November 2, 2026 in finance
The Anthropic S-1

So, now that the S-1 is out it's time to revisit the thoughts I had on AI unit economics, what checks out and what was wrong?

All numbers in USD:

So, H1 revenue was 16.28B and 13B of that was from pay per token customers, only 20% came from customers like me. I only have a sample of one company that I know of personally that uses Claude Code extensively and they pay per seat with no overage arrangement, ie if a dev hits the cap then he's done unless he gets upgraded to eg 100 that month. So that company is effectively in the 20% bucket. This is a long winded way of saying that even enterprise customers can have fixed subscription deals without being forced to pay per token if limits are reached, we'll circle back to this fact later.

The S-1 claims that cost of revenue for serving both buckets is 2.38B, that is less than the revenue from my bucket which was 3.2B. So, on a pure cost of goods basis Anthropic is profitable even if they would have only customers like me, ie the meme in the above linked article does not check out after all, on a pure COGS basis at least.

There are two places I went wrong in my COGS assessment:

1.) My utilization, even though Claude continues chugging along when I'm AFK the roughly 5 agents I typically juggle will eventually stop for input from me, and if that happens when I'm sleeping it could be up to eight hours of downtime per agent. Furthermore, I was assuming that my 80% credit usage would equate to 80% of my $100 / month plan, ie costing Anthropic $80, or even more. That is probably not the case, it is ofc what Anthropic is comfortable with, which could for instance be 80% of $50 bucks for a guaranteed 50% COGS margin. That was a brainfart on my part.

2.) Internal optimizations, the example I was using from InferenceX used a comparable Chinese model via a direct prompt, no harness. Agentic coding, where Anthropic has full control over both the harness and the server side logic that the harness is sending calls to is ofc opening up for all kinds of optimizations to get the cost per generated token down.

I think both of the points are in play here in order to achieve that fantastic COGS margin. I'm impressed, the foundations for a wildly profitable business are definitely there.

This is also why a 139k depreciation cost per month per rack is doable, it is simply not punishing enough to make a dent in this business at the current token prices that Anthropic is able to charge.

Valuation

Any kind of future projection or conjecture about whether or not the IPO price of +2T is reasonable must take these things into account:

1.) Competition from Gemini, ChatGPT, Muse, MAI, Composer / Grok and Chinese models.
2.) The cost of staying ahead of the pack in the software engineering vertical.
3.) Pay per token vs subscription fee.

Let's take them one by one:

Competition

Google: Gemini comes off as a joke to me atm, it hallucinates, fabulates, acts sycophantically and suddenly starts writing answers in Norwegian even though I strictly communicate with it in English. It's a joke of a product that is clearly just designed to be a wrapper around Google search to preempt OpenAI from completely stealing that whole ad business from them, and for that purpose I suppose it's doing a fairly good job, but this doesn't have much to do with Anthropic at this point in time so digression over. Caveat: I switched over to Claude from Gemini, for coding, before the release of Argon so I have not tried out Argon when writing this. Maybe a Claude Code killer, maybe not...

OpenAI: I haven't tried GPT-6 Sol yet, and why would I? I'm happy with my Claude Max subscription for the same price as the Sol subscription, but I note that Sol is $2 / $10 per 1M on OpenRouter atm, and Opus 5.5 is $4 / $20 per 1M, that is 50% and should be highly interesting for anybody who is paying per token for Opus 5.5.

Meta: Muse Wonder Full is supposed to be good for coding but I'm not switching, I'm happy with Opus 5.5.

Microsoft: AFAIK there's nothing here atm that directly competes but the problem is that MSFT has got control over distribution via GitHub, it's in their interest to try and switch out Claude with the future MAI-Base-2.

SpaceX: I haven't tried Grok for coding and Composer is untestable in isolation, but we have the same issue here as we do with GitHub, control over distribution via Cursor and trust me, Elon has zero plans on keeping Claude as the intelligence behind Cursor for a second longer than necessary.

Chinese models: I've tried DeepSeek 4 Pro via OpenRouter, it does good work but stumbles when it comes to harder work. That could be more a function of the harness I used (Goose) than the model though. But the token cost here is the thing, it is currently at $0.8719/M input tokens and $1.744/M output tokens versus $4 / $20 per 1M for Opus 5.5, again very interesting numbers for anyone paying Anthropic per token.

So, we have cheaper American models, dramatically cheaper Chinese models and competitors aiming for larger shares of the coding vertical that also have control over some of the distribution. And it's not just Cursor and GitHub, any gateway company would want to get their own costs down whilst at the same time keeping the subscription fees where they are, and the way to do that is by for instance trying to switch out Claude for DeepSeek, at least for stuff like boilerplate code generation etc. I am pessimistic when it comes to competition in the future.

R&D Costs

Of the 16B income almost 10B was eaten up by internal R&D, read: training Fable / Mythos. If this cost can stay constant while the growth continues at the current clip then Anthropic will end up becoming a massive cash gusher due to the high COGS margins we've already discussed. If on the other hand the cost of staying at the frontier continues growing then the math gets ugly, because then you're looking at a situation where you have to keep running faster just to stay in the same place. In this scenario Anthropic is forced to constantly grow revenues in order to keep up with the guaranteed growth of the training costs which will create constant anxiety on Wall Street.

My take is that this cost will start to plateau and I base that take on the limitations of the technology, both hardware and software, simply making these things larger (which is the thing that would increase costs) will come with diminishing returns. World models could change that but Anthropic's main business is software development and that is text only and staying on top of that particular niche will probably involve cheaper optimizations than training ever larger frontier models. I'm optimistic with regards to future costs.

Revenue Mix

either the OpenAI / Anthropic pay per token setup is fantastically, insanely profitable, or the subscription business is losing a nauseating and horrible amount of money.

That is from my prior post on this subject. And I think we can now conclude that: Anthropic has a fantastic business in selling pay as you go, per token, and an OK business in selling subscriptions.

We have already concluded that most of Anthropic's revenue is from customers paying per token, and if we now circle back to my example medium sized enterprise that I mention at the beginning of this post, that was only paying for subscriptions, then I wonder how long the pay per token money fountain can continue. More and more enterprises will similarly angle to move from pay as you go to subscriptions.

Presumably a lot of the pay per token revenue is coming in from entities like GitHub and Cursor and we've already concluded that entities like that are incentivized to get rid of expensive Claude usage. Other uses are ofc as integrations of AI straight into the applications, that would be pay per token customers by the nature of the usage, think MSFT Copilot et al, but these customers are also incentivized to get the token cost down. I am pessimistic when it comes to the future revenue mix.

Scenarios

Let's assume that the whole adventure will play out between now and 2036, so ten years into the future.

Steady state, zero growth end state with a 2T market cap and P/E of 10: here we're then looking at profits of 200B / year, let's slap a 50% margin on that so then we're looking at 400B of revenue, this definitely assumes that R&D stays fairly constant, which I believe is probable in the software development vertical. In order to reach this state they have to more than 10x the revenue from here which frankly does not seem impossible.

To maintain this position customers like me need to stay on the $100 plans and not switch supplier which seems achievable, we have already concluded that I can't be bothered to try Muse, Argon nor Grok as the pricing is roughly the same and I'm happy with Opus 5.5. What would move the needle for me is an equivalent plan for $50 / month, that would entice me to try out alternatives but then they would need to also be just as good or better. There is zero chance that I would stomach even 1% worse quality output for $50 / month.

Given the competition etc I consider this scenario to be middling and probable, then the market cap ends up in the same place it started, 10 lost years. Let's take that scenario and slap an 80% margin on it, this kind of margin could be achieved by COGS trending towards virtually zero and customers like me continuing to pay the same subscription fees, not impossible given the specs of Rubin and if Feynman also achieves a 10x reduction in generation cost then why not?

So, 80% results in profits of 320B and if we apply a P/E of 10 to that we get an mcap of 3.2T, 50% up from an assumed 2T IPO price, 5% / year, not great.

Max TAM achieved with a P/E of 10: JetBrains estimated that there were 13.4M paid developers in the world in 2024. SlashData claimed 36.5M in 2025. Let's just do (36.5+13.4) / 2 = 25M. If we slap a yearly $40k average salary on that number then we get 1T in total cost that Claude could capture, with 80% margins on that we arrive at 800B in profits for a 8T valuation, a quadrupling from the IPO.

I'm not going to belabor the absurdity in the assumptions that make me arrive at an 8T terminal valuation. The crazy part is that even if this scenario plays out it only represents a yield of 40% / year.

Conclusion: anyone buying at the IPO price is providing exit liquidity to the pre-IPO investors. The big gains are not to be made by the investors in the public company, they were already made by the investors in the private company.

avatar henrik