This is just general comment to all the excitable kids over here.

Anthropic is one company in a gigantic ecosystem.

You want to understand the eco system. Got to flyover the jungle, not just stare at one elephant herd.

Read annual reports of NVIDIA, atleast one from Microsoft/Amazon/Google and atleast one representing activities of Enterprise projects (Accenture/IBM/TCS). Then look at govts budgets/programmes for AI US/EU/UAE/S.Korea.

Do it with a group like a book club.

Assuming you have done that, what's your conclusion?

Isn't that worse because you then get into neocloud companies like Coreweave and Nebius who have taken insane debt to build out data centers and if the demand sinks for those they'll go under

[delayed]

It gets worse. CoreWeave has more than $50 billion in debt. rising costs will make financing economically unfeasible. And a demand shock will push the company into bankruptcy. This will affect NVIDIA and others involved in this entire AI supply chain.

CoreWeave defaulting vs say Alphabet or Meta or MS or Amazon is very different thing. According to Pachai right now they are severely compute constrained they have 2x the demand they can serve. So while CoreWeave bondholders might take a haircut one of the hyperscalars will likely gladly buy their assets.

As a kid maybe I don’t get it, a giant ecosystem? Of frontier model providers?

A "gigantic ecosystem" of two frontier model providers

Looks like "Non-GAAP Adjusted Operating Income" is still +4bn$ ... if i read this correctly.

AFAIK GAAP Operating Income has stock compensation packages in it and is susceptible to valuation changes(?)

https://www.sec.gov/Archives/edgar/data/2130357/000162828026...

Liabilities in stocks are still liabilities. It's money they will not get if they go public. You can't just ignore these.

GAAP rules exist for a reason. One of the main reasons being the investors should have a clear and consistent view of the business before they decide to invest. The "non-GAAP adjusted" might as well read "But if we exclude the numbers we don't like our future looks so much brighter".

Big money (Wall Street) will be looking at GAAP equivalent accounts for their real valuations. They'll happily go along with public figures looking healthier than reality, providing they can exit investments or lock themselves in to be paid back first, when offloading to the public or in the case of bankruptcy.

> naively assuming the demand will also grow as technology matures.

I don’t think this is naive at all; it’s quite rational that demand increases as price decreases, which does seem inevitable.

> I don’t think this is naive at all; it’s quite rational that demand increases as price decreases, which does seem inevitable.

But that's not what OC is saying, isn't it? Frontier models becoming better and cheaper at the same time are not a given.

Near-frontier models that have 99% of the capabilities of frontier models, at a tenth of the cost of frontier models, might ironically be what pops the bubble.

Is there seriously a non-frontier model out there that can do what Opus 5.5 does? I can say: go look at a ticket in Jira and figure out how to fix it, right into end tests for it, make sure it aligns with previous architectural decisions we've recorded in documents, and then test it, summarize all the testing you did, and open up a PR for review?

Yeah Opus 5.5 doesn't do that either a lot of the time you have to review it yourself or you're a moron.

Anthropic, for one, doesn't seem to be replacing their human engineers and data scientists with Opus 5.5 yet even when they would be better positioned to do so than any other company. https://www.anthropic.com/careers/jobs

But they're all using Claude (with unlimited tokens) every day, multiplying their effectiveness for sure over what even the best coders can produce.

This is satire, right?

A 10x will always be beat by 10x+ai. But ai+10x is better than just ai. The world has an over supply of ai (or soon will) but a finite number of 10x coders.

CPUs (and a million other things) got better and cheaper with time, I see no reason that AI would be any different.

The real question is not if the technology is going to get better and cheaper, but rather if it's going to meet the market expectations to justify trillions of USD in valuation.

1. What you want changes as the market does, in the way people use agentic coding now rather than just cheaper versions of Cursor autocomplete 2. Frontier labs are cheaper than open-source models too

Most of the "laws" of economy in this form say often as a footnote "all else being equal", which remains to be seen. Otherwise you can justify any attempts at price decrease in any industry the same way.

Unless the number of users is already pretty much saturated. Which it probably is if we are honest by this point.

I absolutely disagree number of users is saturated. Fortune 500 companies, for all of their 'AI accelerating' shareholder speeches, are doing very little to actually integrate or spend on AI for the average employee. I've talked to a lot of people who work white collar jobs that could absolutely be accelerated by AI, and they just... aren't using it at all. Some of the companies make Copilot available. Some people just copy paste their tasks into the free version of chatGPT and spend a lot of time correcting the dumb model. A lot of orgs are concerned about data leakage and just have a blanket AI ban.

I recently joined a software dev team in which nobody is using any AI at all. They weren't conceptually against it, they just didn't feel the need to change their process.

There's a lot of room to grow.

That doesn't make them more money. Those people who aren't maximizing their AI subscriptions already have an AI subscription. If anything this would hurt AI companies if people were actually using the compute they have access to with their subscription.

Revenue per user is very much not saturated

If you are subscribed that's it, no more money possible to make.

[delayed]

  The author talks about revenue growth on the basis of Anthropic's S-1 filing, but ignores the third quarter results which were posted later.

But the AI industry doesn't consist of just Anthropic.

As far as I know, Q3 for Anthropic wasn't the best because:

* OpenAI killed it with their models in Q3

* US Government ban on Anthropic models, including temp restricting Mythos/Fable

* They decided to dedicate more compute to training in Q3 instead of inference

* Opus 5 was a disappointment and Fable 5 was too expensive to inference

* They saturated all their available compute in Q3 and didn't have anymore to sell

Their growth rate has ramped up in Q4 since Opus 5.5 if OpenRouter data is a good proxy.

The competition is also a problem tho. Both of them are betting the house pretty much on winner takes all. If competition reduces growth, that's a problem.

Sounds to me like the market isn't big enough for both of them. So how much growth is left?

For the AI bubble, competition is good because it would mean better cheaper services and products.

> But the AI industry doesn't consist of just Anthropic.

That is not the claim. Author uses Anthropic's filing as an example indicating growth of the whole industry. While questionable on its own, it is also not given in full, which I pointed out.

I know this is a bit of an ad hominem attack, but it is hard to take the authors arguments seriously as they also argue that the imperial system is superior to the metric system because people in Italy apparently state their height as 183cm (in real life everyone would use 1.83m) and that calculating things using the imperial system is easier as well...

Demand will also grow as technology matures, if that maturing means more efficient and cheaper models. We're already at a stage where only a very small niche cares for better more complex models.

But which companies will capture that value? It could be like semiconductors where TSMC makes most of them but captures very little of the value.

People keep repeating this but is it true? In natural sciences, for example, there are endless applications. Anthropic models have only recently reached the level what I would call useful. Still, they are far from perfect and there's a lot of room for growth. Is science a very small niche? I don't think so.