Anthropic can still cancel or drop 186 billion dollars of its computing power deals
Anthropic has committed to 900.9 billion dollars in spending, mostly on computing power. About 186 billion of that can be cancelled or still depends on conditions, its updated stock market filing shows. Chief executive Dario Amodei sees buying extra capacity at short notice as a kind of insurance.
Anthropic, the maker of Claude, is going public this week, and investors want to buy far more shares than are on offer. the company submitted a second updated version of the document it uses to register with the US securities regulator for its stock market listing. It shows how many of its computing power agreements Anthropic can still cancel.
The previous version, from 2 November, already showed that Anthropic took in almost 17 billion dollars in revenue in three months and had committed to about 895 billion dollars in contracts, mostly for computing power and data centres. In the new version that figure is 900.9 billion dollars (about 806 billion euros). Around 71 billion dollars of it has to be paid within a year after the end of September 2026; almost all of the rest falls between 2031 and 2045.
What can still be cancelled
Of all the commitments together, roughly 186 billion dollars (about 168 billion euros) can be cancelled or still depends on conditions, according to Anthropic. One example of such a condition is that a supplier first has to secure the financing or a building for a data centre. In the version of 2 November this was still almost 214 billion dollars. The large contracts with Google, Amazon, Microsoft and Broadcom are mostly fixed: there, Anthropic has to buy at least a set amount, even if it uses less.
A notable part that can be cancelled is the computing power Anthropic rents from SpaceXAI, formerly Elon Musk's xAI. As SpaceX already revealed at the end of October, this is worth up to 120.5 billion dollars (about 109 billion euros) until 2029. From January and March 2027, either side can end these agreements with ninety days' notice. an analyst claimed on X that Anthropic does not want to extend the contract beyond mid-2027. Neither company has confirmed this. Musk replied that SpaceX would be happy to take the computing power back, but that he expects the opposite to happen.
Buying at short notice as insurance
, in a conversation with Jamie Dimon, the head of the bank JPMorgan, chief executive Dario Amodei compared buying computing power at short notice to insurance. Anthropic commits to relatively little computing power far in advance and buys more when demand is higher than planned. So far, that computing power has always been available, he said. JPMorgan is helping to manage part of the stock market listing.
Computing power bought at short notice does cost considerably more than computing power from contracts that run for several years. The loss of 567 million dollars on its normal business in the third quarter came mainly from 5 billion dollars of such short-term purchases.
The share price is expected to be set . A day later, trading starts on the US stock exchange Nasdaq, under the code ANT.
What does this mean for the road to the singularity?
Computing power is increasingly traded the way electricity is: a large part is fixed years in advance, and a smaller part is bought at short notice when demand peaks. Agreements that can be cancelled, such as the one with SpaceX, sit in between. They give Anthropic a choice in 2027: carry on if demand keeps growing, or hand the chips back if the other contracts deliver enough. That freedom also works the other way. SpaceX builds its own AI models with xAI, so chips that Anthropic hands back can go straight to its own AI company.
For AI companies this is a tricky calculation. Those who commit too much in advance pay for chips that sit idle. Those who commit too little have to buy expensive extra capacity in busy times, or fall behind the competition. In a peak, computing power can end up with whoever pays the most, leaving smaller companies at the back of the queue. But the more easily computing power can move between companies, the fewer chips stay unused. That lowers the cost of every answer an AI model gives, and makes it easier to train ever more powerful models and make them available to many people.
What is the technological singularity? Read the explanation →