Meta and CoreWeave Sign 21 Billion Dollar Deal to Expand Specialized AI Infrastructure

Meta Platforms and CoreWeave have announced a landmark expansion of their partnership, with Meta committing an additional $21 billion for dedicated cloud capacity. This agreement ensures Meta has access to cutting-edge hardware through December 2032 to support its growing artificial intelligence ambitions. The deal solidifies CoreWeave's position as a dominant force in the specialized cloud market.

This fresh order builds on a previous $14.2 billion contract, bringing Meta's total investment with the provider to roughly $35 billion. The expanded capacity will focus heavily on AI inference workloads, which are necessary for running consumer-facing models like Llama in real-time. Meta's strategy involves a portfolio-based approach to mitigate supply chain risks while diversifying its compute resources.

A key highlight of the partnership is the early commercial deployment of NVIDIA's next-generation Vera Rubin architecture. These systems promise significantly faster processing speeds compared to current Blackwell chips. CoreWeave will also integrate GB300 systems across its global data center network to meet the high-density compute requirements of modern AI applications.

CoreWeave, which went public on the Nasdaq as CRWV in early 2025, continues to see explosive growth in its contracted revenue backlog. To fund this massive expansion, the company is raising over $4 billion in new debt to support an estimated $30 billion in infrastructure expenditures. These financial moves are designed to keep pace with the unprecedented demand for specialized GPU compute services.

For CoreWeave, the deal marks a significant step in diversifying its customer base and reducing its historical reliance on Microsoft. CEO Michael Intrator noted that no single client will account for more than 35 percent of total sales following this expansion. The firm has also secured multi-billion dollar deals with other major players including OpenAI and Anthropic to power their respective models.