LOT of stuff A LOT better and A LOT faster than A LOT of people around me already.
I'm lost on how you can be snarky while others already are loosing their jobs because of Anthropic/OpenAI
My point is that Anthropic is not special. Neither is OpenAI. They are just really good at marketing.
Job displacement is real, but based on real world usage it would owe just as much to Grok, Gemini, DeepSeek, Kimi, and whatever else that is close to frontier, but cheaper.
Your original comment was just snarky though :P
Sure that argument is a lot stronger. But the market might be big enough for Gemini, Anthropic and OpenAI.
Grok is a lot worse then Frontier and still is owned by a company for very little trust.
The chinese model are intersting, they do leverage Frontiermodels to train, this might be a moat for a while.
The / hints at the problem.
If in a couple years they're losing their jobs because of Anthropic/OpenAI/10+ other enterprise token resellers using open models then their value proposition is much weaker even if people are still losing jobs.
Yes could be but we don't know yet how big the overall market will be.
If you have in every single industry AI tokens, it will replace wages. In USA alone thats 15 Trillion. If its just 10% the USA market alone are 1.5T without the value added.
That is the only future where the valuation makes sense. But I'm afraid to think about how a society where that[1] happens will behave. My guess is that it would become _highly_ unstable.
[1] Say >10% of people of employable age suddenly becomes superfluous with not much prospect of finding new jobs (since AI eats those too).
do you think we live in the alien universe or something? I seriously sometimes try to argue with some of these comments but jesus christ
If you give me ANY idea of why you think my argument is stupid, we could actually discuss it?
My company is building an agentic layer in its product, google and the other hyperscalers do that too.
I already want to get rid of one person in my team because he was slow before but was able to contribute something, now I want to have his position as AI tokens.
you think 10% of the entire usa market is gonna go to anthropic and I'm supposed to somehow find a way with words for you to understand how stupid that sounds?
You are still not discussing or talking in good faith at all.
No i don't say that 10% of the entire USA market is gonna go to anthropic.
I said what the market overall could look like. USA is not the only country out there.
Why has Walmart such a high evaluation despite them only being in USA?
Why is Apple at 5 Trillion? They only make Laptops and Smartphones, everyone else is doing this too.
Why is Space-X 1.8 Trillion? Because of 10 Million Starlink customers?
I think AI is a disruptive technology.
The OP is snarky but he makes a valid point: if Claude is so intelligent, why can't it just "do something" to get whatever valuation its owners want? Because that's what we've been told, this is AGI, no it's ASI, it will kill us all.
Currently the progress is real and there is no ceiling in sight. If you just follow the line, there will be / should be your scenario coming up sooner than later.
But no one knows if there is a ceiling or progress might slow down massivly.
Is warning people so wrong if you interpolate the current progress and just look ahead a few years? Might just be as stupid as some North Korean/Russian hacker throwing some abliberated model against some computer system and having some cascading effect on water or energy grid.
For plenty of people AI already took their jobs, jobs they liked and enjoyed doing.
Get your tickets to Layoffpalooza 2027 now!
Has already started with companies offering buyouts.
I fire U, electric boogaloo.
The future is not local models because by the time you think of what to pipe through the LLM, the super intelligent AGI owned by Anthropic or OpenAI (but not both) would have already mapped it out, generated it and had it in production. You can, however, still pursue living as a hobby.
Simple as. I hope this helps.
The Facebook valuation some 10 years ago was like… 300 million? I mean, I get it, but jeez, when I see headlines like this “as low as(!)”, I need to pinch myself.
And after they went way out of their way to state in their S-1 that they would never dream of valuing the company at $1 more than it was worth?
> Anthropic IPO buyers see valuation as low as $1.1T
AS LOW AS?! This wording has to be intentional... Because I would use "as high as", myself.
Gotta give it to OAI for waiting. They smelled it.
Well they’re struggling to raise cash via new rounds, having slashed both the raise amount and valuation themselves. So in that situation they may go bust before having a chance to go public.
OAI is in the middle of another funding round at $1.25T. If Anthropic with allegedly better financials can't get above $1.1T then why would anyone buy OAI at $1.25T
Or then any numbers they would have publicly presented with even previous scrutiny would have been so bad no one would have bought them at price they want.
Is waiting better? Surely it would be better to be early, before faith is lost? I think being late to a party as it’s deflating is the worst case
My hunch is that it’s better to be late in this scenario because now your moves are under public scrutiny or at least neck-deep. People can pull out, it could slow them down and at the end look worse.
As a privately held company, it functions under different dynamics that in the environment we’re seeing feels better to stay private for longer.
Compare to SpaceXSI, they lack missiles, datacenters, and a profitable internet business so... yeah.
As "low" as 1.1T. What a world we live in where 1.1T is considered "low"...
Seems like a great argument for anchoring high in negotiations.
Low ???
Still at least $1.1T too high.
People seem unnecessarily deluded here in a supposed hacker forum. If they are really created actual digital intelligence smarter than every human and figure it out how to control it, then 1T dollars means this is the most undervalued business in the entire world. Maybe by 100x or 1000x. You have billions of super genius slaves in data centers that do whatever you want, then if a single human developer can earn $100k/year pessimisticly let alone a super genius one. Multiply that by a billion and do the math at 80% gross margins
…and what if they haven’t? Given that we haven’t seen evidence of one so far.
You’re also missing the competition aspect. What if other companies have their own super genius slaves to sell at a cheaper price? What if open models allow anyone to have their own moderately genius slave at far less expense?
Yours is a reasonable estimate at the _value_ of their product. They will only be able to capture that value if their competition cannot catch up AND they can reduce spend on R&D. I personally do not believe they’ll defeat distillation.
> If they are really created actual digital intelligence smarter than every human
Whoever is claiming that is an idiot.
If a frog had wings...
You are making claims with no basis, caveat-ed by "if," and then calling everyone here fools for not basing their beliefs on your wishful thinking.
Every market cap is a prediction of future profits. If they had trillions in profits every year already, then the price wouldn't be 1.1T, it would be 100T and we wouldn't have this discussion. 1T is a option to buy it, and it's probably a good bet even if there's a 10% chance that it happens.
What if they have billions of super genius slaves, but Chinese companies have just slightly less super genius (and catching up) slaves for 1/4th of the price, while all prices are already dropping?
I see nothing delusional in speculating that that’s a very big if.
And if they haven’t created an artificial superintelligence then they start to look more like just another tech startup with no moat[1] and a terrible balance sheet.
[1] I don’t know about you, but I switch model providers multiple times per day.
Every market cap is a prediction of expected future profits. In their mind, I'd say that's what the whole "race" is that they are so desperate to win. The race to win is the race (in their minds) to get such powerful beings to get a permanent lead with ASI. Getting to powerful enough beings to wipe out all the competition and become emperors of this new world.
But we're so close to RSI!
A lot of people have had it for decades. It really hurts.
"as low as $1.1T"
the absurdity
wasn't 1.3 just a couple of hours ago and 1.5 yesterday?
that's low enough Musky could just buy it out right, jeez.
As high as $1.1T. Where is the supposed growth going to come from?
Claude has been pushed by force and coercive marketing into most software organizations and many open source projects. In 2026 the companies buying the narrative were exuberant and told employees to tokenmaxx.
This will stop in 2027 because companies will find out that nothing has improved. Nothing at all. They'll evaluate the expenditures.
So the market is saturated and usage is going down. Anthropic should be valued at $200 billion at best.
Usage is not going down. Take a look at their published revenue, it has a nice hockey stick shape.
I’m not making a bigger argument other than I disagree with that one datapoint you mentioned.
When you sell people 6 quarters for $1 they tend to buy a lot. Useage is a poor metric here given the capital intensive nature of this business. What matters is can they do this in a way that makes money to the degree that justifies the valuation, and there they have not yet demonstrated this is possible.
The internal metric that they are using is they get $2 for every $1 in inference. So, if you discount model development costs, they are selling the tokens for more than it costs them to generate a token. And this is before they start integrating their advertisment networks into the model. They see something as high as $8-11 per $1 of inference once purchasing and advertisments are incorporated.
The “but ads” logic fails basic sense checks. There isn’t enough ad money to go around to make up the gaps. With the big ad players like Google, Meta, and Amazon all shutting Anthropic out of their ecosystem there’s very little evidence that Anthropic can steal away enough of that as revenue to save itself.
In the “but they’ll slow down training and so will become more profitable” thinking is in big trouble given the pace of open models. They’ll then be even less competitive with free options and so even less reason to buy from Anthropic.
There's not enough ad money? What? lol. there will always be ad money as long as a click costs below my profit. Every single one of those companies you listed was excluded from the Ad system of the ones before it. They built their own.
> So, if you discount model development costs, they are selling the tokens for more than it costs them to generate a token.
If you discount my expenses I am actually swimming in cash too.
It is absolutely moronic to discount the value of developing the thing from the profitability of the thing. Models to depreciate with time.
For sure, models definitely depreciate but so does CapEx. but you don't include CapEx, you have an amortization schedule over it's useful lifetime. However, when you start mixing CapEx and model depreciation, you get an interesting curve. As CapEx increases, model development is spread across multiple generations across that same CapEx. So, assuming you've amortized your CapEx, successive model generation eventually has a horizontal assimtote at OpEx and depreciation.
The real question is "Are frontier training costs growing faster than the productive lifetime value of the GPU infrastructure used to create those models?" Which is a hard question to answer and is something we'll find out it early 2030.
Also, Anthropic is known to lie through their teeth to make themselves look better, so any claims that they make a profit on inference should be treated as false until proven otherwise.
Any “non-GAAP” measure of “profit” should generally be viewed with a giant red flag.
In nearly all cases these metrics just say “if you ignore all the reasons why we’re horribly unprofitable, we’re profitable.” WeWork’s “community adjusted EBITDA” was a textbook example of this.
usage may shift to models run locally, albeit trained by the frontier labs, increasingly over time, though.
A floppy hockey stick.
Revenue growth q2 to q3 was significantly lower percentage wise than q1 to q2 and they swung back to net loss.
Yeah, the hockey stick is flipped in the wrong direction if you're looking at the past 4 months, plateauing after steady gains.
I introduced it at our admittedly small company not because of coercive marketing or force of any kind. I did it because rolling new stuff out takes days instead of months. Not only that but we can make products above our previous level.
Unsafe, comprehension debt, vibecoding, slop? Have you seen actual lowly developers’ code? AI is a vast, vast improvement on all relevant axes if you ask me.
If you think this is going to slow down I have no idea what to tell you.
They made a typo, it's more like $11B
1793898170 | Michael Burry: Anthropic Is "Betting Its Existence on Optimism" | https://mrkt30.com/michael-burry-anthropic-s1-commitments-op... | https://news.ycombinator.com/item?id=50381531
1793902395 | So, about that Anthropic IPO | https://www.ft.com/content/2fbd60ee-060a-4957-b9c5-3b5a0ce88... | https://news.ycombinator.com/item?id=50382331
1793918987 | Michael Burry says Anthropic is 'betting its existence' on optimism | https://www.businessinsider.com/michael-burry-anthropic-ipo-... | https://news.ycombinator.com/item?id=50385642
1793923469 | Major Investors Are Skeptical Anthropic IPO Can Achieve $1.5T Valuation | https://www.wsj.com/finance/major-investors-are-skeptical-an... | https://news.ycombinator.com/item?id=50386375
1793923694 | Anthropic warns that it's close to RSI capability, a.k.a. self-building AI | https://techcrunch.com/2026/11/02/anthropic-warns-that-its-v... | https://news.ycombinator.com/item?id=50386406
1793924267 | Major Investors Are Skeptical Anthropic IPO Can Achieve $1.5T Valuation | https://www.wsj.com/finance/major-investors-are-skeptical-an... | https://news.ycombinator.com/item?id=50386483
1793925071 | Major Investors Are Skeptical Anthropic IPO Can Achieve $1.5T Valuation | https://www.wsj.com/finance/major-investors-are-skeptical-an... | https://news.ycombinator.com/item?id=50386597
1793934067 | About That Anthropic IPO | https://www.wsj.com/livecoverage/stock-market-today-dow-sp-5... | https://news.ycombinator.com/item?id=50387632
1793965329 | Is Anthropic About to Delay the Biggest IPO? | https://www.zerohedge.com/ai/anthropic-about-delay-biggest-i... | https://news.ycombinator.com/item?id=50391165
It's not worth 500 billion
What? Imho that is still way too much, I wouldn't go above 400B, considering if z.ai is about 128B
Talk about terrible timing. Had their IPO been 2 weeks earlier they'd have hit $2T easy.
make it 100b
Imagine 'as low as' next to 1.1 Trillion Dollars.
Good drop. Should be less than a trillion amid current midterm fallout. The chances of more AI regulations are coming.
Colossus 1 & Colossus 2 are located on Tennessee and Mississippi, and Anthropic rents both of them earlier this year.
Now that Democrats are surely going after Trump and Elon Musk and the sidekicks, and both Tennessee and Mississippi turned blue as well, this directly impacts the way Anthropic to source cheap compute and data in the future due to uncertainty around the continuation of SpaceX data centers. And other data center buildouts are likely to be paused for things like power grid connection and environmental impacts and land purpose evaluation, the compute capacity constraints are likely going south, too.
Worse, the proposals around superintelligence is exactly what Anthropic is looking for: slow the fuck down. Now they got the killshot they long desired for, it obviously makes a dent in the capability and capacity of their AI, hence a drop in valuation. I think $800-900B is somewhere I would say okay, and $600-700B is really where it is fair, considering the long term outlook for Anthropic is looming with a greater AI massacre caused by political factors.
Wasn’t the regulation main hope to sustain the value? Otherwise open models come and eat the lunch.
This is a "not like that" situation. They wanted "AI safety" legislation guaranteeing them rent money. What they're likely getting is legislation to slow down data center buildout and make them more environmentally sane (more expensive upfront cost).
And open models are coming whether they like it or not.
Oh, and if the Dems take OpenAI and Anthropic's AI safety rhetoric at face value, nationalization and other forms of substantial government interference is on the table. A predictable outcome of irresponsible fear mongering!
Nothing ever happens. Elon and SpaceX will be fine.
[dead]
[dead]
[dead]
I don't see how Anthro is worth 15x( Gemini estimates based off WSJ source )as much as Deepseek( 76 billion) .
If anything is going to pull the rug out, it's going to be a model from Deepseek or another Chinese lab. I'll admit I'm very uncomfortable with so much of our economy riding on AI hype.
To be fair to China, if AI is really this revolutionary technology, they need to be able to have home grown peers.
Every nation ( or bloc of nations) needs alternatives to American tech.
[dead]
Anthropic intends to rule humanity with its ASI. Perhaps this is the optimal outcome, as there are many problems AI could fix that it wouldn't be able to as efficiently without eliminating the human bottlenecks which caused those issues, but I fear the more financially powerful it gets, the fewer checks there will be on any misalignment.
What's ASI?
Artificial super intelligence
As opposed to the CCP, which is essentially what the Chinese model companies map to, who are only interested in freedom and liberty for humanity right?
If they release weights, then no, it's not concentrating power and control in a single company like the vision of Anthropic and OpenAI.
Straight from the CCP playbook. Clever girl.
I don't think being ruled by ASI is by default safe whether it's Chinese or American.
Agreed.
You think it's a good idea to trustmebro some Silicon Valley jerks with world domination? I... don't.
Perhaps $50B would be a fair valuation to aim for?
Imagine that - 50 (fifty) times more than an average unicorn start-up!
Something they could be very proud of, and many if not all employees would become multi-hundred-thousandaires practically overnight!
Reportedly Anthropic's ARR is $80B and gross margin 85%, so $50B sounds low.
Sure, if you ignore all RnD costs. If you don't; they were $80B in the hole the first half of 2026.
Almost all of that loss is accounting charges due to valuation increases, not R&D or other expenses.
No
A trillion in debt and no profit. It ain’t worth a trillion. It has no moat and no guarantee of future revenue. It doesn’t own anything.
50b would be a fair valuation.
[flagged]
and I behind 100M, in that case everyone who buys if they are worth more including them will get more right?
No - the company is initially offering shares, so a higher IPO valuation means that the company gets more liquidity.