How the AI Boom Is Really Financed: Nvidia Backs $250B So OpenAI Can Buy Nvidia’s Chips.

Behind every model release, every benchmark, every product launch gafam.ai has covered this summer sits a financial machine that is rarely examined directly. This week, a Wall Street Journal report pulled back the curtain on how that machine actually works — and the structure it revealed is a circle so large it should make everyone, in America and Europe alike, look twice. gafam.ai reads the money.

What Was Reported

On the evening of July 26, the Wall Street Journal reported, with Bloomberg and Reuters confirming, that Nvidia is in talks to provide roughly $250 billion in financing guarantees to help OpenAI lease a 10-gigawatt data center campus in southern Ohio. The site sits on the grounds of a decommissioned uranium-enrichment complex near Piketon, about 50 miles south of Columbus, and is being developed by SB Energy, the energy subsidiary of Japan's SoftBank.

The scale is difficult to hold in the mind. The full project could cost more than $500 billion once the computing hardware is included, making it the largest data center project ever announced. And Nvidia's involvement does not stop at the $250 billion guarantee, which covers only the lease and construction debt. In a parallel negotiation, Nvidia is discussing financing OpenAI's purchase of the chips themselves — a figure the Journal put at up to $350 billion.

These talks are early-stage and unfinalised; the terms could change or the deal could collapse, and Nvidia, OpenAI and the Commerce Department declined to comment. But the structure being discussed is the story, whether or not this specific deal closes.

The Circle

Read the arrangement slowly and its shape becomes unmistakable. Nvidia sells AI chips. OpenAI needs those chips but, having never turned a profit, cannot obtain an investment-grade credit rating and therefore cannot borrow at reasonable rates on its own. So Nvidia steps in to guarantee the financing — allowing the project's developer to raise debt against Nvidia's balance sheet rather than OpenAI's — and separately offers to finance OpenAI's purchase of Nvidia's own chips. Nvidia has already invested $30 billion directly in OpenAI.

The chip vendor is underwriting its customer's ability to buy the vendor's product. Nvidia's willingness to backstop demand before the infrastructure exists secures a multi-decade stream of its own hardware sales. Bank of America analysts have flagged this expanded role — Nvidia as financier, matchmaker and equipment vendor simultaneously — as a multi-year revenue driver. Others see something more fragile.

The Two Readings — Both Serious

gafam.ai will not pretend the interpretation is settled, because two credible readings genuinely compete.

The constructive reading: this is how transformational infrastructure has always been built. The interstate highway system, transoceanic cables, the early cloud — all required financing arrangements that backstopped demand before the physical asset existed. AI compute is following the same pattern, only faster and larger. On this view, Nvidia is doing the rational thing a dominant supplier does when demand for its product is constrained only by its customers' ability to finance the buildout: it removes the financing constraint. The infrastructure gets built, the AI economy scales, and the guarantees are never called because the demand is real.

The skeptical reading, voiced by commentators including Ed Zitron and echoed at Tom's Hardware, is blunter: OpenAI cannot get a bank loan to lease the data centers it needs to produce a product it currently sells at a loss, so the company that profits from OpenAI's expansion is manufacturing the financing to keep the expansion going. On this view, the circularity is the warning sign — vendor-financed demand can inflate an infrastructure bubble that looks like organic growth until the underlying revenue fails to materialise. One sobering data point: Nvidia's own most recent quarterly filing capped its maximum gross exposure across all partner facility lease guarantees at $3.5 billion. A $250 billion guarantee would be a scarcely imaginable escalation of that commitment.

Both readings cannot be right, and gafam.ai does not know which is. But the honest observation is that the entire edifice rests on an assumption — that AI revenue will grow into the half-trillion-dollar infrastructure being built for it — and that assumption is a forecast, not a fact.

The Detail That Extends a Pattern We Have Tracked

One element deserves particular attention from readers who have followed gafam.ai's coverage of US control over AI. The electricity powering this campus is not a private matter. The power is controlled by the US government and funded by Japan — some $33.3 billion committed by Tokyo under its recent trade agreement with Washington — and Commerce Secretary Howard Lutnick, the same official at the center of the Anthropic export saga and the Annex A access lists, controls the allocation of the site's power. Anthropic, Microsoft and Google have all reportedly spoken to Lutnick about access to it.

Consider what that means. The access-gating we documented at the model layer — Washington deciding which entities may use frontier AI — now extends down to the physical foundation: electricity. The US government is allocating the raw power on which the AI buildout depends, entity by entity. Even the kilowatts are becoming a strategically administered resource, distributed by the same official, on federal land, funded by an allied government's trade concession.

The European Perspective