Nvidia's $250 billion OpenAI backstop lands as a warning, not a win

Nvidia fell roughly 4.5% in early trading on news that the chipmaker is weighing a ~$250 billion backstop for an OpenAI-led data-center buildout. On a day a deal of that size should have looked like a vote of confidence in the AI buildout, the tape priced it as a warning. That verdict is the news.

The terms, as Kobeissi Letter first reported, point to a 10-gigawatt Ohio project being developed by SoftBank at a total cost north of $500 billion, with power allocation run through US Commerce Secretary Howard Lutnick and a separate funding line from Japan under a recent trade deal. A backstop, in plain terms, is a financial guarantee: Nvidia would be on the hook for roughly two-thirds of the chip cost if the project stumbles. The market is treating that exposure as a red flag rather than a moat.

The pattern this deal sits inside is what skeptics have taken to calling circular financing. The simplest translation: the same dollars keep cycling through the AI supply chain, with chipmakers funding model labs, model labs committing to buy chips, hyperscalers signing capacity offtake, and creditors absorbing the gap through off-balance-sheet vehicles. When that loop is expanding, the market rewards every leg of it. When the loop starts to look overextended, each new commitment reads as evidence that the previous one is shakier than advertised, and the discounts start compounding.

Gary Marcus, the AI researcher whose Substack post surfaces the deal terms, frames the loop bluntly: "The GenAI boom hasn't been sustained by profits; it has been sustained by hope and circular financing. That may no longer be enough." Marcus is openly skeptical of the AI boom, and his framing is commentary, not neutral reporting. The facts, though, are the facts: short-seller Jim Chanos, whose job is to find cracks, points to Nvidia guaranteeing roughly two-thirds of the chip cost to a single data-center project. Antitrust writer Matt Stoller cuts to the demand question: "I don't understand why Nvidia is backstopping data centers unless there's just not enough real demand for compute."

In September 2025, Oracle announced a roughly $300 billion OpenAI agreement, and the stock briefly popped about 43% to ~$307 before climbing to roughly $328 within ten days. As of Marcus's writing, ORCL trades around $120, well below the deal-day peak. The market is not repricing announcements of AI demand. It is repricing the financing behind them.

That repricing has a second, less visible leg. The same Substack cites an analysis aggregated by HedgieMarkets from a Nikkei investigation into off-balance-sheet AI capex debt. The analysis, sourced through secondary reporting, finds that Alphabet, Microsoft, Amazon, Meta, and Oracle together carry roughly $1.65 trillion in off-balance-sheet AI-related debt against about $1.35 trillion reported on their balance sheets. Meta's gap, on that read, appears to be in the range of $420 billion, roughly three times its reported figure. Oracle's off-balance-sheet load appears to have grown substantially, roughly 30x in four years. Marcus flags a related Axios thread describing creditor wariness toward "creative financing" structures. Whether or not the exact Nikkei numbers survive a primary-source check, the directional point is hard to dismiss: the AI buildout is being funded substantially outside the reported balance sheet, and the credit market is starting to notice.

SpaceX's decline, down more than 25% over the past month and more than 50% from its June high, is scene-setting, not the story. It belongs here only as a reminder that the AI-adjacent capital cycle is no longer in one-direction mode and that the private mark is starting to catch up to the public tape.

The strongest counterargument is also the simplest: this is just deal-day noise, the deal is fundamentally bullish, and the backstop reflects Nvidia's confidence in OpenAI's demand curve. The falsifier is concrete. If Nvidia closes up on the week and creditors keep financing off-balance-sheet AI capex at the pace of the past two years, the read here is wrong. If instead the discount on the second leg of the deal widens and a second hyperscaler follows Oracle from deal-day peak to deal-day hangover, the read here is just early.

What changes, regardless of which side the tape ultimately lands on, is the financing model. Announced gigawatts are no longer enough. The market now wants to see funded gigawatts, the credit behind them, and the identity of the entity left holding the bag if demand disappoints. Until those three answers are clean, every new "backstop" deal of this size will trade like the one Nvidia tried to print on Tuesday.