Anthropic in $10bn compute talks with rival Meta

Anthropic is in preliminary discussions to lease computing power from Meta in a deal valued at roughly $10 billion. The negotiations follow Anthropic's recent agreement to pay SpaceX $1.25 billion per month for access to Nvidia GPUs at the Colossus 1 data centre.

For Anthropic, the pursuit of external capacity underscores a critical bottleneck in the artificial intelligence sector. Despite approaching a $1 trillion valuation, the company cannot build data centres fast enough to satisfy surging demand for its Claude models, forcing it to impose usage limits on its most advanced software.

A contract with Anthropic would represent the first major step in Meta’s long-telegraphed entry into the cloud computing market. Chief executive Mark Zuckerberg has spent months signalling this pivot, noting last October that companies regularly ask to buy Meta's excess capacity "at some premium." The company is preparing to spend as much as $145 billion on capital expenditure in 2026, and monetising this infrastructure is becoming essential to justify that outlay to investors.

To execute this strategy, Meta is bringing in outside expertise, with former senior Amazon Web Services executive Dave Brown set to join the company. The infrastructure push follows a painful restructuring in May, where Meta cut 8,000 jobs specifically to redirect billions of dollars toward AI hardware.

The proposed arrangement blurs the traditional lines of competition in the AI industry. Meta would effectively be supplying the computing power needed to run Claude, a product that competes directly with Meta's own Llama models. However, the sheer scale of the compute shortage has made these unusual alliances routine across the sector.

The same hardware constraints are reshaping corporate relationships elsewhere in Silicon Valley. Google is paying SpaceX $920 million per month for GPU access while simultaneously selling its Gemini models to Meta and rationing that access due to its own capacity limits. The fundamental reality is that no single company can currently build enough computing power to meet the demands of the generative AI boom alone.