Anthropic’s business plan: Build God or die trying

The Anthropic S-1 seems to have provided the financial justification for the business to expend as much as it perchance can on compute. If you are the benevolent of individual who wonders if Anthropic is doing its flat finest to build a device god during convincing Wall Street it’s in the endeavor of marketing IT contracts, Anthropic has fair offered a helpful solution: Doing a bunch of things that appearance akin you’re trying to build the device god is the lone way to evade going bankrupt.

Forget concerning the device god for a minute. At smallest in my peruse of the S-1, it seems akin Anthropic is pinning the forthcoming of its business to theatrical growth; it has to, in command to encounter its fixed costs. You don’t have to be an AI skeptic to lift an eyebrow at these numbers. But if you are skeptical that theatrical recursive self-improvement is imaginable — in the way the endeavor scheme seems to necessitate — this doesn’t appearance akin a fine investment.

Let’s commencement alongside the fine news for Anthropic fans. The company’s income grew fourteen times to $16.3 milliard between the archetypal fractional of 2026 and the archetypal fractional of 2025. There’s an $80 milliard loss, but it’s mostly an accounting item that isn’t applicable to the underlying business. If we eliminate that, and several another charges, including employee compensation, we’re looking at $1.62 milliard in adjusted earnings in the archetypal fractional of 2026. I’ve asked before if AI can make adequate prosperity to pay for itself, and the answer from Anthropic’s S-1 appears to be, at smallest for this company, at smallest correct now, yes.

Anthropic was considerably additional disciplined than OpenAI or Elon Musk, and it seems to be paying off. (Anthropic names lone OpenAI and Google as rivals. Better fortune next time, Elon!) The business has never released an AI depiction generator, an AI TikTok clone, or an AI waifu. Instead, it specialized in tasks akin coding and cybersecurity, staying laser-focused on the item it is easiest for a device to do: additional computer. And it focused on the endeavor market, which is small fickle than consumers. Its net income preservation is 900 percent. Pretty good! That income isn’t consumers, either. About three-quarters of customers in the archetypal fractional of 2026 are spending at smallest $100,000 a year.

On the another hand, Anthropic is under serious rivalrous force — from the two another frontier labs and open-source models. The another frontier labs are all rushing for the cutting edge, too. Claude Opus 5.5 currently sits atop the Epoch Capabilities Index, but that leaderboard changes frequently. And for anyone who doesn’t attention that their application is four months behind, there are the open-source ones.

What’s more, Anthropic has a disadvantage against a competitor akin Google. It doesn’t have multiple paths to income — and Google does. Google can additionally cover its AI into another products businesses already use. Remember whenever Microsoft Teams ate Slack? It wasn’t since Teams was the improved product; Microsoft fair had improved distribution, since Teams was bundled alongside another programs companies were already paying for. Having the finest example matters additional for cybersecurity — one area Anthropic is targeting for growth — than for instant messaging software, but it’s difficult to vie on distribution.

And also, income comes in spurts, normally following a new example is released. Between the competition and growth in use following all novel model, Anthropic can never halt training, and it has to do so additional efficiently than its much larger competition.

So one crucial inquiry is: Can Anthropic continue to provision the results we see in the S-1?

There are several long-term commitments that emphasize how serious the enlargement force is on Anthropic. Sure, there’s a convenient blank on its acquisition commitments and lease agreements on leaf 64, and another kicky small blank on the amounts it’s expecting to expend before December 31st, 2030, on leaf 137. However, we cognize the commitments run through 2045.

Fortunately, there’s a “Commitments and contingencies” division that can assistance us out a little. There’s a fixed minimum of non-cancelable acquisition commitments that total $265.6 billion, alongside conditions between one and 20 years. The business has lease agreements “primarily connected to compute and infrastructure capacity” that haven’t yet commenced of $321.4 billion, as of the unaudited June 30th figures. After that, Anthropic entered new compute and infrastructure agreements totaling concerning $245 billion, again alongside conditions between one and 20 years. There’s several another transactions that haven’t yet been finalized, which Anthropic estimates as costing $125 milliard to $150 billion. There are additionally another compute commitments that are variable.

If I add all that up, I’m looking at a minimum of about $850 milliard of commitments complete the next 20 years, which volition increase to concerning a trillion dollars if the transactions that haven’t been finalized are added in at the current estimated cost. And that’s assuming I’m not missing anything, which I may extremely fine be.

So to encounter those commitments, Anthropic’s growth must continue at breakneck speed. Anthropic doesn’t provision us a fee agenda for its forthcoming commitments, but analyst Harrison Rolfes of PitchBook models it at concerning $90 milliard a twelvemonth from 2027 to 2030, which method $188 milliard in income to keep the current proportion of disbursal to compute stable. “Growth is not optional for this company,” he writes. “The contracts assume it.”

Compute and infrastructure (and a few another items) connected alongside customers disbursal $2.3 milliard in the archetypal fractional of this year. Model betterment (and a few another items) was considerably additional expensive; Anthropic spent $9.7 milliard on it. But Anthropic simultaneously got additional productive — spending small prosperity on compute to bring in all dollar of revenue, according to Rolfes.

Is that a long-standing trend? Hard to say. We don’t have the third fourth numbers in forefront of us. “The S-1 does freeze in period a blessed duration whenever their first models skyrocketed in use and they didn’t have adequate compute so they are capable to accusation additional for what they had,” says Brendan Burke, an expert at Futurum Group.

As endeavor software, it makes awareness for Anthropic to mark companies’ IT budgets. Conveniently, cybersecurity is a location anywhere you volition entirely get your crap rekt if you are not continually up to date — everyone there has to keep operating fair to remain in the identical place, exactly akin Anthropic. Of course, AI is part of the problem; hackers can use it to brute-force material they couldn’t do before. Plus, the products themselves are additionally causing headaches gratitude to a partner’s careless fuck-up of security protocols about AI agents.

Frontier labs need high-margin products to evade a competition to the base on token pricing, notes Burke. That made item alternatively remain out to me: Opus 5.5 reduced client expenses by 40 percent complete Opus 5. Now, Anthropic says that lesser disbursal method additional usage, which lets the business maintain its margins. If you assume that disbursal decrease happens all period a new example is released, beautiful shortly you breeze up alongside tasks at a fraction of the price.

This isn’t necessarily an insurmountable problem, in part since risen effectiveness doesn’t continually slash margins, says Gil Luria, an expert alongside DA Davidson. Amazon Web Services and another haze providers ”have rotate into far additional productive all twelvemonth and decrease their prices all twelvemonth and they have had dependable growth and rising margins,” he says.

But the force for growth is beautiful undeniable. That’s anywhere recursive self-improvement, alongside minimal or no supervision, comes in. “More capable models assistance our researchers detect these improvements faster; in turn, additional mighty compute multipliers allow us to train equal additional capable models and deploy them additional efficiently.” Based on the broad rhetoric Anthropic has concerning the apocalypse or whatever, it expects to hit theatrical recursive self-improvement, perchance soon.

Claude is already construction Claude under individual direction. In its hazard factors, Anthropic warns that it’s locked in to that approach. That comes alongside a apparent tradeoff, since using AI to create additional AI “may logic the capabilities of our models to advancement additional quickly than our capability to develop, validate, and execute corresponding safety, alignment, and safety measures.” If I am study that right, Anthropic is saying that its endeavor relies on construction new models so fast, using activity it may not completely understand, that it is basically forced, for endeavor reasons, to do material that may be risky to everyone. Which is why we’ve got a bunch of weird material concerning the existential hazards of “catastrophic” harm and the betterment of “forms of sentience” in the hazard factors.

Is theatrical recursive self-improvement that escapes the limits of individual supervision possible? Dunno. But if it isn’t, that makes it much harder for Anthropic to encounter its existing compute commitments and rotate a profit.