Anthropic is in early talks to lease GPU computing capacity from Meta in a deal valued at roughly $10 billion over two years, according to a report from MLQ and people familiar with the matter. The agreement, if completed, would be the largest single compute-lease transaction in the AI industry, and it would give Anthropic a stockpile of capacity as it prepares for what is expected to be one of the largest technology listings in years.
The talks are at an early stage and could still fall apart, the people said. The companies have not commented on the report, and the terms, including how much capacity would be covered and whether the deal would include power and network access, have not been settled. What is clear is why Anthropic wants the machines.
Training frontier models requires tens of thousands of accelerators at a time, and serving them to customers requires a fleet that keeps growing with demand. Anthropic has long relied on cloud deals with Amazon and Google, which provide the bulk of its capacity, but its chief executive has said publicly that compute is the binding constraint on what the company can build. A two-year lease at Meta’s scale would lock in capacity without the risk of building data centers from scratch.
Meta, for its part, has built some of the largest AI clusters in the world to train its Llama models. The company has spent tens of billions of dollars on data centers and accelerators, and analysts have long noted that its capacity exceeds what its own models consume at any given moment. Selling spare capacity to a rival would turn a cost center into a revenue stream, an unusual move for a company that has treated its infrastructure as a competitive asset.
The deal would also be a statement about the compute market. Nvidia’s accelerators were the scarce resource of the past two years, and companies hoarded them like gold. Supply has loosened since, but the largest players still sign multi-billion-dollar commitments years in advance. A $10 billion lease would signal that capacity itself has become a financial instrument, with owners renting it like real estate and users buying it like insurance.
For Anthropic, the timing points to the listing. The company has been reported to be worth about $965 billion in private markets, and investors have pushed it to show that it can secure the compute needed to hold its position at the frontier. A lease of this size would answer the question that has followed every AI company planning to go public: do you have enough machines? The company’s valuation, already among the highest in the industry, depends on the answer.
There is also a competitive angle. Meta and Anthropic are rivals in the model business: Meta’s Llama is the leading open-weight family, and Anthropic’s Claude competes with it for developers. A compute deal between them would be a business arrangement between competitors, the kind of thing the industry is starting to see more of as AI’s costs concentrate among a handful of giants.
Analysts were split on what the talks reveal. Some read them as a sign that GPU supply is abundant enough that even rivals can trade capacity. Others read them as evidence that the largest AI companies are desperate for machines despite record spending. “The interesting part is that a company with Anthropic’s resources would rather rent from a competitor than wait for its own capacity,” one analyst said.
The structure matters as much as the size. Leases of this kind typically bundle accelerators, power, and network access, and they run for years to match the useful life of the hardware. A two-year term is short for such an arrangement, which suggests the talks could evolve into something longer, or into a broader partnership between the two companies, according to people familiar with the matter. Meta has also been expanding its own model efforts, and any deal would need to protect Meta’s ability to train its own frontier models.
For Meta, the lease would be a first: renting capacity to a direct competitor at industrial scale. The company’s chief executive has said Meta’s investment in AI infrastructure is a multi-year commitment, and the economics of that spending improve if the machines are working for someone else when Meta’s own models do not need them.
The outcome will depend on price, terms, and whether Meta’s own models need the capacity first. What is already clear is that the AI industry has entered a phase where the machines themselves are traded like commodities. Anthropic’s reported willingness to pay Meta $10 billion for two years of compute shows that the race is no longer just about models. It is about who controls the hardware.
