Anthropic compute costs rise to 58% of revenue as demand surges

The high cost of scaling

Anthropic recently released updated financial data as it prepares for an initial public offering. The new filing provides a clearer look at the company's internal economics, specifically how much of its revenue is consumed by the power of computing. While the company saw significant growth, the data shows that the cost of running its models is becoming more expensive to manage.

The company reported that its revenue reached $17 billion in the third quarter. This was a notable increase from the $11.5 billion it reported in the previous quarter. Despite this durable growth, the company's operating losses expanded because the expenses tied to compute hardware rose even faster than the income.

The most striking detail is the shift in compute margins. In the latest quarter, compute costs took up 58% of Anthropic's total revenue. This is a sharp increase from the 48% it represented in the quarter before. This trend suggests that the company is finding it harder to maintain high margins as it scales its operations.

Scrambling for hardware

The primary reason for the shrinking margins is a surge in user demand. To keep up with the pace of requests, Anthropic had to spend $10 billion on compute in a single quarter. A significant portion of that spending went toward short-term commitments, which are much more expensive than long-term contracts.

Specifically, the filing notes that $5.0 billion of that spending was for short-term compute. These types of deals cost significantly more than the multi-year arrangements the company prefers to use. Essentially, the company had to pay a premium to ensure it had enough hardware available to serve its customers immediately.

This scramble has led Anthropic to partner with what the industry calls neoclouds. These are smaller outfits that offer more flexible hardware deals but at higher rates than the massive hyperscalers. Some of these public deals include a $1.25 billion monthly payment to SpaceX, along with multi-billion dollar agreements with companies like Akamai and Volta.

The revenue-to-compute ratio

The company's financial filings include a chart that tracks revenue for every dollar spent on compute. This metric provides a clear visual of the company's efficiency. Since the start of 2025, this ratio has been steadily growing, but it recently hit a peak and then pulled back.

At its peak, the company was generating just over 2x in revenue for every dollar spent on compute. However, that figure fell back down to 1.72x in the most recent quarter. This decline highlights the difficulty of balancing rapid expansion with the high costs of the underlying infrastructure required to power AI models.

While some analysts suggest this might be a temporary blip, it reveals the current reality of running a major AI lab. The company is still trying to find ways to lower these costs, but the immediate need for capacity is currently outweighing the desire for better margins. As the filing states, compute and infrastructure will remain the company's largest operating expense for the foreseeable future.

Future infrastructure risks

Anthropic is attempting to address these costs through a project called Anthropic Direct. The goal of this initiative is to build out its own dedicated compute base to reduce reliance on external providers. By owning more of the infrastructure, the company hopes to eventually secure better pricing and more stability.

However, the current pace of demand means that building out private infrastructure cannot keep up with the immediate needs of the market. For now, the company remains reliant on neoclouds to bridge the gap. This dependency makes the company's financial health sensitive to the prices set by these hardware providers.

There are also broader risks to the infrastructure buildout that Anthropic cannot control. If bond markets become unstable or local communities oppose the construction of new data centers, the supply of compute could slow down. If that happens, the company's messy compute picture could become even more difficult to manage as it tries to maintain its growth trajectory.