Do the "current stockholders" listed on that document need to file the SEC form 4 (insider disclosure) ? Or because this is technically pre-IPO they are exempt?
I would be curious to know who they are. But says something about this IPO. If I hold stock for what I think it's a good company, I wouldn't be selling on day 0
Why do their models always get Nerf'ed? Is Anthropic doing planned obsolescence or are they trying to encourage use of their latest models? Opus 5.5 was clearly Nerf'ed last week. I think they are at risk of getting a class action lawsuit for deceptive practices with this unannounced "model optimization" that actually changes performance as compared to the advertised performance in the official model cards.
It's frustrating they can't be transparent that they're even doing it, let alone why they're doing it.
They still charge you full price and don't offer any quota resets even though you're getting less value for money while you're working with a nerfed model.
>Why do their models always get Nerf'ed? Is Anthropic doing planned obsolescence or are they trying to encourage use of their latest models? Opus 5.5 was clearly Nerf'ed last week. I think they are at risk of getting a class action lawsuit for deceptive practices with this unannounced "model optimization" that actually changes performance as compared to the advertised performance in the official model cards.
You have psychosis my friend. They don't nerf the models.
I'm sure they'll do well, but I just can't get behind the investment thesis. Eventually the LLMs will plateau. Specialized expert models will dominate each knowledge area, because they can handle specific tasks without as much overhead. Then the swarm of agents will be burned onto silicon to reduce interference costs. Every X time intervals both the models and silicon will evolve, but less frequently than in the wild-west environment of today. Instead of new models every 3 days, we'll see a new set of models on a hardware release cadence. You'll be able to customize your model by mixing and matching expert model chips.
Software > hardware is a hell of a pivot, and Alphabet and Meta already do hardware. Unless they are building a huge IP moat and have proven themselves competitive with hardware, I don't see the rosy future.
The bet behind Anthropic is whether they will be the first to reach recursive self-improvement (RSI). Some experts thinks that if anyone achieve RSI, then it would be easy for them to reach superintelligence.
I don't think RSI is possible with LLMs. When they train on their own data, it becomes an ungrounded random walk into model collapse. RSI needs to generate and verify novel knowledge and avoid death spirals. RSI is going to come from the physical world, likely via robotics. It needs novel stimuli. Language data scrapes have a ceiling, and we're already hitting it.
I thought like you a few months ago! But it looks like decoder-only transformers are _really_ good at learning nearly anything, and current LLMs are not language only anymore... (part of it comes from how stupidly easy it is to scale pre-training for these things)
LLMs are also quite good at robotics for language models even if their current capabilities come from SFT and RL-maxxing they will play a lot for robotics (look at VLAs with the Physical Intelligence startup!).
There are also efforts going on to replace the old BPE tokenizer with a learnable tokenizer for every data structure possible.
It sure feels like the obvious target for increasing revenue is to charge the software engineers more. Most other users that I see are either using the tech as a novelty, which will wear out really quick once the thumb screws come out, or users who cannot explain to me to save their life what the ROI is beyond "disruption".
The software engineers are also the most likely to pivot to local inference. I know several engineers playing around with local as a hobby. After the first bill comes in that means nobody gets the Christmas bonus this year, they'll resolve the issue and start figuring out how to adapt their workflow.
And somehow I doubt that they'll need frontier for the majority of customers who just need crud and libraries hooked up. Yeah it won't be as easy as with frontier but it just means the party is over and it's time to get back to work.
So ... Maybe security audits need frontier ... And they can pass the price increase onto their customers, I presume, easily enough.
I'm not convinced thats sufficient to pay off the bill that it took to get frontier in the first place.
[And so far, I haven't met a single software engineering customer, although surely they have to exist somewhere, who was actually comfortable vibe coding their crown jewels or even anything vaguely auxiliary. Please engineer ask the magic box to generate code for us. I think they'll be just as happy checking a box they says 'save money by using agentic development' if they box is hooked up to some future price gouging frontier or hypothetically cheaper local setup]
They're selling shares but it doesn't look like they're cashing out to me. Dario is going from 3.6% to 3.2% and then also losing 2% of his voting power.
Sure, he's putting aside the possibility of being a 10 trillionaire in exchange for merely being a billionaire and then maybe being a 100 billionaire later. But, I mean at some point there's diminishing returns on having a mountain of cash that would make a dragon blush.
EDIT. Despite evidence to the contrary, I remain bearish on frontier models. However, unless I'm missing something the evidence also shows that Dario et al remain bullish. Except they're taking the reasonable route of also having a giant pile of money.
I just think it sends the wrong message. The only people who get to sell any stock at the opening are the 7 musketeers… and given how poorly IPOs have been doing it seems extremely unfair.
It's not a scam if you do it in the open and don't break any laws. Look at SPCX. A lot of retail investors are just dumb. There are people here who will go off about how great a company it is-- fine, no argument. That's not what makes a good investment.
Americans want to invest how they want. They want to strike it big and the way the world works is most of them won't. I think we need to be at peace with that and maybe adjust the tax code some.
"The proposed offering consists of 825,000,000 shares of Class A common stock. Of these, 801,351,354 shares are being offered by Anthropic and 23,648,646 shares are being offered by certain existing stockholders. Anthropic will not receive any proceeds from the sale of shares by the selling stockholders."
WTF - so basically insiders / investors are dumping their stock on dumb money retail investors / our 401ks and the company is getting NOTHING?
They're offering about $65Bn, of which about $1.8Bn is from existing stock holders. That's a tiny amount of shares from existing stock holders and probably not anyone who works at the company.
Wouldn't "by Anthropic" mean by the company itself - new shares to raise money? In that case the 23M shares being sold "by existing stockholders" would (might) include the founders, and seems surprisingly small ~1.7B at $75/share.
Maybe founders and employees aren't allowed to sell yet?
All these valuations seem nuts to me, but who knows. Maybe there's some reason big enterprise will keep shoveling money into the Anthropic furnace to avoid downloading GLM.
Past data is irrelevant given the company is completely changing its strategy moving forward. Their S-1 is very clear, they will be shifting to a capex heavy strategy to build their own compute infrastructure.
Don’t do investment based on a series of heuristics, the actual details are what matters.
Also: Annualized Run Rate isn’t a standard metric. It’s a bullshit metric that doesn’t mean anything without the actual definition the company is using. It’s what startups use to impress gullible investors, not something a serious company relies upon…
As someone who uses a Claude enterprise plan at work and DeepSeek on my personal time, I can't imagine why my employer chooses to pay for Claude Code and I assume the accountants will wake up eventually.
But again, who knows? Oracle is still making money somehow.
Do the "current stockholders" listed on that document need to file the SEC form 4 (insider disclosure) ? Or because this is technically pre-IPO they are exempt?
I would be curious to know who they are. But says something about this IPO. If I hold stock for what I think it's a good company, I wouldn't be selling on day 0
page 299 of S-1
https://www.sec.gov/Archives/edgar/data/2130357/000162828026...
Why do their models always get Nerf'ed? Is Anthropic doing planned obsolescence or are they trying to encourage use of their latest models? Opus 5.5 was clearly Nerf'ed last week. I think they are at risk of getting a class action lawsuit for deceptive practices with this unannounced "model optimization" that actually changes performance as compared to the advertised performance in the official model cards.
> Why do their models always get Nerf'ed?
It's frustrating they can't be transparent that they're even doing it, let alone why they're doing it.
They still charge you full price and don't offer any quota resets even though you're getting less value for money while you're working with a nerfed model.
they don't have the chips and need the compute
This is the answer. Old model makes way for new
And their customers seem to be okay paying
probably the only way to make their economics work
>Why do their models always get Nerf'ed? Is Anthropic doing planned obsolescence or are they trying to encourage use of their latest models? Opus 5.5 was clearly Nerf'ed last week. I think they are at risk of getting a class action lawsuit for deceptive practices with this unannounced "model optimization" that actually changes performance as compared to the advertised performance in the official model cards.
You have psychosis my friend. They don't nerf the models.
Are they saying what the target valuation is?
They’re offering 825,000,000 shares for $70 and $78 per share. That would raise about $57B-$64B.
It doesn’t say what percentage of the total shares that would be.
Is it crazy that they don't have to explain that? How can you know if that price makes sense without knowing the total?
up to $1.405T valuation on a fully diluted basis
Thanks. Can you explain how you arrived at that number?
Some more structured stats on the IPO, valuation, and compute commitments: https://www.kadoa.com/companies/insights/anthropic-ipo-prici...
Valuation at 968B plus green shoe?
I'm sure they'll do well, but I just can't get behind the investment thesis. Eventually the LLMs will plateau. Specialized expert models will dominate each knowledge area, because they can handle specific tasks without as much overhead. Then the swarm of agents will be burned onto silicon to reduce interference costs. Every X time intervals both the models and silicon will evolve, but less frequently than in the wild-west environment of today. Instead of new models every 3 days, we'll see a new set of models on a hardware release cadence. You'll be able to customize your model by mixing and matching expert model chips.
Software > hardware is a hell of a pivot, and Alphabet and Meta already do hardware. Unless they are building a huge IP moat and have proven themselves competitive with hardware, I don't see the rosy future.
The bet behind Anthropic is whether they will be the first to reach recursive self-improvement (RSI). Some experts thinks that if anyone achieve RSI, then it would be easy for them to reach superintelligence.
That's my understanding of things at least.
I don't think RSI is possible with LLMs. When they train on their own data, it becomes an ungrounded random walk into model collapse. RSI needs to generate and verify novel knowledge and avoid death spirals. RSI is going to come from the physical world, likely via robotics. It needs novel stimuli. Language data scrapes have a ceiling, and we're already hitting it.
I thought like you a few months ago! But it looks like decoder-only transformers are _really_ good at learning nearly anything, and current LLMs are not language only anymore... (part of it comes from how stupidly easy it is to scale pre-training for these things)
LLMs are also quite good at robotics for language models even if their current capabilities come from SFT and RL-maxxing they will play a lot for robotics (look at VLAs with the Physical Intelligence startup!).
There are also efforts going on to replace the old BPE tokenizer with a learnable tokenizer for every data structure possible.
The future is quite exciting!
You can make that argument for just about any field, pre and post LLM. A lot of investing is point-in-time.
I'm sure there was a point in time when everyone thought horses had a great moat.
I mean they kinda did for a few thousand years, no?
It sure feels like the obvious target for increasing revenue is to charge the software engineers more. Most other users that I see are either using the tech as a novelty, which will wear out really quick once the thumb screws come out, or users who cannot explain to me to save their life what the ROI is beyond "disruption".
The software engineers are also the most likely to pivot to local inference. I know several engineers playing around with local as a hobby. After the first bill comes in that means nobody gets the Christmas bonus this year, they'll resolve the issue and start figuring out how to adapt their workflow.
And somehow I doubt that they'll need frontier for the majority of customers who just need crud and libraries hooked up. Yeah it won't be as easy as with frontier but it just means the party is over and it's time to get back to work.
So ... Maybe security audits need frontier ... And they can pass the price increase onto their customers, I presume, easily enough.
I'm not convinced thats sufficient to pay off the bill that it took to get frontier in the first place.
[And so far, I haven't met a single software engineering customer, although surely they have to exist somewhere, who was actually comfortable vibe coding their crown jewels or even anything vaguely auxiliary. Please engineer ask the magic box to generate code for us. I think they'll be just as happy checking a box they says 'save money by using agentic development' if they box is hooked up to some future price gouging frontier or hypothetically cheaper local setup]
They'll be fine. They just need to prompt a swarm of Claudes to pivot to hardware /s
Anyone else feel incredibly left out?
In which way? If you want in you will have in after the IPO
Left out of what? You can still lose your money if you wish to do so
> Anyone else feel incredibly left out?
I'm sure there were a lot of people without enough money to invest in Madoff's scheme that felt that way as well.
> The initial public offering price is expected to be between $70.00 and $78.00 per share.
finally, a number analysts can obsess about until the IPO launches
doesnt mean anything on its own though
Am I reading this right that the 7 cofounders are all cashing out day one? Lol.
Do you have a source for that? The linked article just says “certain existing stockholders.”
Yeah sorry it’s page 299 of the amended S1
https://www.sec.gov/Archives/edgar/data/2130357/000162828026...
They're selling shares but it doesn't look like they're cashing out to me. Dario is going from 3.6% to 3.2% and then also losing 2% of his voting power.
Sure, he's putting aside the possibility of being a 10 trillionaire in exchange for merely being a billionaire and then maybe being a 100 billionaire later. But, I mean at some point there's diminishing returns on having a mountain of cash that would make a dragon blush.
EDIT. Despite evidence to the contrary, I remain bearish on frontier models. However, unless I'm missing something the evidence also shows that Dario et al remain bullish. Except they're taking the reasonable route of also having a giant pile of money.
Unless of course I'm missing something.
I just think it sends the wrong message. The only people who get to sell any stock at the opening are the 7 musketeers… and given how poorly IPOs have been doing it seems extremely unfair.
Lol, can you quote that part?
And yes, of course, the scam is almost done. Time to go the bank $$$.
It's not a scam if you do it in the open and don't break any laws. Look at SPCX. A lot of retail investors are just dumb. There are people here who will go off about how great a company it is-- fine, no argument. That's not what makes a good investment.
Americans want to invest how they want. They want to strike it big and the way the world works is most of them won't. I think we need to be at peace with that and maybe adjust the tax code some.
Normal people are not reading an S1 and don’t understand the various levers that can be manipulated. The SEC is dropping the ball imo.
SEC has 2 of what should be 5 commissioners, it's hardly functional
The SEC is working exactly as it is bribed to work right now
Griftfest 2026 is on the road!
"The proposed offering consists of 825,000,000 shares of Class A common stock. Of these, 801,351,354 shares are being offered by Anthropic and 23,648,646 shares are being offered by certain existing stockholders. Anthropic will not receive any proceeds from the sale of shares by the selling stockholders."
WTF - so basically insiders / investors are dumping their stock on dumb money retail investors / our 401ks and the company is getting NOTHING?
They're offering about $65Bn, of which about $1.8Bn is from existing stock holders. That's a tiny amount of shares from existing stock holders and probably not anyone who works at the company.
See my other comment, I believe it’s the 7 founders.
Wouldn't "by Anthropic" mean by the company itself - new shares to raise money? In that case the 23M shares being sold "by existing stockholders" would (might) include the founders, and seems surprisingly small ~1.7B at $75/share.
Maybe founders and employees aren't allowed to sell yet?
Why would the company get anything from someone else selling? It isn't the company's share.
The company is making a lot of money from the 800M shares they do own that they are selling.
[dead]
https://x.com/findatasets/status/2120145211733180753
All these valuations seem nuts to me, but who knows. Maybe there's some reason big enterprise will keep shoveling money into the Anthropic furnace to avoid downloading GLM.
what value would you assign a company that has grown it's revenue run rate $1B-$10B-$100B in 2 years?
Past data is irrelevant given the company is completely changing its strategy moving forward. Their S-1 is very clear, they will be shifting to a capex heavy strategy to build their own compute infrastructure.
Don’t do investment based on a series of heuristics, the actual details are what matters.
Also: Annualized Run Rate isn’t a standard metric. It’s a bullshit metric that doesn’t mean anything without the actual definition the company is using. It’s what startups use to impress gullible investors, not something a serious company relies upon…
It's very easy to grow revenue if you're selling $10 bills for $5.50 https://www.statista.com/chart/12520/moviepass-subscriber-gr...
As someone who uses a Claude enterprise plan at work and DeepSeek on my personal time, I can't imagine why my employer chooses to pay for Claude Code and I assume the accountants will wake up eventually.
But again, who knows? Oracle is still making money somehow.