Anthropic's Compute Costs Consumed 58% of Revenue in Q3

Anthropic spent 58% of its revenue on compute in the quarter ending September 30th, up from 48% the quarter before, according to a new filing released Monday. The number lands at the worst possible moment for the company: it is preparing for an IPO, and investors pricing the offering will read the filing closely.

The filing is an update to the S-1 that Anthropic released last week. The earlier document disclosed financial results only through June 30th. The new version extends the disclosure by one full quarter, to September 30th, and the added period shows a company growing quickly while losing more money on the infrastructure that powers its models.

What do the new numbers show?

The headline figures point in two directions at once. Revenue reached $17 billion in the quarter, up from $11.5 billion in the previous quarter — growth of roughly 48% in three months, which suggests durable demand for Anthropic's products. But the operating loss grew over the same period, driven in large part by rising compute costs.

The margin picture is the more troubling half of the disclosure. Compute's share of revenue climbed from 48% to 58% in a single quarter. At a moment when Anthropic could win over IPO investors by showing margins improving, they moved in the opposite direction.

A chart in the filing tracks the relationship from a different angle: revenue generated for every dollar spent on compute. That ratio had been growing steadily since the start of 2025 and peaked at just over 2x this past quarter. It then fell back to 1.72x. Each dollar of compute still produces more than a dollar seventy in revenue, but the efficiency trend reversed at the exact moment the company opened its books to public-market scrutiny.

Why did the compute bill jump?

Anthropic's compute bill for the quarter reached $10 billion. The filing breaks out a specific line within that total: "$5.0 billion of compute under short-term commitments, which currently costs significantly more than comparable compute procured under our multi-year commitments."

In plain terms, demand grew faster than Anthropic's contracted capacity, and the company had to scramble for compute on the open market at premium rates. Half of the quarter's compute spend came through these expensive short-term arrangements rather than the cheaper multi-year deals Anthropic prefers.

Some of those supply deals have already surfaced publicly. Anthropic has been paying SpaceX $1.25 billion a month for compute. It has also signed multi-billion dollar agreements with Nscale, Akamai and Volta. In industry parlance, these providers are neoclouds rather than hyperscalers — smaller outfits that offer more flexible compute deals, but at higher rates than the large cloud providers command under long-term contracts.

Is the margin decline permanent?

The evidence in the filing suggests the shift is more of a blip than a structural problem, but it still reveals something important about how a major AI lab operates in 2025. When demand surges faster than planned, the only way to serve it is to buy compute at spot-like prices, and those prices erode margins immediately.

Anthropic does not expect the pressure to ease soon. The filing states: "We expect compute and infrastructure expenses to increase in absolute dollars and remain our largest operating expense for the foreseeable future as additional committed compute becomes available to us under multi-year arrangements with CSPs, and as we continue to invest in additional compute."

That sentence carries two messages. First, Anthropic is locking in more multi-year capacity with cloud service providers, which should bring per-unit costs down over time. Second, the company will keep pouring money into additional compute regardless, so the absolute expense line will keep climbing even as the terms improve.

What is Anthropic doing about it?

The company discusses at length its effort to build its own compute base under a project it calls "Anthropic Direct." Owning infrastructure rather than renting it is the long-term answer to the margin problem. But demand is growing so fast that these efforts cannot keep pace, and the filing gives no indication they will replace external suppliers soon. For now, neoclouds remain the best option Anthropic has for bridging the gap between contracted capacity and actual usage.

Why does this matter beyond one company?

In one sense, these are champagne problems. Revenue growing at nearly 50% a quarter will quiet most investor concerns about high compute costs. If demand ever slows, the compute picture will likely look a lot sunnier by then, because the short-term commitments will have rolled off and cheaper multi-year capacity will have come online. Either scenario is survivable; the uncomfortable middle ground — heavy spending on expensive compute while growth cools — is the one the filing does not resolve.

The disclosure also quantifies how much the major AI labs depend on the trillion-dollar data center buildout, and how vulnerable they are to hiccups in that process. Anthropic's margins deteriorated in Q3 because of the deals it struck with neoclouds, and there is no sign it can do without them anytime soon.

The risks sit largely outside the company's control. The filing arrives as the broader infrastructure buildout faces two identifiable threats: bond markets could turn skittish about financing data center construction, and local opposition to new facilities could intensify. Either development over the next year would slow the buildout. There is not much Anthropic can do to address either risk, and either one would make a messy compute picture considerably messier — pushing more of the company's spend toward the short-term, high-price arrangements that hurt its margins this quarter.

The S-1 update is the clearest public accounting yet of the economics inside a major AI lab, a class of company whose finances have remained largely opaque. It shows a business that converts compute into revenue at a healthy multiple, grows revenue at extraordinary speed, and still cannot fully control its largest cost line.

Until Anthropic's own infrastructure under "Anthropic Direct" reaches scale, and until multi-year cloud commitments cover a larger share of its needs, the future of Anthropic and other labs will be at the mercy of compute prices.