There has long been mystery surrounding the financial condition of major AI labs, and as Anthropic draws closer to an IPO, some answers are beginning to emerge. In particular, the company has finally disclosed more data on the economics of compute power and how much of its soaring revenue is being consumed by compute costs.
Data disclosed by Anthropic in a new filing on Monday indicates that its compute-related profit margins are actually worsening. The filing is an update to the S-1 released last week—but last week’s financial disclosures covered only through June 30 of this year, while the new data extends the period to September 30. Overall, Anthropic performed well this quarter: revenue reached $17 billion, up from $11.5 billion in the previous quarter, showing continued growth. However, operating losses also expanded, primarily due to rising compute costs.
From a hash rate profitability perspective, the situation is worse. In the new quarter, hash rate accounted for 58% of the company’s revenue, up from 48% in the previous quarter. At a time when the company could have attracted IPO investors by demonstrating improved profit margins, the situation is deteriorating.
A chart in the document illustrates this more clearly: it tracks the revenue generated for every dollar spent on computing power. Since early 2025, this metric has steadily increased, briefly surpassing 2x during the most recent quarter before settling at 1.72x.
Several indicators suggest that changes in compute profitability are more indicative of short-term fluctuations than long-term issues, yet they still reveal important insights into operating a large-scale lab business. To meet demand, Anthropic’s compute bills surged to $10 billion this quarter, including “$5 billion in short-term committed compute, which currently costs significantly more than comparable compute acquired through multi-year commitments.” In simple terms, demand grew too quickly, forcing Anthropic to temporarily source compute at higher prices.
Some of these transactions have already become public. Anthropic has been paying SpaceX $1.25 billion per month for computing power, while also entering multi-billion-dollar agreements with Nscale, Akamai, and Volta. In industry terminology, these companies are classified as neoclouds rather than hyperscalers—smaller firms that offer more flexible computing agreements but at higher costs.
Anthropic's third-quarter profit margin deteriorated due to exactly these types of transactions, and there is currently no indication that the company will be able to escape them in the near term. New documents state: “We expect compute and infrastructure costs to continue rising in absolute dollar terms as we secure additional committed compute through multi-year arrangements with CSPs and continue investing in more compute, and these costs will remain our largest operating expense for the foreseeable future.”
Anthropic is still working to reduce compute costs. In particular, the company has devoted significant attention to its efforts to build its own compute infrastructure under a project called "Anthropic Direct." However, given the rapid growth in demand, these efforts are struggling to keep pace, and neocloud remains their best option for now.
In a sense, these are all “champagne problems.” As long as revenue continues to grow at nearly 50% per quarter, few investors worry about high compute costs; but should demand slow down, investors might instead find comfort in the clear improvement in compute availability. Yet this also underscores how heavily AI labs still rely on trillion-dollar infrastructure buildouts—and how vulnerable they are to any disruptions in this process. It’s easy to imagine that over the next year, data center construction could slow if bond markets grow cautious or local resistance intensifies. Anthropic has little control over either of these risks, and either scenario would make an already complex compute landscape even more challenging.
Meanwhile, Anthropic’s and other labs’ futures will still be constrained by the cost of computing power.