Meta Wants to Rent Anthropic $10B of Compute

Meta is in early talks to lease computing power to Anthropic in a deal worth up to $10 billion over two years, first reported by the New York Times on July 17, 2026 and confirmed by Reuters and Bloomberg. Anthropic pitched the arrangement in June and would pay Meta in monthly increments. The interesting part is not the number. It is that Meta, which builds its own Llama models, would become the infrastructure landlord for a company it competes with directly.

• Roughly a third the size of Anthropic's SpaceX deal. The $45 billion, three-year Colossus 1 arrangement signed in May runs about $1.25 billion a month. Meta's would be closer to $415 million a month.
• Meta is building a cloud business in public. Zuckerberg said at May's shareholder meeting that cloud computing was "definitely on the table," and Bloomberg reported earlier in July that Meta was standing up a unit to sell excess capacity.
• Both sides can exit early. Terms are fluid, discussions are early stage, and the deal may not happen at all.
• Anthropic would be running on five silicon programs at once. Nvidia, AWS Trainium, Google TPUs, Azure, and now Meta's fleet.

What is actually being negotiated?

This is a lease, not an investment and not a partnership. Anthropic rents capacity on hardware Meta already owns and operates, pays monthly, and either party can walk away early. That structure matters because it is the cheapest possible way for Anthropic to add capacity: no capex, no multi-year lock-in of the kind the SpaceX and AWS deals carry, and no equity given up.

RelatedAnthropic Weighs a Samsung-Built Custom Chip for Claude

For Meta the economics run the other way. Meta plans up to $145 billion in 2026 capital spending, more than double the $72 billion it spent in 2025. Data centers built for training runs and ad ranking do not stay fully utilised every hour of every day. Selling the gaps to a customer that will absorb whatever is available is the classic cloud origin story, and it is precisely how AWS started inside Amazon.

Why would Meta arm a competitor?

Because the alternative is idle silicon. Zuckerberg has said companies approach Meta "almost every week" wanting to buy access to its models or spare compute. Meta has already hired former senior AWS executive Dave Brown, which is not the kind of hire you make to run an internal-only fleet.

There is also a strategic read. Meta's Llama line has not kept pace with the frontier, and the company's own numbers make its infrastructure more valuable than its models right now. Renting the fleet converts a depreciating asset into recognised revenue and gives Meta a second business that Wall Street knows how to value. Competing with CoreWeave and Nebius on capacity is a fight Meta can win on balance sheet alone.

What does this say about Anthropic's compute strategy?

This is the part most coverage skips. Add the Meta talks to the existing contracts and Anthropic would be running frontier inference across Nvidia GPUs, AWS Trainium, Google TPUs, Azure, and Meta's fleet simultaneously. No other lab operates at contracted gigawatt scale across that many distinct silicon programs.

That is expensive and technically painful. Every accelerator family needs its own kernels, its own quantisation behaviour, its own serving stack. Labs do not take on that cost for fun. They take it on when supply is the binding constraint and no single vendor can deliver enough. The multi-vendor sprawl is the tell: Anthropic is buying optionality against a market where being unable to serve tokens is a worse outcome than paying a premium for them.

What it means for the market

For Meta (META), a signed deal would be the first hard evidence that the $145 billion capex line has a revenue path attached to it, which is the exact question bears have been asking. Meta shares closed down more than 2% on July 17 in a broad tech selloff, so the market has not priced this in as a positive yet. The signal for investors is whether Meta discloses a cloud revenue line at all in coming quarters, not the headline deal value.

For the neocloud group, notably CoreWeave and Nebius, a hyperscaler with $145 billion of annual capex entering the spare-capacity market is a direct margin threat. Those businesses exist because hyperscalers were not selling raw capacity to labs. That gap is what is closing.

Related375ai wants to be the data layer for physical AI

The broader read is that compute contracts are becoming the real balance sheet of the AI industry. Anthropic's disclosed commitments now exceed $175 billion in headline value. That is not a research budget. That is an infrastructure company's obligation book, and it is worth tracking alongside the model releases on our AI coding leaderboard.

Our take

The $10 billion number is the least interesting thing here. What this signals is that the AI buildout has produced so much capacity, so fast, that the companies who financed it are now motivated sellers of their own strategic advantage. Meta spent two years arguing its data centers were a moat. Renting them to Anthropic reprices that moat as a utility.

It also tells you something uncomfortable about Anthropic. A lab with a $45 billion SpaceX contract and a $100 billion AWS arrangement does not go shopping for another $10 billion of two-year capacity unless it expects demand to outrun everything already signed. Either the inference market is growing faster than the public numbers suggest, or these contracts convert to capacity more slowly than the headlines imply. Both readings are worth more attention than the deal itself.

Treat this as early stage. It was pitched in June, reported in July, and either side can leave. But the direction of travel, hyperscaler fleets becoming rentable and frontier labs buying from everyone at once, is now unmistakable.

Original analysis by GenZTech. Reporting first published by the New York Times and confirmed by CNBC.