Some news tells you more about the state of an industry than any benchmark. This is one of those: Anthropic is in talks with Meta to rent compute capacity — worth around $10 billion over two years, according to CNBC.
What we know
The talks are still very early and could fall through. The idea for the data center deal was reportedly floated by Anthropic back in June. One telling detail: Anthropic apparently wants to negotiate an early-exit option — committing only as far as necessary.
The deal wouldn’t be an outlier so much as a pattern. Just weeks ago, Anthropic struck a similar arrangement with Elon Musk’s SpaceX: $1.25 billion a month for the Colossus supercomputers, structured as a 180-day lease that either side can end with 90 days’ notice. On top of that came a $19 billion deal in early July with data center operator TeraWulf.
Why it’s awkward
Meta isn’t a neutral vendor — Meta builds frontier models of its own. So Anthropic would be renting capacity from a direct competitor. The exact same odd setup already exists: Anthropic and Google rent compute on SpaceXAI’s Colossus clusters, which train the rival Grok models.
For Meta, it still makes sense. Mark Zuckerberg said back in May that Meta was considering entering the cloud business — partly to show investors that its enormous AI spending can pay off. Meta is pouring $125 to $145 billion into infrastructure this year. An Anthropic contract would be the anchor customer for that business.
My take
What fascinates me about this story isn’t the number — it’s the structure. The leading AI labs are becoming each other’s biggest customers and biggest competitors at the same time. That only works while compute stays scarce. If you’ve got spare GPUs, you’ll sell them even to your rival.
For Anthropic, one thing matters above all: predictable costs. The company is preparing to go public, and nothing scares off investors like unpredictable compute bills. Locking in capacity long-term outweighs the awkwardness of sourcing it from a competitor.
And the real lesson? Data center construction has outgrown the AI industry itself. No single company can feed its own compute appetite fast enough. That this turns into rival-rents-from-rival deals is the most honest signal of where the bottleneck actually sits.
Sources: