Why NVIDIA can't let the music stop
Vendor financing built the telecom boom, then buried it. Nvidia is now running the same play at seventy-five times the size.
The guarantee, when the number surfaced over the weekend, was for $250 billion. A paragraph deeper into the same Wall Street Journal reporting there was a second number, $350 billion, for the chips themselves. Nvidia, the company whose accelerators the AI industry is built on, is in talks to stand behind that much of OpenAI's spending: the lease on a ten-gigawatt data center in southern Ohio, then the GPUs to fill it. It would let OpenAI, the maker of ChatGPT and the most expensive private company Silicon Valley has produced, occupy a building it could never finance on its own name. The talks are early and could collapse, reprice, or shrink by the time you read this. The direction is the thing. The arrangement would be the largest vendor-financing commitment the technology industry has attempted, and it came together, like every important deal this year, over a weekend, with not one new dollar of revenue attached.
The market did the obvious thing on Monday. Nvidia fell more than four percent. Michael Burry, the investor who shorted the subprime mortgage in 2007 and has spent the years since looking for the next thing quietly on fire, added to his short and posted four words: "around and around we go." Ed Zitron, a critic who has made the boom's accounting his beat, called it "such an insane thing to do on so many levels," and pointed out that the building is going up by SoftBank. Everyone reached for the same two words. Circular financing. It has been the designated worry of the boom for two quarters, and it is correct, and it is not the useful part.
Here is what the phrase misses. The money does move in a loop: Nvidia invests in OpenAI (a reported $30 billion earlier this year), OpenAI buys and rents Nvidia silicon, Nvidia guarantees the debt on the building the silicon sits in. Drawing the loop proves the system is entangled. It does not tell you the one thing an allocator needs before pricing any of it, which is where the loss lands first when the demand comes in soft. The loop has a floor.
Watch the mechanism one storey down, where it runs quieter. CoreWeave, the "neocloud" that rents out GPU capacity by the hour, closed the first quarter with more than $17 billion of debt against a $740 million loss, and it borrows at investment-grade rates regardless. Its flagship facility carries an A3 from Moody's, the first investment-grade rating ever stamped on debt backed by graphics chips. The rating measures whoever signed the customer contract on the other side, a Microsoft or a Meta whose promise to pay is the collateral that counts, not CoreWeave's own credit, which is speculative. The lender underwrites the name the borrower can point to.
That is the whole game, and the Ohio guarantee is the game at full volume. OpenAI cannot raise $250 billion against a building on its own signature, because it has no investment-grade rating and is losing, by its own internal projection, roughly $14 billion this year. The honest cash burn, once you stop excluding what that figure excludes, runs closer to a reported $27 billion. So a rated balance sheet is slid underneath the debt in place of the one OpenAI lacks. Nvidia's. The lenders who said no to OpenAI directly say yes to OpenAI-with-Nvidia-behind-it, and the building gets built.
Which leaves the question the loop-drawers never reach. If the borrower's credit is not what holds up the debt, and a stronger balance sheet has been slid into its place, that balance sheet is sitting in the first-loss seat. Whose is it, and how wide is the seat?
By our reckoning it is Nvidia's, and the seat is about $600 billion wide.
This article is for Vector members. Start a 7-day free trial to keep reading.
Start your free trial