Meta Platforms has committed $21 billion to CoreWeave for AI cloud infrastructure across a multi-year term, the largest single commitment between a hyperscale consumer platform and a specialized compute provider. The arrangement expands an existing partnership announced last year and positions CoreWeave as Meta's primary external GPU supplier for training and inference workloads tied to Llama models and generative AI product rollouts.
CoreWeave disclosed the figure Tuesday without publishing specific capacity or timeline details. The company, which went public in January at a $19 billion valuation, has now secured more than $30 billion in forward commitments since late 2023. Meta's stake accounts for roughly two-thirds of that backlog. The deal does not appear to include equity participation or exclusivity, but the scale suggests Meta is underwriting a meaningful portion of CoreWeave's datacenter buildout through 2027. CoreWeave operates 28 datacenters globally and has confirmed plans to add 14 facilities by year-end 2025, many co-located near Meta's own regional hubs.
The move reflects a structural shift in how large platforms source compute. Meta continues to operate proprietary clusters for its core recommendation and ads infrastructure, but the Llama ecosystem—now powering third-party applications and consumer-facing generative tools—requires elastic capacity that internal capital cycles cannot match. CoreWeave offers Meta access to Nvidia H100 and H200 GPUs on shorter lead times than traditional cloud providers, with contract terms that allow Meta to scale inference workloads without committing to long-term reserved instance pricing. The company's focus on bare-metal GPU orchestration and its willingness to offer fixed-price multi-year contracts make it a cleaner counterparty than AWS, Google Cloud, or Azure for workloads where margin pressure and utilization volatility are high.
The $21 billion figure is notable less for its size than for its specificity. CoreWeave has traded on forward commitments to reassure public equity holders that its revenue pipeline is durable, but the company has not previously disclosed customer-level contract values. Meta's willingness to have the number public suggests confidence in its AI product roadmap and a desire to signal to the market that it is securing capacity ahead of anticipated demand from consumer-facing generative features launching later this year. The arrangement also derisks CoreWeave's expansion into European and Asian markets, where Meta's regional presence provides natural anchor tenancy for new builds.
Allocators should track three developments over the next six months. First, whether CoreWeave's datacenter utilization rates—currently estimated near 78 percent—hold steady as the Meta capacity comes online, or whether the company begins to show signs of overbuild relative to near-term demand. Second, whether other large language model operators follow Meta's lead in committing to long-term infrastructure partnerships rather than building proprietary clusters. Third, whether CoreWeave's cost of capital improves as its backlog becomes more concentrated. The company raised $7.5 billion in debt last year at rates near 8 percent, and a visible anchor customer may allow it to refinance at lower spreads.
Meta's infrastructure spending for 2025 is now tracking above $60 billion, with roughly one-third allocated to external cloud and colocation. That ratio is climbing.