Anthropic has promised 518 billion dollars for compute, and most of it cannot be canceled

When people talk about the AI race, they usually talk about models. Which one writes better code, which one reasons better, which one is cheaper. Underneath every model, though, there is a much less glamorous contest that decides who can even compete: the race for computing power. A new look inside Anthropic's books shows how far one company is willing to go to win that contest, and how much of the bill it can never walk away from.

According to a confidential IPO prospectus seen by Reuters and reported on September 29, Anthropic expects to spend at least 518 billion dollars over roughly a decade on AI infrastructure with six partners. The company told investors that about 80 percent of that amount is either non cancelable or has to be paid regardless of usage. In plain words: if demand for Claude grows more slowly than Anthropic expects, most of these bills still arrive. The figures were summarized in the Reuters report, republished by Investing.com, and the related SpaceX numbers were laid out by Wccftech.

Who gets the money

The prospectus breaks the commitment down by partner, at least partly. Anthropic plans to spend at least 111.1 billion dollars with Google, 110 billion dollars with Amazon and 31.4 billion dollars with Microsoft, all under long term infrastructure service agreements that run seven to ten years and apply "regardless of usage", in the words quoted by Reuters.

On top of the cloud contracts, Anthropic carries about 161.2 billion dollars of Broadcom related equipment lease obligations, which are also described as largely non cancelable. Broadcom designs custom AI chips, and this line suggests Anthropic is not only renting general cloud capacity but also locking in dedicated hardware for years.

Then there is the most surprising name on the list. Anthropic disclosed agreements with Elon Musk's xAI, which is now part of SpaceX, that could lead to up to 84.5 billion dollars of spending on Nvidia based computing capacity through 2029.

If you add up the named items, you get roughly 498 billion dollars. The remaining gap of about 20 billion dollars to the 518 billion total is not itemized in the public reporting, so treat the breakdown as partial until the document itself becomes public.

Pay even if you do not use it

The most important detail is not the size. It is the structure. Reuters quotes the prospectus saying that if Anthropic's actual spend with Google falls short, "we must pay Google the difference", and that similar terms apply to the Amazon agreement. The Microsoft commitment "is non-cancelable except in the event of Microsoft's uncured material breach". That is lawyer language for: Anthropic can only get out if Microsoft seriously fails to deliver and does not fix it.

This kind of take or pay contract is common in energy and heavy industry. A steel mill signs for gas years in advance, a utility signs for coal. What is new is seeing it at this scale for a software company. Anthropic is essentially behaving like a power plant operator: it secures the fuel first and hopes the customers follow.

The company explains why. It told investors that future demand for advanced AI will likely exceed available supply and will be "limited principally by the availability of compute". If that is true, the companies that reserved capacity early win, and the ones that waited will be stuck in line. If it is wrong, Anthropic is left holding hundreds of billions of dollars in obligations for capacity it does not need.

How long the bills run

The time frames are long, even by tech standards. The Google commitment runs from April 2026 to July 2033. Amazon runs from May 2026 all the way to April 2036. Microsoft starts in November 2026 and runs to May 2033. The xAI agreements end in 2029.

Think about what those dates mean. In 2036 the models we use today will look ancient. The chips inside those data centers will have been replaced several times. Yet the contracts assume that Anthropic will still need, and still pay for, massive amounts of compute from the same partners. That is a bet not just on Claude, but on Anthropic still being a top tier lab a decade from now.

The SpaceX twist

The xAI line is different from all the others, and it is the one that caught the most attention. According to Reuters, these agreements are largely cancelable with 90 days notice. Anthropic can walk away from most of that capacity in about three months.

There is also a mismatch in the numbers. When SpaceX filed its own registration statement for its IPO, it said Anthropic had agreed to pay 1.25 billion dollars per month through May 2029 for capacity at the Colossus 1 and Colossus 2 data centers, with a reduced fee during the ramp in May and June 2026. As Wccftech calculated, that works out to roughly 42.5 billion dollars if you leave out the ramp months. Anthropic's prospectus now names a ceiling of 84.5 billion dollars, about twice as much.

Both figures can be true at the same time. One describes the base agreement, the other the maximum that could be spent if options are used. But the gap shows how hard it is for outsiders to judge these deals. A headline number of 84.5 billion dollars sounds like a done deal, while the fine print says much of it could disappear with a three month notice.

For SpaceX, the relationship matters a lot. Musk is racing to bring more than a million GPUs online in the Colossus 2 project, mostly Nvidia Blackwell chips, and has projected hundreds of billions of dollars in revenue from AI infrastructure. A paying customer like Anthropic helps fill that capacity. For Anthropic, it is a flexible top up, useful while demand is surging, easy to drop if it is not.

It is also a strange pairing. xAI builds Grok, a direct competitor to Claude. Musk has publicly criticized Anthropic in the past. Yet when it comes to compute, rivals rent from each other. The same is true for Google and Amazon, which both invest in Anthropic, sell it cloud capacity, and build their own competing models.

Why this matters for the IPO

Anthropic filed confidentially for an IPO with the SEC in June. The paperwork is not public yet, and Reuters reported that the listing is likely to come after the November US midterm elections. Earlier reporting from the same prospectus, covered by CNBC, showed a net loss of about 42 billion dollars for 2025. We covered a different part of the same filing, the section on catastrophic AI risks, in an earlier article on Anthropic's IPO risk warnings.

Put the loss and the commitments side by side and you see the core question every future shareholder will have to answer. Can revenue grow fast enough to cover hundreds of billions of dollars in fixed costs that are already signed? Anthropic's revenue has grown very quickly, and business demand for Claude, especially for coding, is strong. But many of its own customers can reduce or stop their spending at any time, while most of Anthropic's suppliers cannot be dropped. That asymmetry is the real risk in this story.

There is also a dependency problem the prospectus itself flags. Amazon, Google and Microsoft supply Anthropic's computing infrastructure and help distribute Claude, while they also build rival AI models. Anthropic is tying its future to companies that are partners and competitors at the same time.

What the numbers tell us about the whole industry

Anthropic is not alone. OpenAI has signed its own enormous infrastructure deals, and the biggest cloud providers are spending record amounts on data centers. Memory makers say AI demand is soaking up supply for years to come. The 518 billion dollar figure is one of the largest AI buildout commitments ever disclosed, but it fits a pattern: the leading labs are locking in capacity years ahead, and they are doing it with contracts that look more like infrastructure finance than software.

For the rest of us, there are three practical takeaways.

• Prices may stay competitive. Labs that have prepaid for huge amounts of compute have a strong incentive to keep that capacity busy, which can mean aggressive pricing for developers and businesses.
• Lock in goes both ways. Customers worry about being locked into one AI vendor. It turns out the vendors are locked in too, to their cloud and chip suppliers.
• The bubble question is now a contract question. If AI demand keeps rising, these deals will look smart. If it stalls, the companies with the most rigid commitments will feel it first. Anthropic has made a lot of rigid commitments, plus one large flexible one.

The bottom line

Anthropic has promised to spend at least 518 billion dollars on compute, and about four fifths of it is money it has to pay no matter what. The only big exit door is the deal with Elon Musk's company, which can be cut back on short notice. It is a clear statement of belief that compute, not ideas, will be the bottleneck of AI for the next decade. When the full prospectus becomes public, we will find out how carefully that belief was priced.

Sources

The commitment figures come from Anthropic's confidential IPO prospectus as reported by Reuters on September 29, 2026, republished by Investing.com and StreetInsider. The SpaceX filing comparison is from Wccftech. Loss figures are from Reuters via CNBC. The prospectus is not public, so all numbers are as reported and were not independently verified.