CoreWeave beat on nearly every line that Wall Street models and the stock jumped as much as 14% in extended trading. The number that deserves more scrutiny than the beat is the one CEO Michael Intrator offered almost in passing: that six-year-old Nvidia hardware is being booked through 2029 at full price, a claim that carries far more weight for the AI credit market than for CoreWeave’s quarter.
The Print
The second quarter, reported on Tuesday, was a straightforward beat on a business that is still compounding at a rate few public companies have ever sustained:
- Revenue of $2.58 billion, up 112% year over year, against a $2.56 billion consensus
- Adjusted operating income of $128 million, roughly double the $66 million analysts modeled
- An adjusted loss of $1.03 per share, narrower than the $1.20 loss expected
- Adjusted EBITDA of $1.51 billion, with margin compressing to 59% from 62%
- Revenue backlog of $104.2 billion as of June 30, up 246%, excluding more than $25 billion in commitments signed in the first weeks of the third quarter
The customer names attached to that backlog are the part that makes it credible rather than aspirational. Meta committed an additional $21 billion. Anthropic signed a multi-year agreement. Jane Street put down $6 billion. These are counterparties with the balance sheets to honor what they signed, which is not something you could say about every AI contract written in the last two years.
Why the Useful-Life Argument Was the Real Headline
Michael Burry has spent much of this year arguing that the AI boom rests on generous accounting. His specific charge is that hyperscalers depreciate Nvidia-based hardware over four to six years when the real replacement cycle is closer to two or three, and that closing that gap would take more than $176 billion out of reported earnings across 2026 through 2028. Nvidia has pushed back publicly, saying observed utilization supports the longer schedules.
Intrator walked onto CNBC and answered that charge with contract data rather than theory.
His argument has three legs: Nvidia’s hardware remains the best available, CUDA makes the chips fungible across workloads, and CoreWeave’s own software layer keeps older fleets productive. The evidence he offered is that 2020-vintage architecture has been contracted out to 2029 at what he called full freight. If customers are willingly paying undiscounted rates for nine-year-old silicon, the two-to-three-year economic life at the center of the bear case is wrong.
CNBC’s Investing Club read the exchange as a direct counter to a key bear case on the AI trade. That is right, and it is also why the claim deserves pressure rather than applause. Intrator is not a neutral party to this question. He is the most leveraged possible party to it.
The Debt Underneath the Backlog
Here is the structure, stated plainly. CoreWeave raises debt. The debt buys GPUs. The GPUs fill contracts. The contracts become backlog. The backlog is what persuades lenders to fund the next round of GPUs. Every link in that chain is priced off an assumption about how long the hardware keeps earning.
The cost of running that chain is visible in the same earnings call transcript that carried the beat. Interest expense hit $640 million for the quarter, up from $267 million a year earlier. Capital expenditure ran $9.4 billion in three months, and full-year capex guidance was raised to a range of $35 billion to $39 billion. Against that, full-year revenue is guided to $12.4 billion to $13.2 billion. CoreWeave will spend roughly three times its annual revenue on infrastructure this year, funded substantially by debt, on the expectation that the assets keep producing well past the point the skeptics assign them.
This is the same machinery we described when Nvidia turned its GPUs into collateral for a $500 billion financing program backed by Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs and KKR, and it rhymes with Morgan Stanley’s estimate of the AI debt stack building across the hyperscalers. Useful life is not an accounting footnote in this system. It is the variable that sets recovery values on hundreds of billions of dollars of paper.
What Would Actually Settle This
Intrator has the better evidence right now, and it is not close. A signed contract running 2020 hardware to 2029 at undiscounted rates is a harder fact than a modeled depreciation schedule. Burry is arguing from what should happen to chip values; Intrator is arguing from what customers have actually agreed to pay.
The catch is that both statements can be true in sequence. Older GPUs hold value while compute demand outruns supply, because a slower chip you can rent today beats a faster one that arrives in eighteen months. That condition is doing quiet work in every model on both sides of this argument, and it is a market condition, not a property of the hardware. Cheaper Chinese accelerators reaching scale, or a genuine slowdown in frontier training demand, would loosen it.
Watch the renewal rates on the oldest fleets rather than the backlog headline. Backlog measures what customers promised when compute was scarce. Renewals measure what they will pay once it is not.