Is it difficult to make money with AI? Anthropic lowers gross margin guidance: Revenue increased 12 times, but cannot offset a 23% surge in reasoning costs

The developer of the Claude chatbot, Anthropic, is facing profitability challenges even as its revenue soars.

According to The Information on Wednesday, the company lowered its gross margin expectation for 2025 to 40% last month, a decrease of 10 percentage points from previous optimistic forecasts, due to the inference costs of running AI models on Google and Amazon servers being 23% higher than expected. This data highlights the difficulty AI companies face in escaping reliance on cloud service providers and the cost pressures that come with it while pursuing scale growth.

Despite the gross margin falling short of expectations, Anthropic's financial performance still shows strong growth momentum. According to insiders, the company expects its revenue to reach $4.5 billion in 2025, nearly a 12-fold increase from $381 million in 2024. This growth is primarily attributed to the success of its Claude Code coding tool and Cowork office assistant in the enterprise market, with at least nine clients expected to spend over $100 million annually, including Microsoft's spending, which is projected to reach $500 million.

In contrast, competitor OpenAI expects a gross margin of about 46% in 2025, including the inference costs for both paid and free ChatGPT users. OpenAI's revenue is projected to exceed $13 billion in 2025, and while it still leads in absolute scale, Anthropic is narrowing the gap—current monthly revenues for the two companies are $1.7 billion and $750 million, respectively. This comparison indicates that operational efficiency and cost control have become key factors in determining the outcome of the AI commercialization race.

These figures are significant for investors. Although Anthropic expects an EBITDA loss of about $5.2 billion in 2025, and OpenAI anticipates a pre-tax loss of $21.2 billion, equity investors remain optimistic. Anthropic is negotiating to raise over $10 billion at a pre-money valuation of $350 billion, led by Singapore's GIC and Coatue Management; OpenAI is seeking to raise up to $10 billion at a valuation of about $750 billion.

Inference Costs Erode Profit Margins

Anthropic's gross margin dilemma stems from its reliance on cloud service provider infrastructure. The company calculates its gross margin by deducting inference costs and other sales costs, and the inference costs—i.e., the expenses of running AI models for paying customers on Google and Amazon servers—are 23% higher than expected. If the inference costs for free Claude chatbot users are also included, Anthropic's gross margin would further drop to about 38%.

This performance, while far below traditional software companies, shows significant improvement compared to a gross margin of -94% in 2024. Anthropic had previously projected that its gross margin would exceed 70% by 2027, and OpenAI also expects to reach at least a 70% gross margin by 2029, approaching the levels of publicly traded software and cloud computing companies.

However, in addition to inference costs, both companies also face substantial model training costs—Anthropic expects its training costs to be about $4.1 billion in 2025, an increase of about 5% from summer forecasts. OpenAI's spending on model training computing is expected to reach $9.4 billion last year. These training costs are not included in the gross margin calculation, but they further increase the difficulty of achieving net profit.

To control costs, both companies are taking measures to manage hardware. Anthropic recently agreed to purchase Google's tensor processing units for $21 billion to reduce computing costs. OpenAI is developing server chips for inference, primarily used to run ChatGPT, as an alternative to expensive Nvidia chips. OpenAI's Chief Financial Officer Sarah Friar stated on Wednesday that the inference chip has been "taped out," meaning the company has delivered the final design to the chip manufacturer.

Intense Competition in the Enterprise Market

Anthropic's revenue growth is primarily driven by enterprise customers. It is estimated that about 86% of the company's revenue in 2025 will come from selling AI models to enterprises via API, with the remainder from subscriptions to the Claude chatbot. The recent success of Claude Code and Cowork in coding and office scenarios has been compared by industry insiders to the hype generated by OpenAI in early 2023. Microsoft is expected to pay Anthropic up to $500 million for GitHub CoPilot, and popular coding tools like Cursor and Cognition are also major clients.

OpenAI's enterprise business may still be larger. According to insiders, about 40% of OpenAI's revenue comes from enterprise customers, including API and enterprise chatbot sales, which means its enterprise customer revenue is about $5.2 billion, while Anthropic's is $3.9 billion. OpenAI's ChatGPT has approximately 900 million weekly active users, of which about 95% are free users. The company has just announced plans to launch an advertising business to subsidize the costs for free users.

Despite the unprecedented sales growth of both companies, some lending institutions are cautious about providing loans for data center projects involving these companies, as they expect it will take until the end of the century to generate free cash flow. However, this has not dampened the enthusiasm of equity investors. Nvidia and Microsoft have previously committed to invest up to $10 billion and $5 billion in Anthropic, respectively, and investors valued the company at $170 billion in the $13 billion funding round last September