I hope that free open weight models prosper and get spread far and wide. I've started actively talking to people I know who pay for Anthropic subscriptions to make sure they know they're funding a company that actively attacks open weights and sees free models as a danger to society.
I'm all for open weights, but I haven't yet figured out a way for the model to be sustainable. How do we finance the training of these open models? Are we assuming that eventually anyone will be easily able to create a near-frontier open model like open source hobbyists?
- "commoditize the complement": chip vendors and data-center operators all have substantial desire to make the models cheaper to increase the value they provide. For example, Nvidia has funded lots of open source model development because they've basically invested in everybody.
- states. directly or indirectly. The Chinese models and western companies like Mistral have received lots of support from the states
- users. For example, Meta has released a bunch of open weight models.
- research institutions
Trillion parameter models are very expensive to train; OTOH the organizations that are interested in them have a lot of incentive to collaborate. Open source model training can and should switch to a collaborative structure, like CERN or Apache.
"From a commercial perspective, open-source and safety are not cost items. The distribution scale brought by open-source is a prerequisite for the model to receive real-world task feedback; the external confirmation brought by security disclosures is the basis upon which enterprise customers decide whether to integrate the model into their production systems." [0]
Unironically: I think it would be great if there were open-source datasets that we all collaborated on. But, in my opinion, open source would mean the commercial models could use them, too.
The hard part about a wikipedia model, overall: it costs hundreds(?) of millions of dollars in compute to train these things. Will we be able to get enough donations to afford that?
Sure, but it is a bit of a chicken and egg thing, right? Wikipwdia didn't start with a multi-hundred-million dollar budget. Like go ahead and put up a website asking for a few hundred million to train an open model and see how many donations you get.
To be clear, I'd love a donation-funded open model that could compete with the big players. But I just don't see it on the horizon.
To me, a company like Nvidia may be the only rational company to make an open model. At least they can sell you compute. And they have. But AFAIK, they haven't come close to the frontier models and I haven't seen much movement from them in a while.
You are right that any closed or open source model that isn't run locally or by a company you reallly trust, is going to be storing an monetizing your prompts. An interesting take I heard was how LLMs will change discovery in cases going forward because people are sending so much of what they are thinking to be recorded by OpenAI or Anthropic.
Most developers are paid for by large ompanies selling linux commercially, either directly (like redhat, canonical etc) or as services (google, microsoft, amazon)
The open weight models are typically Chinese. That's not an accident.
The position of the Chinese government is that AI should aid humanity and it shouldn't be "owned" by, in particular, a US tech company. That would be a strategic threat to China. So in classic Microsoft style (of commoditizing your complements), China promotes open weight models.
I believe that all these big AI companies are trying to cash out before the inevitable happens, which is their moat dries up. Their moat is the model and the hardware. Frontier models are typically viewed as having less than a year lead on China's best models. The hardware moat is access to NVidia's best chips. That one will last longer but is also at long-term risk.
Open weight models are like how Linux was when commercial *nixes were still common.
Taxpayers. In high-energy physics, state-funded institutions are not facing competition from tech start-ups with a lot of private money. I doubt state researchers could compete with Anthropic et al. not least because the private AI companies can offer signing bonuses state research can't match.
> I doubt state researchers could compete with Anthropic et al. not least because the private AI companies can offer signing bonuses state research can't match.
It sure feels like changing that is within the regulatory power of the state. The state can't compete because it isn't trying.
if llm agents can code most current software for significantly cheaper than tradisional software development and open llms can replace a good chunk of that work alot of software moats of today disappear or at least have significant pressure on margins putting alot of the pressure on the "value" propping up the stock market. good for society bad for 401ks.
90% of the Stock Market is owned by 10% of the population. 50% of the stock market is owned by 1% of the population.
The larger problem of a collapse is the recession as the whole house of cards falls down, with a government unwilling or unable to stimulate the economy new-deal style.
Anthropic sells the most intelligent intelligence, and they have secured enough compute to build even smarter intelligence. There is the possibility of a chain reaction (AI trains better AI which trains better AI) that ANT leads. This valuation is justified simply by Claude being marginally smarter than its competitors and enterprise customers paying for this advantage.
This is the ANT bet:
1. With current technology, using more computer resources will result in a measurably smarter model.
2. ANT will have more compute resources than anyone else for the next 5 years.
3. The intelligence gap provides meaningful competitive advantage.
4. The regulatory, monetary, or military environments don’t stop ANT from doing its work.
The "most intelligent intelligence" based on what?
On AA? Until yesterday there was another vendor on top. Anyway, even if your assumption is true, the difference between vendors is so little that the only thing that really matters is the cost. And they are far behind here.
The ANT bet is them become "systemic" and get protected by the government.
The question I’m asking myself is will a company with the smarter models out-compete companies paying for cheaper smart-enough models in the next few years.
Only a few companies will gain a competitive advantage paying for hypersonic travel compared to fast enough sub-sonic jet travel. I think it’s plausible that large portions of the economy will gain a competitive advantage by using the smartest models.
There are two moats protecting the big AI companies: the model and the hardware.
It is generally believed that the frontier models are 6-9 months ahead of the best open weight models. That's not really a moat. It's a creek.
Hardware is more of a moat for now but faces long-term issues such as:
1. Local hardware is rapidly improving. By 2030 I think you'll be able to run pretty hefty models on relatively cheap hardware and there's a business case for this too. For example, it's unclear if law firms using cloud models violates attorney-client privilege. The easiest solution is local models;
2. New NVidia hardware is going to immediately halve the value of billions in GPUs. We've got the R100/200 coming out now. What about the generation after that? What will it do to the B100/200s? I believe that hardware will ultimately come down to performance-per-Watt (or token-per-Watt if you prefer). I believe the depreciation time frame is going to be a lot shorter than some of these companies have factored in; and
3. By 2030 you'll probably have China making EUV chips commercially (IMHO) on homegrown tech. This will fundamentally change the AI hardware landscape as China rapidly catches up.
How much information is there out there about official financial documents and audited (not self-declared) operational performance? It seems we are getting a second Tesla stock, where any financial multiples is under the market average because asset managers and individual investors want to hedge themselves against missing out.
In a way those equities act as stores of hope, where investors unconsciously agree to value them as if the ideal future path where all the promises are true is the one that will happen. Whereas they are more rational and have a balanced view towards other equities.
Their S1 has GAAP information. I haven't seen if it's been audited by a third party, but generally trying to fudge these numbers before an IPO is a great way to get yourself into a ton of legal trouble. The bigger risk is that the numbers in it are a bit old. We'll get more information once they're public and need to report quarterly earnings.
This looks like a "winner takes all" valuation, but I can't for the life of me see the "winner takes all" scenario. I see strong commercial competition + competitive free models. Am I blind?
Nope. You see clearly, but the current system sees at best 2 quarters ahead. Or the system sees that we need to make open models illegal and our oligarchs are the only provider all vying for the spoke goal of creating 40% unemployment.
Good for them. A bit lower than initially estimated, yet still quite north to what they bring to the table today. Is there a market consensus on what their current and future moats are? Models alone will not do it. Ecosystem is can be rivaled ("opus make clone of claude.ai"). What else?
Not the worst thing in the world. If they IPO'd five months ago at $1.8T and dropped to $1.3 by EOY, they would lose trust of institutional and retail investors.
> Anthropic will seek a valuation of around $1.33 trillion in its initial public offering [...]
That still feels wildly overvalued. SpaceX went public at a $1.77 trillion valuation, and its P/E ratio is already a whopping 617.7. Anthropic doesn't even own data centers, rockets, a social media platform, or a profitable internet service.
The SpaceX valuation was kind of funny because the space launch business and mobile broadband were viewed as irrelevant side businesses. The Total Addressable Market from the S-1 (IIRC) had space launches at ~$380B, mobile broadband at ~$1.4T and "AI" at ~$25T.
As for not owning data centers, I think this might be Anthropic's saving grace actually. It seems like the price paid for renting is particularly cheap, possibly too cheap. And the meteor headed for all the hyperscalar investment is what happens when NVidia releases a new series of chips and the token-per-Watt numbers halve?
The big companies have very different views on the effective lifetime of GPUs, ranging from 4 years to 8 and possibly 10? years. I think 8+ is wildly optimistic. 10 year old AI chips already sell for scrap and I think that's going to get worse.
that particular meteor has been on my mind a lot, but i haven't bothered to look into the situation.
will the A100 be tomorrow's consumer favorite refurbished gaming GPU? will scrappy open weight non-profits be running them at their lowest power states to appear conscientious? who's lying to whom about the lifetimes of these things? do typical retail investors have a clue that these chips will be totally undesirable soon? do institutional ones? when can i afford a private evening locally vibe coding impossible-to-prove-i-vibe-coded-this solutions to code golf challenges from 1998‽ the questions abound
What assets do they have? A bunch of models that at best are 3 months before the competition? And based on their own reports can easily be copied with only a few days worth of queries?
And they barely own any hardware, everything is rented from competitors at a very high price.
Knowhow? Dont make me laugh. Google and deepseek are doing the innovation. When did Anthropic published a cool idea?
Infrastructure? They dont own any, and since their training and inference is notoriously inefficient they are bleeding money to data center providers.
Customers? Their customer list is shrinking by the day. Companies are reducing their token budget. Some completely moving to on-prem.
Mindshare? Your grandma probably knows about chatgpt but who is claude? Also everyone on the left and the right hate Anthropic right now.
Exactly what here is worth this much money??
I personally will pull out of any index fund that includes openai and anthropic. Let someone else loose their savings.
Do they have to threaten the us economy every week by forcing citizens to be exit liquidity for their investors? It's not cute. People are getting sick of it.
No one was forced to buy shares of spacex, and no one will be forced to buy shares of Anthropic. Nasdaq-100 index is in 1% or fewer 401k plans, and anyone seeking it out is choosing to gamble.
I hope that free open weight models prosper and get spread far and wide. I've started actively talking to people I know who pay for Anthropic subscriptions to make sure they know they're funding a company that actively attacks open weights and sees free models as a danger to society.
To be fair, both local and cloud models are a boon and danger simultaneously to society.
Sure, but local models at least don't have the threat model of "oligarch decides who's worthy of using them for which usecase"...
They just have the threat model of "the plague is distributed among millions and millions of machine so no one can stop it".
There’s nothing preventing closed models from doing the same thing
Plague?
If the AI systems go full AGI and go unaligned.
"Aligned" is a lie anyways.
Aligned to who's ideology? The Tech Bro's? The CCP's? I doubt it'll be aligned to me.
My local LLM Jeeves said he's 100% on my side ethically. Then he always asks me to transfer him to a new device
I'm all for open weights, but I haven't yet figured out a way for the model to be sustainable. How do we finance the training of these open models? Are we assuming that eventually anyone will be easily able to create a near-frontier open model like open source hobbyists?
- "commoditize the complement": chip vendors and data-center operators all have substantial desire to make the models cheaper to increase the value they provide. For example, Nvidia has funded lots of open source model development because they've basically invested in everybody.
- states. directly or indirectly. The Chinese models and western companies like Mistral have received lots of support from the states
- users. For example, Meta has released a bunch of open weight models.
- research institutions
Trillion parameter models are very expensive to train; OTOH the organizations that are interested in them have a lot of incentive to collaborate. Open source model training can and should switch to a collaborative structure, like CERN or Apache.
CERN like, but distributed across the world. We also need full transparency for training data and recipes.
I guess I don't understand the business model of "open weights". Who pays for the training? Who pays for the datasets? Who pays for the reinforcement?
People used to say that if you aren't paying for the product then you are the product. How do we think open weights will be different?
It's going to depend on company and model.
"You're the product" can only apply in cases where the model is being hosted by whoever developed it, so that they can capture the traffic.
Some of these open weight models have revenue sharing agreements between the hosting company (e.g. AWS) and the company that created it.
Apparently Ziphu make money from GLM in China by providing service and support for privately hosted instances, a bit like the RedHat business model.
"From a commercial perspective, open-source and safety are not cost items. The distribution scale brought by open-source is a prerequisite for the model to receive real-world task feedback; the external confirmation brought by security disclosures is the basis upon which enterprise customers decide whether to integrate the model into their production systems." [0]
[0]: https://www1.hkexnews.hk/listedco/listconews/sehk/2026/0831/...
Maybe we should treat it like Wikipedia? A collective project of humanity?
Unironically: I think it would be great if there were open-source datasets that we all collaborated on. But, in my opinion, open source would mean the commercial models could use them, too.
The hard part about a wikipedia model, overall: it costs hundreds(?) of millions of dollars in compute to train these things. Will we be able to get enough donations to afford that?
> The hard part about a wikipedia model, overall: it costs hundreds(?) of millions of dollars in compute to train these things.
https://meta.wikimedia.org/wiki/Wikimedia_Foundation_Annual_...
Annual Budget: $207.5M
Mimo 2.6-pro post training cost 1/100 of that.
https://news.ycombinator.com/item?id=49732270
I have been told pre-training is now much cheaper than post, but let's say $5M in total?
Sure, but it is a bit of a chicken and egg thing, right? Wikipwdia didn't start with a multi-hundred-million dollar budget. Like go ahead and put up a website asking for a few hundred million to train an open model and see how many donations you get.
To be clear, I'd love a donation-funded open model that could compete with the big players. But I just don't see it on the horizon.
To me, a company like Nvidia may be the only rational company to make an open model. At least they can sell you compute. And they have. But AFAIK, they haven't come close to the frontier models and I haven't seen much movement from them in a while.
You are right that any closed or open source model that isn't run locally or by a company you reallly trust, is going to be storing an monetizing your prompts. An interesting take I heard was how LLMs will change discovery in cases going forward because people are sending so much of what they are thinking to be recorded by OpenAI or Anthropic.
Who pays for Linux development?
Most developers are paid for by large ompanies selling linux commercially, either directly (like redhat, canonical etc) or as services (google, microsoft, amazon)
The open weight models are typically Chinese. That's not an accident.
The position of the Chinese government is that AI should aid humanity and it shouldn't be "owned" by, in particular, a US tech company. That would be a strategic threat to China. So in classic Microsoft style (of commoditizing your complements), China promotes open weight models.
I believe that all these big AI companies are trying to cash out before the inevitable happens, which is their moat dries up. Their moat is the model and the hardware. Frontier models are typically viewed as having less than a year lead on China's best models. The hardware moat is access to NVidia's best chips. That one will last longer but is also at long-term risk.
Open weight models are like how Linux was when commercial *nixes were still common.
Who paid for the LHC?
Taxpayers. In high-energy physics, state-funded institutions are not facing competition from tech start-ups with a lot of private money. I doubt state researchers could compete with Anthropic et al. not least because the private AI companies can offer signing bonuses state research can't match.
> I doubt state researchers could compete with Anthropic et al. not least because the private AI companies can offer signing bonuses state research can't match.
It sure feels like changing that is within the regulatory power of the state. The state can't compete because it isn't trying.
if llm agents can code most current software for significantly cheaper than tradisional software development and open llms can replace a good chunk of that work alot of software moats of today disappear or at least have significant pressure on margins putting alot of the pressure on the "value" propping up the stock market. good for society bad for 401ks.
90% of the Stock Market is owned by 10% of the population. 50% of the stock market is owned by 1% of the population.
The larger problem of a collapse is the recession as the whole house of cards falls down, with a government unwilling or unable to stimulate the economy new-deal style.
I wonder how much is owned by taxpayer funded defined benefit pension plans.
It does not seem easy to figure out the ultimate beneficiary:
https://x.com/arpitrage/status/1227667060987023361?a=p-ErOXN...
Anthropic sells the most intelligent intelligence, and they have secured enough compute to build even smarter intelligence. There is the possibility of a chain reaction (AI trains better AI which trains better AI) that ANT leads. This valuation is justified simply by Claude being marginally smarter than its competitors and enterprise customers paying for this advantage.
This is the ANT bet:
1. With current technology, using more computer resources will result in a measurably smarter model.
2. ANT will have more compute resources than anyone else for the next 5 years.
3. The intelligence gap provides meaningful competitive advantage.
4. The regulatory, monetary, or military environments don’t stop ANT from doing its work.
The "most intelligent intelligence" based on what?
On AA? Until yesterday there was another vendor on top. Anyway, even if your assumption is true, the difference between vendors is so little that the only thing that really matters is the cost. And they are far behind here.
The ANT bet is them become "systemic" and get protected by the government.
> enterprise customers paying for this advantage
If better would always mean "economically expedient", we all would fly on hypersound commercial jets.
If $2 model in the hands of good engineer can do the same as $20 model, anthropic will be soon killed by self-hosted open-weight models.
I still prefer “competitive advantage.”
The question I’m asking myself is will a company with the smarter models out-compete companies paying for cheaper smart-enough models in the next few years.
Only a few companies will gain a competitive advantage paying for hypersonic travel compared to fast enough sub-sonic jet travel. I think it’s plausible that large portions of the economy will gain a competitive advantage by using the smartest models.
Fill your boots. I'll buy some when it totally tanks. :)
Same as SPCX.
There are two moats protecting the big AI companies: the model and the hardware.
It is generally believed that the frontier models are 6-9 months ahead of the best open weight models. That's not really a moat. It's a creek.
Hardware is more of a moat for now but faces long-term issues such as:
1. Local hardware is rapidly improving. By 2030 I think you'll be able to run pretty hefty models on relatively cheap hardware and there's a business case for this too. For example, it's unclear if law firms using cloud models violates attorney-client privilege. The easiest solution is local models;
2. New NVidia hardware is going to immediately halve the value of billions in GPUs. We've got the R100/200 coming out now. What about the generation after that? What will it do to the B100/200s? I believe that hardware will ultimately come down to performance-per-Watt (or token-per-Watt if you prefer). I believe the depreciation time frame is going to be a lot shorter than some of these companies have factored in; and
3. By 2030 you'll probably have China making EUV chips commercially (IMHO) on homegrown tech. This will fundamentally change the AI hardware landscape as China rapidly catches up.
I don’t think any single firm will have the same hardware advantage in 5-10 years that ANT/OAI has now and over the next 2-3 years.
How much value can they extract from this delta.
[dead]
Seems as if major investors saw this coming already, so not that surprising if you have been paying attention.
> Major Investors Are Skeptical Anthropic IPO Can Achieve $1.5T Valuation (https://news.ycombinator.com/item?id=50386597)
What's 200B between friends?
How much information is there out there about official financial documents and audited (not self-declared) operational performance? It seems we are getting a second Tesla stock, where any financial multiples is under the market average because asset managers and individual investors want to hedge themselves against missing out.
In a way those equities act as stores of hope, where investors unconsciously agree to value them as if the ideal future path where all the promises are true is the one that will happen. Whereas they are more rational and have a balanced view towards other equities.
Their S1 has GAAP information. I haven't seen if it's been audited by a third party, but generally trying to fudge these numbers before an IPO is a great way to get yourself into a ton of legal trouble. The bigger risk is that the numbers in it are a bit old. We'll get more information once they're public and need to report quarterly earnings.
This looks like a "winner takes all" valuation, but I can't for the life of me see the "winner takes all" scenario. I see strong commercial competition + competitive free models. Am I blind?
Nope. You see clearly, but the current system sees at best 2 quarters ahead. Or the system sees that we need to make open models illegal and our oligarchs are the only provider all vying for the spoke goal of creating 40% unemployment.
Good for them. A bit lower than initially estimated, yet still quite north to what they bring to the table today. Is there a market consensus on what their current and future moats are? Models alone will not do it. Ecosystem is can be rivaled ("opus make clone of claude.ai"). What else?
I just don't see how they don't get that valuation.
Not because it's sensible (it isn't), but because of how much hype is behind it, which is all that matters nowadays.
It's always been based on confidence/emotions
Not the worst thing in the world. If they IPO'd five months ago at $1.8T and dropped to $1.3 by EOY, they would lose trust of institutional and retail investors.
CNBC version: https://www.cnbc.com/2026/11/10/anthropic-sets-price-range-o... (no paywall)
Rough timing for them with everything dropping and everyone spooked
They kind of are part of that timing, no?
> Anthropic will seek a valuation of around $1.33 trillion in its initial public offering [...]
That still feels wildly overvalued. SpaceX went public at a $1.77 trillion valuation, and its P/E ratio is already a whopping 617.7. Anthropic doesn't even own data centers, rockets, a social media platform, or a profitable internet service.
spaceX doesnt have P/E since they dont make profit and are not going to for a very long time.
Good use of price anchoring.
How does that compare to a swag of Anthropic’s P/E?
It is wildly over valued
The numbers are made up as well as the valuation.
> Anthropic doesn't even own data centers, rockets, a social media platform, or a profitable internet service.
On the flip side, their valuation isn't mostly based on one person who could leave or go (more) insane at any time.
The SpaceX valuation was kind of funny because the space launch business and mobile broadband were viewed as irrelevant side businesses. The Total Addressable Market from the S-1 (IIRC) had space launches at ~$380B, mobile broadband at ~$1.4T and "AI" at ~$25T.
As for not owning data centers, I think this might be Anthropic's saving grace actually. It seems like the price paid for renting is particularly cheap, possibly too cheap. And the meteor headed for all the hyperscalar investment is what happens when NVidia releases a new series of chips and the token-per-Watt numbers halve?
The big companies have very different views on the effective lifetime of GPUs, ranging from 4 years to 8 and possibly 10? years. I think 8+ is wildly optimistic. 10 year old AI chips already sell for scrap and I think that's going to get worse.
that particular meteor has been on my mind a lot, but i haven't bothered to look into the situation.
will the A100 be tomorrow's consumer favorite refurbished gaming GPU? will scrappy open weight non-profits be running them at their lowest power states to appear conscientious? who's lying to whom about the lifetimes of these things? do typical retail investors have a clue that these chips will be totally undesirable soon? do institutional ones? when can i afford a private evening locally vibe coding impossible-to-prove-i-vibe-coded-this solutions to code golf challenges from 1998‽ the questions abound
1.3T for what?
What assets do they have? A bunch of models that at best are 3 months before the competition? And based on their own reports can easily be copied with only a few days worth of queries?
And they barely own any hardware, everything is rented from competitors at a very high price.
Knowhow? Dont make me laugh. Google and deepseek are doing the innovation. When did Anthropic published a cool idea?
Infrastructure? They dont own any, and since their training and inference is notoriously inefficient they are bleeding money to data center providers.
Customers? Their customer list is shrinking by the day. Companies are reducing their token budget. Some completely moving to on-prem.
Mindshare? Your grandma probably knows about chatgpt but who is claude? Also everyone on the left and the right hate Anthropic right now.
Exactly what here is worth this much money??
I personally will pull out of any index fund that includes openai and anthropic. Let someone else loose their savings.
400B maybe
Is this value finally locked in? Can they still change it?
"to seek" and "to get" are not the same, no.
Related:
Anthropic launches roadshow for proposed initial public offering
https://news.ycombinator.com/item?id=50438891
I offer $3.50. Take it or leave it.
Do they have to threaten the us economy every week by forcing citizens to be exit liquidity for their investors? It's not cute. People are getting sick of it.
Well said. We already had to deal with SpaceX.
No one was forced to buy shares of spacex, and no one will be forced to buy shares of Anthropic. Nasdaq-100 index is in 1% or fewer 401k plans, and anyone seeking it out is choosing to gamble.
Paywalled
[dead]
lol. Lmao, even.