Meta expanded its cloud infrastructure commitment to CoreWeave to $21 billion, marking one of the largest single-customer AI compute contracts disclosed to date. The partnership, which began in late 2023, now dwarfs CoreWeave's earlier contracts and positions the New Jersey-based provider as Meta's primary external GPU capacity source outside its own data centers. CoreWeave simultaneously closed $1 billion in high-yield debt to finance the buildout, its third bond offering in nine months.
The deal structure commits Meta to multi-year capacity reservations across CoreWeave's NVIDIA H100 and H200 clusters, with delivery timelines extending through 2027. CoreWeave operates roughly 45,000 GPUs today, a fraction of hyperscaler fleets but concentrated entirely on AI workloads. The company's valuation reached $23 billion in its December funding round led by Coatue and Magnetar, up from $7 billion eight months prior. Meta's commitment represents nearly the entire enterprise value of its vendor, an unusual concentration that underscores both parties' confidence in sustained AI infrastructure scarcity.
The expansion matters because it reveals Meta's conclusion that internal capacity alone cannot meet its Llama training roadmap. Meta operates one of the world's largest private GPU fleets—over 600,000 H100-equivalent chips by most estimates—but CEO Mark Zuckerberg has signaled plans to scale Llama 4 to parameter counts requiring exascale compute. External capacity from specialists like CoreWeave offers faster deployment than building owned facilities, which require 18-24 months from lease signing to production. The deal also hedges against NVIDIA allocation risk; CoreWeave's direct supplier relationship and early H200 access give Meta guaranteed delivery slots that hyperscaler customers cannot secure at comparable speed.
The financing structure carries risk worth noting. CoreWeave's $1 billion bond offering priced at yields above 10%, reflecting investor caution around a business model dependent on GPU depreciation cycles and customer concentration. The company's entire revenue base now tilts heavily toward Meta, creating binary exposure if Meta's AI spending moderates or shifts to owned infrastructure after 2027. CoreWeave has mitigated this somewhat by signing smaller contracts with Mistral and Inflection, but Meta represents an estimated 65-70% of contracted revenue through the end of the decade.
Allocators should track three developments over the next six months. First, CoreWeave's planned IPO, expected in mid-2025, will disclose actual utilization rates and margin profiles that remain opaque in private markets. Second, Meta's Q1 earnings in late April will likely address AI capex pacing and whether the CoreWeave commitment displaces or supplements internal buildout. Third, NVIDIA's Blackwell chip availability in Q2 will determine whether CoreWeave can maintain its delivery advantage or whether hyperscalers close the deployment gap.
The $21 billion figure is a capacity reservation, not cash changing hands today, but it locks Meta into a vendor whose equity trades at twenty-three times trailing infrastructure value on the strength of contracts like this one.