CoreWeave’s $14.2B and $21B Meta Deals Don’t Yet Reconcile

On February 24, 2026, Bloomberg reported that GPU-cloud operator CoreWeave was seeking about $8.5 billion from banks against a Meta contract worth up to $14.2 billion. On April 9, CNBC reported an additional $21 billion Meta commitment for 2027–2032, overlapping the earlier contract’s final years. CNBC called it additional; the public record still does not show how the two agreements divide capacity or which cash flows would support the loan.

Five months before Bloomberg’s report, Reuters said Nvidia had promised to buy any CoreWeave cloud capacity left unsold through April 13, 2032, under a $6.3 billion agreement. Nvidia would collect on chip sales and stand behind the unsold output.

CoreWeave’s rack now sits inside four claims: lenders finance chips and facilities, Meta commits demand, Nvidia covers some unused capacity, and grid rules can determine how many hours the equipment runs. The operator keeps what remains after those counterparties set their terms.

CoreWeave made the chip carry the debt

In August 2023, Reuters reported that CoreWeave had raised $2.3 billion in debt collateralized by Nvidia chips. It also raised $421 million in equity that year. Lenders financed the expansion because the GPUs themselves could support a claim.

By February 2026, GPU-backed debt and special-purpose vehicles had spread beyond that first facility. Reporting described CoreWeave moving $2.6 billion of data-center construction debt off its balance sheet through special-purpose vehicles. The same account said a CoreWeave-tied data center raised $900 million through five-year high-yield bonds priced to yield 7.5%.

Lenders could finance selected chips, facilities and cash flows instead of making one indivisible corporate bet. The GPUs became collateral, the facilities became borrowers, and future compute sales serviced the debt.

Meta’s contract does what an hourly price cannot

A GPU-hour can produce revenue, but it cannot promise that the next hour will sell; Meta’s contract could. Bloomberg reported that CoreWeave was seeking the approximately $8.5 billion loan from banks including Morgan Stanley and Mitsubishi UFJ Financial Group. The contract gave them a named customer and a maximum value against which to assess the proposed capacity.

CNBC said the earlier agreement runs through 2031 and the additional commitment runs from 2027 through 2032. That leaves five calendar years of overlap. No disclosed terms show whether the later agreement covers distinct capacity, amends the earlier contract or cross-references it, so the two headline figures cannot yet be summed for underwriting.

A maximum contract value also does not establish the timing or certainty of every cash receipt. The unresolved record is narrower and more consequential: one report identified a contract worth up to $14.2 billion as support for the loan plan, while another announced a further $21 billion without publishing enough terms to reconcile the obligations.

Meta and Nvidia pull risk from opposite sides of the rack. Meta supplies contracted demand; Nvidia makes unused capacity less costly to carry. Together, their agreements strengthen CoreWeave’s case for building capacity before the corresponding revenue arrives.

CoreWeave later reported $2.08 billion in first-quarter revenue, up 112% year over year, and a $99.4 billion revenue backlog. That scale makes the composition of the backlog material: lenders and suppliers need to know which customers signed, how long they committed, and which cancellation or delivery terms govern the cash flow.

Nvidia is moving from vendor to counterparty

On February 26, 2026, The Information reported that Nvidia had disclosed $3.5 billion in guarantees to companies leasing land, power and data-center facilities, four times its third-quarter level. On July 2, it reported that Nvidia would rent back unused GPUs from younger cloud providers Firmus and Sharon AI in exchange for a share of their revenue.

Nvidia can reduce the financing and utilization risks that stop smaller operators from scaling, then collect part of the revenue its guarantees make possible. The supplier’s exposure no longer ends when the chips ship.

PJM can reach the margin before the GPU does

A data center is one computer assembled at building scale. Chips cannot be separated from networking, storage, cooling, electricity and operations because each component constrains the useful output of the others. A full GPU rack without dependable power sits unavailable while financing costs accrue.

PJM Interconnection proposed requiring large data centers either to bring their own generation or curtail electricity use to prevent large-scale outages. A curtailment rule can reach the same hours that customer contracts and lenders expect CoreWeave to sell.

Oracle shows why occupancy alone cannot settle the margin question. Internal documents indicated that its Nvidia cloud server-rental business generated about $900 million in revenue and $125 million in gross profit, a 14% gross margin compared with roughly 70% companywide. Oracle’s scale and cloud experience did not turn expensive GPU rentals into ordinary software economics.

Grid operators and lenders compound the pressure on the same rack. Costlier electricity leaves less cash for debt service, while higher borrowing costs raise the utilization needed to cover a facility. Curtailment can interrupt the hours on which both assumptions depend. CoreWeave needs operating efficiency as well as occupancy.

The unresolved relationship between Meta’s $14.2 billion and $21 billion figures matters because lenders need to know whether they represent separate claims on CoreWeave’s future capacity. A rack that began as GPU collateral now arrives with a customer commitment, a supplier backstop, a power schedule and a bond yield. The cloud icon stayed weightless; underneath it, four counterparties had been bolted to the frame.