Meta Hands CoreWeave $21 Billion Despite $740 Million Quarterly Loss
The deal reveals hyperscalers have abandoned unit economics for capacity monopolization in the race to control AI inference infrastructure.
CoreWeave reported a $740 million net loss in Q1 2026, then secured a $21 billion commitment from Meta through 2032—the clearest signal yet that hyperscalers will absorb massive burn rates to lock in GPU compute capacity years ahead of demand.
The expanded agreement, announced in April, brings Meta’s total CoreWeave commitment to $35.2 billion, according to BranderGroup, and now represents roughly 40% of CoreWeave’s $99.4 billion revenue backlog. Meta is paying for infrastructure that won’t fully deploy until 2028-2030, betting that control over inference capacity—not training clusters—will determine who wins the AI platform war against OpenAI, Google, and Microsoft.
CoreWeave’s economics defy traditional cloud infrastructure logic. The company generated $2.078 billion in revenue with a 112% year-over-year increase, per 24/7 Wall St., yet maintained a 56% adjusted EBITDA margin while burning cash due to massive capital deployment. The paradox resolves when viewing capex as the real cost center: CoreWeave spent $7.695 billion in Q1 alone and raised full-year 2026 guidance to $31-35 billion, meaning the company deploys roughly $2.60 in capital for every dollar of revenue. Traditional cloud providers operate at 0.3-0.5x capex-to-revenue ratios. CoreWeave is building capacity five years ahead of cash conversion.
The Take-or-Pay Bet
Meta’s commitment is structured as a take-or-pay contract: the company will pay CoreWeave whether it uses the capacity or not, effectively financing CoreWeave’s buildout in exchange for guaranteed access. This arrangement makes sense only if Meta believes inference demand—serving billions of daily AI interactions across Facebook, Instagram, and WhatsApp—will exceed what its own datacenters can handle, according to Global Datacenter Hub.
“They’re going to continue to do it themselves, but they’re also going to continue to do it with us. There’s just too much risk not to.”
— Michael Intrator, CEO, CoreWeave
The deal centers on Nvidia’s Vera Rubin platform, which delivers 50 petaflops of FP4 inference per GPU versus Blackwell’s 10—a fivefold improvement specifically optimized for serving models rather than training them. CoreWeave will deploy these systems across multiple geographies starting in H2 2026, with CoreWeave Investor Relations confirming delivery timelines extending through 2032. Meta is paying now for compute it won’t fully utilize until Llama 5 or 6 are serving production traffic at global scale.
Hyperscaler Capex Acceleration
Meta’s $35.2 billion CoreWeave commitment sits atop its own internal AI Infrastructure budget. The company guided 2026 capital expenditures of $115-135 billion in May, with CoreWeave representing supplementary capacity insurance rather than a replacement for owned infrastructure. Combined hyperscaler spending from Google, Amazon, Meta, and Microsoft is expected to approach $700 billion in 2026, up from approximately $365 billion in 2025.
Jensen Huang’s recent assertion that hyperscaler AI spending will reach $3-4 trillion by decade’s end, stated during NVIDIA’s Q1 FY2027 earnings call, provides the macro context. Current combined spend sits around $1 trillion annually and is accelerating, not plateauing. Huang’s thesis—that compute itself is the product, not merely infrastructure supporting a product—explains why Meta is locking in capacity at prices that may prove uneconomical if utilization lags or model efficiency improves faster than expected, per CNBC.
Unit Economics Question
CoreWeave’s model works only if hyperscalers continue paying for capacity years before they need it and AI inference demand grows fast enough to justify current buildout rates. The concern isn’t the quarterly loss in isolation—the company is pre-revenue on much of its contracted backlog—but rather the $31-35 billion annual capex commitment creates a financing treadmill. CoreWeave must either convert backlog to cash faster, raise additional debt (it already carries significant leverage from prior datacenter financings), or accept dilution through equity raises.
- Meta’s $21B commitment validates capacity reservation as strategic insurance, not cost optimization.
- CoreWeave’s 56% EBITDA margin coexists with operational losses because capex ($7.7B/quarter) dwarfs operating expenses.
- Take-or-pay contracts shift risk from provider to customer but require customers to believe utilization will exceed internal capacity.
- NVIDIA Vera Rubin’s 5x inference advantage over Blackwell makes distributed serving economically viable at Meta’s scale.
The broader AI infrastructure market is bifurcating. Hyperscalers are building owned capacity for baseline workloads while using specialized providers like CoreWeave for peak demand, geographic expansion, and risk mitigation. OpenAI committed $22.4 billion to CoreWeave; Microsoft, Google, and Anthropic have similar arrangements. This creates a new category—neocloud infrastructure—that didn’t exist three years ago and now commands $99.4 billion in contracted backlog from a single provider.
CoreWeave’s stock traded at $137.98 as of May 6, up 24% year-to-date while the S&P 500 declined roughly 1%, suggesting public markets believe the capacity land grab justifies current burn rates. The company’s weighted average contract length sits around five years, meaning revenue recognition and cash conversion lag capex deployment by years, not quarters.
What to Watch
CoreWeave’s Q2 2026 earnings in early August will reveal whether capex deployment accelerates further or stabilizes. Meta’s next earnings update will clarify if the company adjusts its $115-135 billion internal capex guidance upward, signaling even tighter capacity constraints. The critical tell: if other hyperscalers announce similar multi-year, take-or-pay deals with CoreWeave or competitors like Lambda Labs or Crusoe, it confirms capacity monopolization is the dominant strategy. If deals slow, it suggests hyperscalers believe their internal buildouts are sufficient or that model efficiency gains reduce inference compute requirements faster than expected.
The unit economics question remains unanswered. Can CoreWeave convert its $99.4 billion backlog to positive cash flow before debt service consumes margins, or does the company require continuous refinancing to bridge the gap between capex deployment and revenue recognition? Meta’s willingness to pay $21 billion for infrastructure that won’t fully deploy until 2028-2030 suggests hyperscalers believe the alternative—losing the AI platform war due to compute constraints—is worse than absorbing losses today.