There remains considerable uncertainty surrounding the financial operations of leading AI laboratories, and as Anthropic approaches its initial public offering (IPO), clarity is beginning to emerge. Notably, the organization is now revealing more details regarding the economics of computing and the extent to which its rapidly increasing revenue is being offset by computing expenses. In a recent filing on Monday, Anthropic disclosed information indicating that the margins associated with computing are, in fact, deteriorating. This filing serves as an update to the S-1 submitted the previous week; however, while the earlier document’s financial disclosures were limited to data through June 30 of this year, the new information extends to results up to September 30.
Overall, Anthropic experienced a positive quarter, with revenue reaching $17 billion, an increase from $11.5 billion in the preceding quarter, suggesting sustainable growth. However, the operating loss simultaneously expanded, primarily due to escalating computing costs. A closer examination of the computing margins reveals an even more concerning scenario. In the latest quarter, computing expenses accounted for 58% of the company’s revenue, rising from 48% in the prior quarter. At a time when the company could attract IPO investors by demonstrating improved margins, the reality is that they are worsening. A chart included in the filing illustrates this trend more vividly, tracking the company’s revenue for every dollar spent on computing. Since the beginning of 2025, this ratio has been on a steady ascent, peaking at just over 2x in the most recent quarter before declining to 1.72x.
Numerous indicators suggest that this shift in computing margins may be a temporary setback rather than a long-term issue; nonetheless, it provides significant insights into the operations of a major laboratory. In an effort to meet demand, Anthropic’s computing expenses surged to $10 billion for the quarter. The filing notes that this includes “$5.0 billion of computing under short-term commitments, which currently incurs significantly higher costs compared to similar computing obtained through our multi-year commitments.” Essentially, the rapid growth in demand forced Anthropic to seek computing resources at elevated rates. Some of these arrangements have already been publicly disclosed. Anthropic has been paying SpaceX $1.25 billion monthly for computing services, in addition to multi-billion dollar agreements with Nscale, Akamai, and Volta. In industry terminology, these are classified as neoclouds rather than hyperscalers—smaller companies that provide more flexible computing arrangements at higher costs.
The deterioration of Anthropic’s margins in Q3 can be attributed to such agreements, and there is no indication that the company will be able to forgo these arrangements in the near future. As stated in the new filing, “We expect computing and infrastructure expenses to increase in absolute terms and remain our largest operating expense for the foreseeable future as additional committed computing becomes available to us under multi-year arrangements with CSPs, and as we continue to invest in additional computing.” Anthropic is actively pursuing strategies to reduce its computing costs. Specifically, the organization has emphasized its initiatives to establish its own computing infrastructure under a project referred to as “Anthropic Direct.” However, given the rapid growth in demand, these initiatives are unlikely to keep pace, and reliance on neoclouds will persist as the most viable option.
In a sense, these challenges can be viewed as “champagne problems.” As long as revenue continues to grow at nearly 50% per quarter, few investors are likely to be concerned about high computing costs. Should demand eventually slow, investors may find reassurance in a computing landscape that has become significantly more favorable. However, this situation serves as a reminder of the extent to which AI laboratories remain dependent on the extensive infrastructure development valued in the trillions, and how susceptible they are to disruptions in that process. It is not difficult to envision a scenario where the data center expansion slows down over the coming year if bond markets become unstable or local opposition intensifies. There is little that Anthropic can do to mitigate either of these risks, and either scenario would exacerbate an already complicated computing landscape. In the interim, the future of Anthropic and similar laboratories will hinge on computing prices.
