Anthropic Spends 58% of Revenue on Compute as Quarterly Revenue Climbs to $17 Billion

Rapid growth in an AI product can conceal an uncomfortable reality: every new user generates revenue while simultaneously consuming more and more infrastructure spending. Judging by Anthropic's updated financials, this is precisely what the company encountered ahead of its public-market debut. Its quarterly revenue rose from $11.5 billion to $17 billion, an increase of nearly 48%. But compute costs climbed from 48% to 58% of revenue, while its operating loss continued to grow.

For developers and founders, the important point is not the stock-market intrigue but a familiar product challenge. Demand can grow faster than infrastructure secured in advance. A team is then left with no good options: restrict customers or urgently purchase expensive additional capacity.

Why Rapid Growth Hurt the Economics

Anthropic spent about $10 billion on compute during the quarter. Half of that amount went toward short-term commitments that, according to the company, are considerably more expensive than comparable multiyear contracts. The reason is straightforward: the capacity reserved in advance was insufficient, while rising demand had to be served immediately.

Before the metrics deteriorated, every dollar of compute spending generated slightly more than $2 in revenue for Anthropic. That return has now fallen to $1.72. This does not mean the product suddenly lost demand. On the contrary, demand was so high that the company had to buy compute wherever it could obtain it quickly rather than cheaply.

Consider an AI service with revenue of 100 rubles. When compute accounts for 48%, 52 rubles remain after infrastructure costs for development, sales, support and all other expenses. At 58%, only 42 rubles remain. Revenue may be rising and the user chart may delight investors, yet the room for hiring and experimentation is shrinking.

What This Changes for Product Teams

The biggest mistake is tracking only the average cost per request. That metric poorly reflects reality when one user sends short messages while another runs long agentic chains involving search, code and retries. It is more useful to calculate the cost of each use case: how many tokens and how much compute time it consumes, how much revenue it generates and how the load changes during peak hours.

Teams need to determine in advance which features can be slowed down, queued or moved to a cheaper model when capacity is scarce. Limits, caching, batch processing and routing requests between models are no longer minor optimizations here; they become part of the business model. Without such mechanisms, a product's sudden success can easily turn into an infrastructure emergency with an ever-growing bill.

Where the Simple Conclusion Breaks Down

A single quarter does not yet prove that Anthropic's economics are deteriorating for the long term. Short-term contracts may serve as a temporary bridge until capacity under more favorable multiyear agreements comes online. The company is also developing its own infrastructure program, Anthropic Direct. If new capacity arrives on schedule, the cost share could decline.

But the dependence on data-center construction, credit markets and permits is not going away. For founders, the practical conclusion is stark: an AI service must be designed to scale before demand takes off. Track not only revenue and user counts, but also the gross margin of each use case, the cost of peak demand and the share of expensive reserve capacity. Without that full picture, growth can look like success right up until the next compute bill arrives.