Anthropic Faces Nearly $1 Trillion in Compute Commitments

Anthropic is under growing financial pressure to keep its revenue climbing as it races to build more capable AI models. An analysis based on the company's financial disclosures estimates its long-term compute and infrastructure commitments at about $850 billion over 20 years, a figure that could approach $1 trillion once unfinished deals are counted.

Revenue surges, commitments dwarf it

The filings show revenue of $16.3 billion in the first half of 2026, 14 times the level of the same period in 2025, driven largely by demand from business customers. Anthropic booked accounting losses of $80 billion, but the analysis says much of that does not directly reflect operating performance. Excluding those losses and some other items, including employee compensation, adjusted income came to roughly $1.62 billion.

The obligations are enormous. As of June 30, 2026, and based on unaudited figures, the company has at least $265.6 billion in non-cancellable purchase agreements, plus $321.4 billion in leases for infrastructure and computing capacity. It has also signed new compute and infrastructure deals worth about $245 billion, with an estimated $125 billion to $150 billion in further agreements still unfinished. Some variable commitments are not included.

Payments tied to these deals could reach around $90 billion a year between 2027 and 2030. Keeping today's ratio of compute costs to revenue would require $188 billion in revenue.

An enterprise focus

Unlike some rivals, Anthropic has concentrated on models for coding, cybersecurity and enterprise needs rather than a broad range of consumer apps. About three-quarters of its customers spent at least $100,000 a year in the first half of 2026. Competition is intensifying, though, from other frontier labs and open-source developers, while companies like Google can embed AI into products businesses already use.

In the first half, Anthropic spent about $9.7 billion on model development, versus roughly $2.3 billion on compute and infrastructure for serving customers.

Part of its strategy relies on using advanced models to speed up the building of better ones, a process known as recursive self-improvement. The company warns in its disclosures that model capabilities could advance faster than humans can develop and verify safety, alignment and security measures. That raises questions if faster development becomes a commercial necessity. If revenue growth stalls or expected capability jumps fail to arrive, meeting the long-term obligations and reaching sustainable profit would get much harder.