Anthropic’s revenue hit $17 billion last quarter, but compute swallowed 58% of it. TechCrunch AI reports that the numbers come from an updated IPO filing released Monday, which extends Anthropic’s financials through September 30. Revenue is climbing fast, and the cost of serving it is climbing faster.
The quick version
- Revenue: $17 billion in Q3, up from $11.5 billion in Q2, or roughly 48% growth in a single quarter.
- Compute share of revenue: 58%, up from 48% the quarter before.
- Compute bill: $10 billion for the quarter.
- Revenue per compute dollar: down to 1.72x, after peaking at just over 2x in Q2.
- Operating loss: it grew, thanks in large part to compute costs.
Why the margins slipped
The filing points to a very specific culprit. Half of the quarter’s compute bill was “$5.0 billion of compute under short-term commitments, which currently costs significantly more than comparable compute procured under our multi-year commitments.”
Think of it like booking a flight the day before you leave. You’ll get a seat, but you’ll pay for the privilege. Demand grew so quickly that Anthropic had to buy capacity wherever it could find it, and that meant paying higher rates.
Some of those deals are already public. Anthropic has been paying SpaceX $1.25 billion a month for compute, which works out to $3.75 billion over a full quarter. It also has multibillion-dollar deals with Nscale, Akamai, and Volta. These providers are often called neoclouds: smaller outfits that offer more flexible deals than hyperscalers like the big cloud giants, usually at higher prices.
Why this filing matters
This is one of our clearest looks yet at the economics of a major AI lab. Those finances have mostly been a mystery. The original S-1, the registration document a company files before going public, landed last week and only ran through June 30. That version showed compute efficiency improving. The extra quarter changes the story.
The timing isn’t great either. Anthropic could’ve won over IPO investors with improving margins, and it’s showing them the opposite.
Still, I wouldn’t read this as a sign the business is breaking. TechCrunch AI calls these “champagne problems,” and that’s fair. Few investors will sweat compute costs while revenue grows nearly 50% a quarter. If demand ever cools, the scramble for expensive capacity should ease too.
The margin drop looks more like a blip than a long-term flaw. But that doesn’t mean the pressure goes away soon.
What comes next
Anthropic isn’t promising relief on the total bill. The filing says, “We expect compute and infrastructure expenses to increase in absolute dollars and remain our largest operating expense for the foreseeable future.”
The company is trying to lower its costs, including through a project called Anthropic Direct, its effort to build its own compute base. According to TechCrunch AI, those efforts can’t keep pace with demand right now, so neoclouds will likely stay in the mix.
Here’s what to watch:
- The revenue-per-compute-dollar ratio: a move back above 2x would support the blip theory. Another drop would make it harder to defend.
- The short-term share of compute: more capacity under multi-year deals should mean better margins.
- Bond markets: if they get skittish, data center financing could slow.
- Local opposition: stronger pushback against data centers could delay new capacity.
Anthropic can’t do much about those last two, and that’s the most important part of this story. Even a lab growing this fast still depends on a trillion-dollar infrastructure buildout that it doesn’t control.
If you build on Claude, I’d keep an eye on pricing and rate limits. The filing doesn’t announce any changes, so this is my read, not Anthropic’s plan. But when compute eats 58 cents of every revenue dollar, capacity costs are a business problem for the whole AI stack, including its customers.
For now, Anthropic’s future and that of its rivals will be at the mercy of compute prices. Next quarter’s numbers will show whether Q3 was a one-off or the new normal. The full breakdown is available at TechCrunch AI.