Anthropic in early talks with Meta for $10B compute deal; third major infrastructure partnership

Anthropic is in preliminary talks with Meta to lease AI computing power in a potential $10 billion deal, according to CNBC reporting on July 17. The discussions represent Anthropic's third major compute-capacity agreement in weeks, following deals with Elon Musk's SpaceX (300 megawatts for $1.25B/month through May 2029) and Google/Broadcom. Meta CEO Mark Zuckerberg said in May that the company was evaluating entry into cloud computing to monetize its AI infrastructure investments and justify massive capex to investors. Meta has signaled $145 billion in 2026 capex, much of it AI-focused.

The talks underscore a critical bottleneck: major AI labs remain capacity-constrained on NVIDIA chips despite record procurement. Anthropic has usage limits on its most advanced models (Fable) and is scrambling to secure compute through partnerships rather than just buying more hardware. SpaceX's earlier disclosure via its S-1 filing revealed xAI's compute arrangement with Anthropic—300MW for $1.25B monthly—suggesting the companies view compute leasing as more reliable than waiting for new chip orders amid geopolitical export restrictions.

The emergence of a secondary market for compute capacity marks a strategic shift: AI labs are now aggregators of third-party compute, not just purchasers. Meta's willingness to monetize its data center capacity signals the company sees a multi-year arbitrage opportunity. The three partnerships (SpaceX, Google, now potentially Meta) reflect Anthropic's strategy to reduce single-source dependency on NVIDIA and lock in capacity ahead of its IPO.

For infrastructure planners, the proliferation of compute-sharing deals between AI labs signals that traditional cloud providers (AWS, Azure, GCP) are losing negotiating leverage to custom partnerships. The $10B ask suggests Meta is pricing its compute near market rates—implying NVIDIA-level pricing power in a secondary GPU market.