Anthropic expects to spend more than $750 billion on cloud and it still isn't enough: its partners now see it as a rival

By TechCrunch · October 20, 2026.

Anthropic is struggling to secure enough cloud capacity from AWS, Google and Microsoft, which increasingly see it as a competitor rather than a customer. That is according to TechCrunch, based on an S-1 (the registration document filed with the US securities regulator to go public) leaked to The Wall Street Journal. Anthropic had not responded to TechCrunch's request for comment.

The key figure: according to the document, the company believes it could spend more than $750 billion on cloud contracts, some still under negotiation. Even so, that amount does not cover its needs. The leak expands on details that Reuters and the Financial Times had already reported in late September.

To make up the capacity, Anthropic leases its own data centers and works with third-party operators. But the document itself warns that these operators have a “limited operating history,” which could make them less reliable than the major cloud providers (the hyperscalers).

The financials, according to the WSJ: a net loss of $80.7 billion in the first half, against $16.3 billion in revenue over the same period. Just two customers account for nearly a fifth of sales. Last week, the company also told prospective investors that third-quarter revenue alone reached $16.8 billion, up 45% from the previous quarter. An IPO is expected in November.

The S-1 also warns, in line with previous disclosures about catastrophic risks, that Anthropic could be liable for damages caused by harmful AI behavior beyond what its insurance covers.

**Our take.** This story is not about money, but about dependence. Anthropic sells a model, but its production relies on infrastructure controlled by companies that are also developing their own models. It is a customer-and-rival relationship at once: the supplier decides how much compute to allocate to its competitor. That is the real strategic risk, more than the accounting loss.

The revenue figure helps size it. If the third quarter ($16.8 billion) already matches the entire first half ($16.3 billion), demand is growing faster than the capacity to serve it. Losing money while scaling like this is common in capital-intensive infrastructure. What is less common is that the ceiling is not demand, but supply. More broadly, this fits with the shift of compute from training to the continuous running of agents: the bottleneck is available energy and efficiency per task.

One caveat: the S-1 is a leaked document and the data comes via the WSJ; it is worth waiting for the final filing.

Winners and losers: the hyperscalers win, getting paid as suppliers while keeping the power to decide; new data center operators also win, if they prove reliable. Anyone dependent on a single cloud loses. The concentration of revenue in two customers adds another short-term vulnerability.

In the long term, the optimistic thesis holds: compute capacity is a supply constraint, and such constraints are solved through investment, as happened with other infrastructure. But the transition will be fraught, with power struggles between those who build the models and those who own the chips and the energy. For readers, the evergreen lesson is this: in frontier AI, power is measured as much in gigawatts and contracts as in benchmarks.