The veil of secrecy surrounding the finances of elite AI labs is beginning to lift as Anthropic moves closer toward an initial public offering. Recent updates to the company’s S-1 filings provide a rare glimpse into the brutal mathematics of running a frontier model, revealing that while revenue is skyrocketing, the cost of the hardware required to power those models is climbing just as fast. Between June and September, Anthropic saw its quarterly revenue jump from 11.5 billion to 17 billion dollars, yet this impressive growth was shadowed by widening operating losses driven primarily by computing expenses.

The numbers suggest a worrying trend for potential investors who typically look for improving profit margins before a company goes public. In the most recent quarter disclosed, compute costs consumed 58 percent of total revenue, a significant increase from 48 percent in the previous period. Essentially, for every dollar spent on compute, the return has dipped from a peak of over two times back down to 1.72 times. This suggests that scaling an AI business isn’t simply a matter of adding more users; it requires an increasingly expensive foundation of silicon and electricity.

This margin squeeze appears to be a byproduct of Anthropic’s own success. Demand surged so rapidly that the company had to scramble for immediate capacity, leading them to enter expensive short-term agreements known as neoclouds rather than relying solely on cheaper, multi-year contracts with giant hyperscalers. These desperate measures included massive payouts to firms like SpaceX and Akamai, with billions allocated to ensure their systems didn’t crash under the weight of new users. While these are often described as champagne problems—since they stem from overwhelming popularity—they highlight a precarious dependency on third party infrastructure providers.

To combat these soaring bills, Anthropic is attempting to build its own internal infrastructure through an initiative called Anthropic Direct. However, given the current pace of global AI adoption, internal projects likely cannot scale quickly enough to replace external rentals entirely. For now, the company remains tethered to a volatile infrastructure market where any hiccup in data center construction or shifts in bond markets could turn today’s messy balance sheet into tomorrow’s crisis. Until they can decouple their growth from these predatory pricing cycles, Anthropic remains largely at the mercy of whoever owns the chips.