Anthropic grew quarterly revenue by 47%, but compute costs rose even faster

Anthropic is rapidly evolving from a research lab into an infrastructure company. In the third quarter, its revenue increased from $11.6 billion to nearly $17 billion, or about 47%. But compute became even more expensive: the corresponding costs rose 77%.

For developers, this is not merely a stock-market curiosity. If your product runs on Claude, Anthropic's financial model affects API pricing, limits, volume discounts, and access to new models. The company is growing extremely quickly, but it is also buying enormous amounts of capacity in advance and temporarily renting some infrastructure at premium rates.

Growth is real, but it must be measured carefully

Anthropic's quarterly filing reports revenue of about $17 billion, up from $11.6 billion in the previous quarter. That represents growth of approximately 47%. The original analysis also cites a 55% increase in “net” revenue after certain cost adjustments. When assessing the business, it is safer to keep both figures in mind without conflating the author's calculation with the metric reported in the S-1 filing.

Consider a simple example. Suppose an AI service receives 100 rubles from customers and spends 40 on model requests. If sales rise to 147 rubles while compute costs climb to 71, revenue will look excellent, but the margin cushion will barely change. That is why sales growth alone does not prove that the economics of an AI platform have stabilized.

Anthropic does have a major strength: over nine months, operating cash flow turned positive and reached $6.6 billion. At the end of September, the company also held $95.6 billion in cash and $36.3 billion in short-term investments. The issue, then, is not a lack of money but the cost of further scaling.

The main risk is hidden in compute contracts

Anthropic disclosed major commitments for future infrastructure purchases. About $69.8 billion in payments is due over the next 12 months. The company obtains some of its capacity through short-term agreements signed during the GPU shortage.

One example is a contract for xAI infrastructure linked to SpaceX. Public data indicates a monthly payment of $1.25 billion for access to roughly 325,000 NVIDIA accelerators. CCIR's calculation puts the cost at about $5.27 per GPU-hour. This is not a universal market price: the chip model was not disclosed, and the agreement can be terminated with 90 days' notice after an initial period. The rate therefore cannot be compared directly with long-term contracts for specific GPU generations.

Short-term leasing helped Anthropic secure capacity quickly while its own infrastructure was not yet ready. But that growth accelerator is expensive. As long-term sites come online, the company will likely replace temporary contracts with cheaper capacity. That would improve margins if demand for Claude continues to grow and the new data centers do not end up underutilized.

What product and engineering teams should do

Do not build a budget on the assumption that the price of a request is fixed forever. Calculate the cost of key use cases separately for your primary and fallback models. Add request routing: send complex tasks to a powerful model and simpler ones to a cheaper model. Cache repeated responses, limit context, and measure the cost of a completed user action rather than the cost of one million tokens.

Another practical step is to test portability in advance. If a product cannot switch to another model without rewriting half the system, every new price or usage limit becomes an infrastructure crisis of your own.

Anthropic's report looks strong: sales are growing, cash flow is positive, and its liquidity reserve is enormous. But compute costs are still rising faster than quarterly revenue. For the market, this tests the company's ability to turn scarce GPUs into sustainable margins. For developers, it is a reminder that AI product architecture must withstand changes in prices, models, and providers.