Anthropic bought $517 billion of compute after warning rivals about the risk

Anthropic has signed compute contracts worth up to $517 billion in eleven months, according to reporting by The Information. Since October 2025 the company has locked in at least 14.8 gigawatts of computing capacity, on top of the one to two gigawatts it already had, and it is planning data centres of its own.

Eight months earlier, its chief executive was warning that buying compute at this pace was the wrong thing to do.

What Amodei actually said

In early 2026, Dario Amodei said competitors racing to secure capacity "don't really understand the risks they're taking". The remark was widely read as aimed at OpenAI, which had spent the preceding year committing to infrastructure at a scale no private company had attempted before.

The figures now reported put Anthropic second in that same race rather than outside it. OpenAI's plan is reported at roughly $750 billion running through 2030, against a 30 gigawatt capacity target. Anthropic's contracted total is smaller, but many of its deals extend well past 2030, which makes a straight comparison awkward rather than simply favourable to one side.

A compute commitment is not a bank balance

The $517 billion figure is a contract ceiling, not money spent. These arrangements work like long term leases. A lab agrees to pay for a defined block of future GPU capacity, data centre space and power across several years, usually on take-or-pay terms, in exchange for guaranteed access rather than competing for scarce chips on the open market.

Anthropic's deal with SpaceX is the concrete version of the pattern. That arrangement covers more than 300 megawatts and around 220,000 Nvidia GPUs reserved for Anthropic's exclusive use. Repeat that structure across a dozen providers over several years and headline totals in the hundreds of billions stop being surprising.

Reporting on the individual contracts has described a lease of roughly one gigawatt of Microsoft Azure capacity signed in November 2025, a compute agreement with SpaceX worth up to $45 billion in May 2026, and further arrangements with Lambda and nScale. Those component figures are less firmly sourced than the total and should be read as indicative.

What changed between the warning and the signing

Anthropic has not explained the shift, and the honest position is that any answer here is inference rather than reporting. The mechanics, though, are not mysterious.

Compute is the binding constraint on both training new models and serving existing ones. In May 2026 Anthropic doubled Claude Code's five hour rate limits permanently, a change tied directly to new SpaceX capacity coming online. That is the clearest public example of infrastructure turning into product: more reserved capacity, fewer people hitting a wall mid-session.

A warning about reckless spending and a decision to spend can sit in the same head without contradiction, if the risk being described is not the cost but running out of room while a competitor does not. Amodei's phrasing was about understanding risk, not avoiding it.

The revenue does not cover it yet

Neither lab can fund these commitments from sales alone. Anthropic's annualised revenue passed $65 billion, according to Bloomberg. OpenAI was above $40 billion as of July. Those are substantial businesses by any normal measure, and they are still an order of magnitude short of the contracted numbers.

That gap is why Sam Altman has been warning about what he called "unsustainable silliness" in the buildout, singling out neo-cloud providers. His argument is that technical progress could turn today's expensive capacity into a bad bet before the contracts expire. It is an odd position for the person running the largest commitment of all, and it rhymes closely with Amodei's warning from earlier in the year.

Treat the number carefully

Anthropic has not published these figures. Neither has OpenAI. Every large lab-versus-lab number circulating in 2026 has traced back to leaked internal projections, investor materials or anonymously sourced reporting rather than an audited filing.

That does not make $517 billion wrong. It makes it directional. The right level of confidence is that it is consistent with everything else these companies have done this year, not that it is settled fact. What would change that is an on-the-record statement from Anthropic, or independent corroboration with named sourcing.

For anyone building on these models, the more useful signal is second order. Committed capacity at this scale is a bet on volume, and volume bets tend to end in aggressive pricing once the hardware is online. But capacity reserved is not capacity available, so published rate limits and prices remain the thing worth watching, not the headline number. The same caution applies to the physical buildout these contracts depend on, which is running into planning and power constraints in several countries at once.