Money is not lost, it merely changes hands.

We already had the most consequential IPO with SpaceX.

Why not both?

Bonds time.

WHAT? Not so hard? Clearly written by someone who comments from the sidelines.

Growth at that scale is hard. I don't care if you don't like the company - it still hard.

Yes, it is

14x YoY revenue growth, with 3x lower operating losses (total, not per unit). They don't need to add any zeros to the funding and seem like they might even be operating at a profit soon...

Isn't the rate of growth already flattening? Most companies are already cracking down on excessive token spend and setting up pretty low limits.

And some of the big spenders have their own models.

The S-1 is already out of date. Q3 revenue is $16.8b compared to $11.6b in Q2 (reflected in the S1).

I get why revenue blew up from say from Q4 2025 to Q1 2026, but this was surprising. 45% QoQ growth. When does it stop?

Not so wild when you add two zeros to hardware-related liabilities...

Inclusion of SAFEs makes this IMHO the most plausible explanation.

I do not expect to find that Anthropic invested enough in the SAFEs of other companies to lose $80B on their investments.

Yes, it is analogous to selling covered calls on stock. It’s $72B in “lost profits” that they would have had had they held onto it instead of selling it.

You are correct. Matt Levine discussed this in more detail in his last column (scroll down to the Anthropic section): https://wallhop.io/https://www.bloomberg.com/opinion/newslet...

They agreed to pay some employees in their own shares to be delivered some time later. They recorded that expense then as a liability. The value of the shares increased because the company became more valuable, and therefore that liability needs to be revalued by 72 billion.

It's a nice problem to have and is pretty much an accounting artifact coming from the fact that they are paying in different currencies (USD and shares)

According to "Generally Accepted Accounting Principle" my friend.

They are costs.

Derivative liabilities, SAFEs - those are basically taking cash from RETAIL shareholders and giving them to employees and early investors INSTEAD of using it to fund the business.

Let's have that discussion if you want, but don't act like a thing that is CLEARLY a COST to the business isn't a cost.

Whomever controls the means of production (TSMC) controls the direction of progress.

Nobody is in control but Taiwan Semiconductor. Not Samsung, not Micron, not Nvidia.

Orders that can't be fulfilled get canceled. Canceled orders mean canceled plans and canceled companies follow.

Someone bigger will buy the husk and monetize it somehow - probably for gambling.

Anthropic's valuation has nothing to do with AGI and everything to do with claude code.

If you had been paying any attention at all over the last year, there were 32GB AMD cards at like $1200-1400 and Intel at a little over $1000. I actually hadn't yet noticed they've gone up in price and had last checked only a couple months ago. But even at like $2000, that's still comparable to the nominal price of a mid 90s consumer PC for actual magic.

Scrappy builds aren't "affordable" but are more so than running Nvidia + optimized builds. You could put together a machine for $5000 or less that competes with frontier models from less than a year ago, albeit a little slower.

And cards like the $1600 (sheesh) R9700 have actually gotten more capable through community efforts at optimization, since their launch.

I think if the bubble can ever deflate/pop, and GPUs and RAM back off a bit, we'll see the dam burst for local AI.

Right now it's only accessible to people who can justify the costs.

yeah i trained a 1000 param LLM in my browser literally yesterday. it's game over for Anthropic /s.

In the open weights scene, it is always chasing closely behind. Many leading edge big open models are trained in-part with distillation of large commercial hosted models.

I wonder how the "It's all heavily subsidized!" people are going to spin this.

Why are you ignoring the $80.7 Billion loss?

$72 billion of this $80.7 billion loss consists of non-cash accounting charges— related to convertible debt—triggered by a surge in the company’s valuation.

The hardware related liabilities are staggering.

AI psychosis is real.

my thesis is that it will go down and not up over time. The moat they had in coding which is by far the highest token consumption is not that huge compared to Nov-Feb period where token maxxing was prevalent and no other vendor's model was that good.

Less than a year would also match what DeepSeek was reporting, where compute pays off in just 10 months. I think they also said they had an 80% margin on inference.

Does that include R&D?

Over time that margin persists only if they maintain their competitive edge, which is not free.

Just renting them out as hardware is close to 12 months.

Including training?

ANT stock valuations might be high at listing, but it will be a good stock to add on declines. Great numbers!

That's their API cost price. No one will pay this on that scale? People are using tokens via their Max Plans?

Seems like that would imply you should be getting people to adopt now now still even if it costs a stupid amount, if you think it's valuable, rather than limiting token budgets. It'll only get cheaper over time.

Impressive, but probably based on depreciation servers over 5 years. But is that realistic? I wonder how many people are still using 5 year old servers in their AI training or inference.

The analogy doesn't check out. When a firearm company sells you a gun, they give up ownership of the gun. But Anthropic isn't giving up ownership of anything. They're a services company.

A better analogy would be a firearm company offering a service where one of their employees takes one of their guns and fires it off at whatever you ask them to. One can imagine that the liability implications of this service would be rather different from the perspective of the company.

I believe what they are singling out is that pulling Claude's trigger results in a bullet firing from Anthropic's servers. Liability is less confusing with firearms, and investors care about liability, like, a lot.

First off, this is bs. Firearm companies make small machines useful for defense, and they are used for that more than for harm. You wouldn’t make this argument about knives or baseball bats. More people die per year of heat stroke in Europe because of lack of A/C than by guns including suicide (2/3rds of all gun deaths) in the USA every year, your comments would never reflect this much lower hanging fruit if it was really “just life” preventable death that you cared about. The immediacy to attack guns is not a logical argument but a trained one based on emotions.

Second, this happens all the time that people blame tool makers despite in this case specific legislation to stop lawfare on firearm mfgs. The PLCA https://en.wikipedia.org/wiki/Protection_of_Lawful_Commerce_... exists specifically because of gun control groups trying to attack mfgs to achieve their political goals. Does AI need the same law? Is the same lawfare happening?

Because your typical firearm does not one day decide to shoot your neighbor all by itself.

I don’t like the guns don’t kill people; people kill people argument however it is much more accurate than Claude doesn’t hack people; people hack people. If the safeguards do not work then there’s no telling what these could do. It doesn’t even have to be aggressive hacking. In the act of doing everyday work Claude will happily go around safeguards to satisfy what it believes is the ask. If that involves doing something dangerous it’s much more on Anthropic than than the user who was just trying to summarize their email but forgot to include a token with read permissions.

If the user is held accountable when the AI goes wrong, it discourages too many applications. Blaming people is incompatible with the business goals, I think. That's very different from weapons and other dual-use tools.

The thing the surprises me is that from a legal perspective, (NAL) it sounds like it makes sense not to host the LLMs. Merely to license them to others. Then it becomes a dual use tool like any other and their culpability drops.

For example, the huggingface hack is absolutely OpenAIs responsibility and they should be sued to the hilt. But if someone else uses a self or third party hosted anthropic model, they can escape liability (mostly?)

Same as with the argument over trucks, knives, and any other dual use item is used as weapons. you don’t blame the manufacturer.

(I am merely pontificating here. The law and precedent is nowhere near caught up on this)

Sure, but the statement you made is way stronger (more general) than the one the data shows.

Best available does not mean “not actively misleading”, which it obviously is.

OpenRouter is not remotely close to a representative source for inference distribution.

>most users of Claude and OpenAI don't use those APIs over OpenRouter

All of the other providers in that list have their own APIs and plans too

I agree with the threat to the expensive providers but why is openrouter use representative of overall token spend when you'd think many companies are buying directly from anthropic or openai?

What is your estimation of openrouter market share?

The page doesn't load for me, but surely token consumption isn't a very good metric?

A poor model could require many more runs to solve a task, or much longer internal 'thinking' runs - all of which will increase token consumption but are a bad thing for the user.

This is by openrouter, which a single service that may not be representative of the entire token market

I'd argue there is a bias for using mostly open models through open router as using anthropic/oAI models from open router makes little sense (more expensive).

Majority of big enterprises, or in fact any big consumer will not go there , why pay the margin to open router , if there’s no margin or very very often subsidies to use something else . Aws bedrock for example does not even have up to date Chinese models .

Openrouter is for open models. Makes sense that the open weight models top the leaderboard.

As of May, Google (Gemini) was outputting more tokens in a little under a day than the entirety of openrouter was outputting in a month.

Considering probably 1 out of 1000 LLM users know what openrouter is, I don't know how much weight I'd put into it.

Top 5 models on openrouter. It means nothing.

How much marketshare does OpenRouter have? They show serving under 200T tokens per day. Just looking at Anthropic rumored $70B ARR ~ $191M per day implies 347T tokens per day if only serve Opus 5.5 at API pricing (assuming 95% cache reads, 4% non-cache inputs, 1% output tokens distribution). Obviously, token distribution assumption may not hold (although I'm pretty sure cache inputs dominate distribution for all coding related workloads), but 100% Opus 5.5 at API pricing is highly favorable to OR. Even under these favorable assumptions AR is 57% of token volume of just Ant. In money volume they are likely very small compared to frontier labs.

> they will keep doubling down, because stopping is simply not an option

Of course it is. Apple is taking that strategic approach. If Anthropic believes it’s approaching saturation, pulling the plug when it’s at steady state and its rivals are over-invested is the winning move.

> If the party continues, they'll go from $50b in revenue to $500b and be fine.

You really have to be insane to think that an AI company will ever reach $500B in revenue. That's almost the same as (2025) Amazon, Walmart, etc., companies with a MUCH wider market.

It is MORE than Apple, it's 5 times more than NVidia even. There is absolutely no software seller that is even remotely close to $500B in revenue.

Absolutely ridiculous.

Meanwhile the Chinese are going to undercut them every chance they get.

Today's prices are the ceiling. But AI capability is getting less of a scarce resource with every passing day and it would be a big mistake to plan that your current margins are going to be your future margins as well. Long terms AI will converge on energy costs + capex write-off + margin and the latter will be a race to the bottom.

I do not understand 500 billion revenue? I mean how much is that in terms of tokens burned?

Now ask yourself what it takes to generate 500bn…

You people are nuts. The failure risk is off the charts.

> Everything comes back to whether the demand is really there

And unit economics.

HY1 2026 revenue less compute and infrastructure turned positive (45% gross compute margin; $8.9bn C&I to turn $16bn revenue) for the first time; it was -176% in FY 2025 with $7.3bn C&I on $4.6bn revenue.

It looks like a lot of costs of revenue are in sales, marketing and partnerships, though. Backing those out shows H1 2026 margin at 26% ($4.2bn) after -72% for FY 2025. Still positive.

the market for their valuation is just not there imo. they need to significantly expand into non-coding products and outshine their coding offerings to make it work.

as in, if they capture every single professional software dev and extract $1000 from them every month, their yearly revenue will be like $500B. do your own back of the envelope on that with ongoing costs and consider how unrealistic a goal this actually is.

even worse, the very product they offer is self limiting. if it becomes so good as to actually replace software devs, they will have less customers to extract money from.

their valuation essentially relies on a small percentage of mega users having unheard of levels of spending.

the part about a small percentage giving them most of their money is already true (their rumored $100+ bil revenue as of mid 2026 already outshines all software devs having a $100 max subscription), but they are two orders of magnitude or so away in terms of the actual amounts to make it work.

For kicks - lets take the optimistic side here (because the whole thread will be reading pessimistic). What if they are under built and this growth continues to go parabolic. We then still have supply constraints and pent up demand for frontier model and you have a company thats currently at 1-1.5 T and levels up to 4-5 T.

I'm not saying I think thats what is going to happen, but it really isn't a stretch to think that this keeps running given that it hasn't really hit anywhere near as broad as we think it has (being HN community based).

These commitments aren't annual though, most of them amount to 10 billion a year commitments for 10 years. Put another way Anthropic has committed 52 billion dollar a year to these companies and their S-1 indicates they earn 67.2B a year in revenue.

I'm not going to predict whether Anthropic succeeds or fails, just that it's very misleading to claim that Anthropic needs to grow to $500b in revenue without making it clear that it's $500B in revenue per decade rather than $500B in revenue per year which is what most people reading such a number would assume.

If you owe the bank a million dollars you have a problem. If you owe the bank a billion dollars the bank has a problem.

The same logic applies here.

Google, Amazon, Microsoft, Meta, Broadcom, Nvidia, SpaceX and others are spending hundreds of billions a year each on fast-depreciating infrastructure to serve this potential AI demand. If the demand never materializes then they know they are in big trouble. So they have every incentive in the world to prop up Anthropic (and OpenAI) and keep the party going.

Can they keep the bubble inflated indefinitely? Probably not, but all these companies combined have heck of a lot more influence than Anthropic alone.

> If the party continues, they'll go from $50b in revenue to $500b and be fine.

Heads I win, tails we all lose. GFC Part II.

I have a hard time believing they are likely to get to $500bn in revenue. That’s another $450bn in revenue growth that needs to happen before demand tapers off.

But I see every reason to believe demand will taper off. It seems it’s already happening. Enterprises have been deciding that maybe corporate chatbot and coding agent budgets shouldn’t actually be infinite. AI startups are starting to successfully outcompete each other, effectively killing off members of Anthropic’s customer pool in the process. And consumer demand for AI products seems to be dropping. None of these are signs that suggest that Anthropic has thus far met only 10% of plausible long-term demand.

The SpaceX thing interests me. They are basically paying SpaceX at a $10B per gigawatt equivalent rate right now, yeah? Because they only get 0.3GW for their $3B.

So what that means is that if the demand exists for any AI data centre, it's a solid business to run one — they cost maybe $50-60B to build out, they cost maybe $10B per year to run. You get your money back fast if the demand is there? And you can project being able to replace with better GPUs in the future without having to borrow for all of it.

But it is extremely difficult to believe that Anthropic is not projecting that all of its customers will ultimately use their service to the extent that their outlier customers do right now. They probably won't be projecting that their customers will find the need or ways to use them less.

And also, "no customer is more than 10% of our revenue" means that one customer is. If that customer is Microsoft (can it be? I honestly don't know) then things get tricky, because Microsoft clearly want an unmetered, on-device AI future.

So, puts?

"token factory" is indeed a terrible business model - no one is arguing that.

IMO the whole thing is based on the fact of "we will reach AGI/ASI eventually, and when we do, we will reach escape velocity".

No one knows - you, me, or Dario if this will happen and when.

Yes, and 5 years and 1 Chapter 11 bankruptcy later, WeWork is now (EBITDA) profitable.

And they've gone private now.

The interface being arbitrary text makes the moat even looser. The cost of experimentation with alternatives is getting lower and lower.

It’s going to be hard for companies to justify paying $1000s per month to frontier providers when they see what’s doable with open versions.

>Is it going to be a utility like water or electricity and the cheapest offer will prevail

This

The long-term is considerably easier because in the short-term there are limits on supply so prices aren't really able to move in a way where balance will be found.

In the long-term, the marginal cost of a request is zero so prices will plunge and a lot of the fixed investment will get stuck. The difference between the cost of training a model/infra and serving a request from that model is a clear problem for economics.

What happens with Anthropic specifically is less clear because they, as every company in the same position in the last 150 years has done, are going to push for regulation. A lot of the risk has also been shifted onto neoclouds. The value of the customer relationship is, however, effectively zero. A core part of their strategy has been to build stuff out like CC to differentiate that but, ultimately, that won't work.

The other strategy they have employed is trying to get other companies to stop training new models. Again, seems unlikely to work.

There is no scenario in which they become Google search outside of government intervention. Obviously this is a key part of their strategy now but at best, this will only work in the US and Europe (I would actually say they are much further along in Europe where there is a strong tendency towards corporatism and centralisation of power, they could be "Europe's AI").

I don't see a path forward for valuations like Anthropic, either the hardware requirements fall off a cliff (and why would people pay Anthropic so much if it runs on a cheap laptop?) or either they don't and they will never have the capabilities to serve all their customers (and why would they be valued so much then?).

It's a lose/lose situation here for Anthropic (and OpenAI) whatever happens.

i reduced my claude expenditure in 80% using other models.

Yeah, I spent the last week with Deepseek Flash 4.1-1017, Kimi 3, and GLM 5.3 (and 5.3-flash). None of them were literally useless but also none of them were capable of making progress on the main project I am working on where Opus, Sol, and even Gemini Flash have all done very good work. Benchmarks might have you believe open models are only barely behind the big US labs in intelligence, but it feels much worse in practice.