CoreWeave's $104B backlog: GPU capacity is pre-sold

The company renting GPUs to Meta and Anthropic just reported the shape of the AI compute market, and it is not a spot market. CoreWeave doubled revenue year over year and is sitting on a $104 billion revenue backlog — capacity that is already sold, on contracts running as far out as 2032. If you are planning AI infrastructure costs for a small business, the takeaway is simple: the good capacity is spoken for, and you are not the one it was reserved for.

What actually happened

CoreWeave reported Q2 2026 results on August 11, per CNBC's coverage:

• Revenue: $2.58 billion, up from $1.21 billion a year earlier — more than double
• Net loss: $626 million, widened from $290 million, driven largely by interest costs
• Revenue backlog: roughly $104 billion as of June 30, up 246% year over year
• More than $25 billion in additional customer commitments signed in the early weeks of Q3
• Active power capacity grew nearly 500 MW to 1.5 GW
• Technology and infrastructure expenses hit $1.51 billion, up from $670 million
• The stock rose about 11% in extended trading

The contracts behind that backlog are the interesting part. CoreWeave signed a $21 billion agreement to supply Meta through 2032, on top of a prior $14 billion commitment, plus a multi-year deal with Anthropic for Claude compute.

Why a pre-sold GPU market matters for your business

Two facts sit uncomfortably next to each other here. Demand is so strong that capacity is booked out to 2032. And the company booking it lost $626 million in a quarter, mostly on interest, while spending $1.51 billion on infrastructure. That is a business converting debt into power capacity and selling the output forward to a handful of very large buyers.

For a small operator, three consequences:

Stop modeling inference costs as a downward curve. Per-token prices have fallen for three years and everyone budgets like that continues automatically. It continues only if capacity outruns demand. A 246% backlog increase says it currently does not. Build your 2027 numbers on today's rate card, and treat any price cut as upside rather than a plan. We made the same point when TSMC posted record July revenue.

Assume you are a residual buyer. Meta and Anthropic have contracts through 2032. You have a credit card and a rate limit. When capacity tightens, the contracted customers do not feel it — the pay-as-you-go tier does, in the form of throttling, waitlists on new model tiers, and quiet regional capacity limits. Design for that: keep prompts and tooling portable, have a second provider already wired up, and know which of your workloads can run on a smaller model without anyone noticing.

Watch the interest line, not just the revenue line. Debt-financed infrastructure is the structural story under all of this. If borrowing costs move against these buildouts, the cost shows up in what you pay per token, because there is nowhere else for it to go.

None of this is a reason to slow down on AI. It is a reason to make sure the AI you deploy is not welded to one vendor's capacity.

Sources: CNBC, Yahoo Finance.