Anthropic updated its S-1 filing on Monday with results through September 30, showing revenue of $17 billion, up from $11.5 billion in the prior quarter. Compute costs absorbed 58% of revenue, up from 48% a quarter earlier, while the operating loss widened. The disclosure matters because the company is approaching an IPO and investors expected evidence of improving unit economics.
Why compute margins fell in the third quarter
Revenue for every dollar spent on compute fell to 1.72x after peaking at just over 2x in the previous quarter, according to a chart in the filing. The total compute bill reached $10 billion for the quarter. That included $5.0 billion of compute under short-term commitments, which the filing says currently costs significantly more than comparable capacity under multi-year commitments. Demand grew so quickly that the company had to secure extra capacity at higher rates.
Anthropic sourced part of that flexible capacity from neoclouds rather than hyperscalers, meaning smaller providers with more flexible terms and higher prices. The source names payments of $1.25 billion per month to SpaceX for compute, alongside multi-billion dollar deals with Nscale, Akamai and Volta. The company is also building its own compute base under a project called Anthropic Direct. Those efforts have not kept pace with demand growth, leaving expensive short-term supply as the available option.
The filing frames rising infrastructure spending as structural rather than temporary. Anthropic states that compute and infrastructure expenses will increase in absolute dollars and remain its largest operating expense for the foreseeable future. Additional committed capacity is expected to become available under multi-year arrangements with CSPs, while investment in extra compute continues. The context is a trillion-dollar data center buildout on which large labs depend, with possible constraints from bond markets or local opposition.
What rising compute costs mean for business
For companies buying AI products, the near-term effect is continued availability paired with unstable supplier costs. Anthropic grew revenue by nearly 50% quarter over quarter, so capacity was found despite the higher bill. Small firms using Claude through APIs or partner platforms will likely see stable access, while large customers with heavy inference loads face greater exposure to price terms, capacity limits and contract renegotiation. Vendor diversification becomes more valuable when one supplier depends on short-term procurement.
The filing does not establish that higher margins are permanent, and the source describes the deterioration as looking more like a blip than a long-term problem. What remains unclear is how quickly multi-year CSP capacity will replace short-term commitments and at what effective price. Buyers should ask vendors what share of capacity is short-term, how inference prices map to infrastructure costs, and what guarantees cover throughput during demand spikes. This report alone does not prove that model economics have broken.
The marker to watch is the next disclosure of revenue per dollar of compute and the share of short-term commitments inside the quarterly compute bill. A return toward 2x and a lower short-term share would show that multi-year capacity caught up with demand. If the ratio stays near 1.72x while revenue growth slows, pressure on pricing and investment discipline will increase.
