The AI labs that built billion-dollar models are quietly admitting they can't afford to run them.
The Summary
- Meta is negotiating to lease up to $10 billion in computing power to Anthropic over two years, opening a potential cloud business line for the social media giant
- The deal would ease Anthropic's "desperate hunt for compute" as AI labs shift from owning infrastructure to leasing it
- Bitcoin miner TeraWulf separately signed a $19 billion lease with Anthropic for AI data center capacity, signaling a broader industry shift
- Meta's infrastructure spending has hit $145 billion, creating excess capacity the company now wants to monetize
The Signal
Meta is in active talks to lease computing infrastructure to Anthropic in a deal that could reach $10 billion over two years. The arrangement would mark Meta's entry into the cloud computing business, directly competing with AWS, Google Cloud, and Microsoft Azure. For Anthropic, it solves an existential problem: how to run Claude without drowning in capital costs.
The numbers tell the real story. Meta has spent $145 billion on infrastructure, much of it for AI training and inference. That's enough compute to power a small country's worth of AI models. Now Meta is looking at those server farms and seeing a new revenue stream. Instead of just running their own models, why not rent the excess capacity to competitors?
"The AI labs that raised billions to build models are now admitting they can't afford to keep the lights on."
This isn't just a Meta-Anthropic story. TeraWulf, a Bitcoin mining company, signed a separate $19 billion lease deal with Anthropic for AI data center capacity. Read that again. A crypto miner is now in the AI infrastructure business, and the deal is nearly double the size of the Meta negotiation. The shift from ownership to leasing is industrywide.
Why the shift? Three reasons:
- Capital efficiency: Building data centers requires billions upfront. Leasing spreads costs over years.
- Flexibility: AI model architectures change fast. Leasing lets labs upgrade without stranded assets.
- Focus: Anthropic wants to build better models, not manage HVAC systems and power contracts.
Prediction markets are pricing Anthropic's valuation at $1.25 trillion by December 31 with 91% confidence. That's higher than Meta's current market cap. If Anthropic hits that number while leasing most of its compute, it proves a new model: you can be the most valuable AI company in the world without owning a single GPU.
For Meta, this is brilliant asset utilization. They've already spent the $145 billion. Every dollar Anthropic pays is incremental margin. Meta gets to compete in foundation models AND infrastructure, hedging against the risk that someone else's model wins. If Claude becomes the standard, Meta still profits from hosting it.
The Implication
The agent economy runs on compute. If the best AI labs are shifting to leased infrastructure, that changes who has power in Web4. It's no longer just about who builds the best model. It's about who controls the GPUs those models run on.
Watch for more deals like this. OpenAI, Cohere, and every other foundation model company faces the same capital constraints Anthropic does. The cloud providers that win this wave won't be the ones with the best API docs. They'll be the ones with the cheapest kilowatt-hour and the most GPUs sitting idle. Meta just became one of them.