The balance sheet is the product

Nvidia has put $2 billion into CoreWeave and $2 billion into Nebius. Both companies turn around and spend tens of billions buying Nvidia GPUs. Nvidia has also agreed to buy back any capacity CoreWeave can't sell, through 2032. Beth Kindig's breakdown of this loop, which made the rounds on Hacker News this weekend, names it plainly: circular financing. She's right. But the circle is the least interesting part of the structure.

The interesting part is what the neoclouds actually sell. The pitch is speed — CoreWeave stands up the newest chips in as little as two weeks from receipt and squeezes more useful work out of each GPU than the hyperscalers manage. Take that at face value; it's probably true. It's also not why Microsoft and Meta have committed up to $122 billion to two companies whose combined revenue this year will be about $16 billion. The load-bearing feature is accounting. A neocloud contract converts capex into opex — the spending leaves the balance sheet and comes back as a service fee, spread over the life of the deal.

The product isn't compute. It's distance between the spending and the financial statements.

The numbers make the incentive legible. Meta expects roughly $136 billion in operating cash flow this year against capex guidance of $125 to $145 billion — building that neocloud capacity in-house would tip it free-cash-flow negative. Instead, up to $62 billion in commitments lands as operating expense through 2032, maybe $10 billion a year. Microsoft runs the same play at larger scale: $190 billion of capex against $200 billion of operating cash flow, with $60 billion parked at neoclouds. The capex doesn't disappear. It reappears at CoreWeave, which spent $7.7 billion last quarter against $2.08 billion of revenue and covered the gap with debt — $24.9 billion of it now, with interest payments already consuming 26 cents of every revenue dollar.

That debt should be expensive. Mostly it isn't, and the reason is the loop again. CoreWeave's $8.5 billion GPU-backed facility carries an investment-grade rating — the first of its kind to get one — not because CoreWeave is investment grade but because the loan is secured by a long-term contract with a customer that is, plus the resale value of the chips. The rating belongs to everyone in the loop except the borrower. And beneath the customer contracts sits Nvidia's own guarantee:

None of this is a scandal, and vendor financing isn't a bubble tell by itself — it built the railroads and most of enterprise IT. If AI demand compounds the way every participant is betting, the loop closes cleanly: backlog converts to revenue, operating cash flow converges with capex, the debt amortizes against real earnings, and Nvidia's equity stakes look like cheap customer acquisition. Each position, taken alone, is rational. The problem is the direction the structure sorts risk. Nvidia collects cash for every chip it ships, whatever happens next. Microsoft and Meta hold options — contracts they can decline to renew when they expire in 2031 and 2032. CoreWeave holds everything else: $33 billion of capex this year, collateral that loses value faster than the loans amortize, and an interest bill growing faster than revenue.

Run the tape to 2031. The contracts come up for renewal, and the GPUs securing the debt will be four generations old. If demand held, everyone wins. If it merely paused — not collapsed, paused — the hyperscalers walk, Nvidia writes down an investment, and CoreWeave meets its lenders. The circle isn't hiding that asymmetry. It's delivering it, one rational deal at a time, to the balance sheet least able to hold it.