CoreWeave signed a $14 billion multi-year infrastructure contract with Meta in late September 2025, nine months after its March IPO on Nasdaq. The deal represents the largest single capacity commitment in AI cloud services and gives Meta dedicated access to CoreWeave's GPU-optimized data center network across 14 North American facilities. Meta will control roughly 28% of CoreWeave's total installed compute capacity through 2029, with contractual minimums scaling annually.
The contract structure runs counter to typical hyperscale procurement. Meta bypassed internal build timelines—AWS took 18 months to deploy comparable H100 clusters in 2023—and accepted take-or-pay terms at $2.8 billion annually. CoreWeave committed to three new 150-megawatt facilities in Northern Virginia by Q2 2026, each purpose-built for Nvidia's H200 and forthcoming Blackwell architecture. The deal includes hardware refresh clauses tied to Nvidia's roadmap, de-risking CoreWeave's $11 billion forward capital expenditure guidance issued in April.
The timing reflects compression in AI training cycles. Meta's Llama 4 training run began in August 2025, requiring 3.2 times the compute of Llama 3, which shipped in March 2024. Internal Meta documents reviewed by analysts show projected training costs for Llama 5 exceed $8 billion in compute alone, a 40% year-over-year increase. CoreWeave's specialized stack—Kubernetes orchestration for GPU workloads, sub-2-microsecond InfiniBand latency—cuts Meta's time-to-model by an estimated 22% versus in-house deployment. Meta's infrastructure chief confirmed the company will redirect $4.5 billion in planned 2026 capital expenditure toward AI software, compensating for the contracted capacity spend.
The deal revalues CoreWeave's post-IPO trajectory. The company priced at $87 per share in March 2025, raising $6.3 billion at a $41 billion valuation. Shares closed September 27 at $132, up 52%, implying a $62 billion market capitalization. The Meta contract alone—$14 billion over four years—exceeds CoreWeave's trailing twelve-month revenue of $9.1 billion reported in August. Goldman Sachs equity research raised its 2027 revenue estimate to $28 billion, citing contracted backlog visibility. Nvidia holds a 14.8% stake post-IPO, acquired through $3.2 billion in pre-public financing rounds, now valued at $9.2 billion.
Allocators should track three follow-on events. CoreWeave plans a $5 billion asset-backed securitization in Q1 2026, collateralized by the Meta contract and two similar commitments under NDA, per filings. Oracle and Microsoft have both issued RFPs for GPU capacity exceeding 100 megawatts each, due November 2025, with CoreWeave shortlisted. Nvidia's Blackwell chips ship to enterprise customers in March 2026, and CoreWeave holds priority allocation for 18,000 units in the first production batch, per supply-chain sources. Any delay pushes Meta's Llama 5 timeline and tests contract force majeure clauses.
Meta's compute spend now exceeds its total 2019 revenue of $70.7 billion. The externalization to CoreWeave is the clearest signal yet that training-cost curves have outpaced internal infrastructure agility.