title: “Nvidia Just Financed Its Biggest Customer for $105 Billion. The AI Bubble Debate Is Now Officially Loud.” date: 2026-08-18 tags: [ai, infrastructure, nvidia, openai, bubble, weekly-briefing] —

Nvidia Just Financed Its Biggest Customer for $105 Billion. The AI Bubble Debate Is Now Officially Loud.

Yesterday (Monday August 17) Nvidia, OpenAI and SoftBank’s SB Energy formally announced what is, by a wide margin, the largest AI infrastructure deal ever put on paper. The numbers are big enough that I had to re-read them twice.

Nvidia has agreed to provide up to $105 billion in “residual value guarantees” to SB Energy, the SoftBank-backed developer that will build, own and operate the new PORTS-Pike Technology Campus in Pike County, Ohio. OpenAI has signed a 20-year lease on the site. The campus is being built on the footprint of the former Portsmouth Gaseous Diffusion Plant — a Cold War–era uranium enrichment facility — and will ultimately scale to 8 gigawatts of AI compute (initial phase 4.25 GW), with first capacity online in 2028 and Phase 2 finished by 2032.

To put 8 GW in human terms: that’s roughly enough electricity to power 6 million US homes at any given moment, dedicated to one site, running one company’s models.

Nvidia is also putting $1.5 billion of equity into SB Energy on top of the guarantee. SoftBank is investing another $4.2 billion in regional grid infrastructure and will build at least 10 GW of new power generation (mostly gas, per the filings). OpenAI is putting $40 million into a community benefits fund for Appalachian Ohio. And the chips going into the site are Nvidia exclusively — roughly 1.5 million GPUs by the time Phase 2 ships.

Total project cost is being reported at $500 billion+. Add this to Nvidia’s prior $30 billion investment in OpenAI and the company is now committed to over $600 billion in compute flowing to OpenAI by 2030.

Why this deal is structurally different

This is not a normal vendor-customer relationship. In the old world, Nvidia sold chips. OpenAI bought them. Both reported revenue on their own balance sheet and the discipline of the market decided whether the demand was real.

The new arrangement is different. Nvidia is now effectively financing its own demand. The structure, as Reuters describes it, works like this:

1. SB Energy builds the data center.
2. OpenAI signs a 20-year lease and pays rent.
3. Nvidia guarantees the residual value of the site — so if OpenAI defaults, Nvidia covers the gap between what SB Energy can recover and the minimum guaranteed value.
4. In exchange, Nvidia becomes the exclusive chip supplier for the entire campus.

This is a clever, defensible structure. It is also, depending on your priors, either the most efficient way to scale AI infrastructure in human history, or a closed loop where the same handful of companies recycle capital between each other and call it growth.

Nvidia preemptively pushed back on the “circular financing” framing in their own blog post: “OpenAI will pay the lease.” That sentence is doing a lot of work.

What I think this actually means

Three things stand out to me.

First, the bottleneck is energy, not chips. The headline number is the $105 billion guarantee, but the binding constraint on the project is power. That’s why SoftBank is also building 10 GW of generation and pouring $4.2 billion into the Ohio grid. The fact that the largest AI infrastructure deal in history has a power-utility attachment that is itself one of the largest US grid investments of the decade tells you where the puck is going. The AI race is now, in practice, an electricity race.

Second, vertical integration is winning. Five years ago the AI stack was a marketplace. Today it’s collapsing into three or four integrated empires. Nvidia owns the chips, the networking, the systems software (DSX), the supply chain, and increasingly the financing. OpenAI owns the model and the distribution. SoftBank owns the energy and the balance sheet. Each one of them is now harder to displace than any of them was 12 months ago. The competitors who are still trying to be one of these things — a model lab, a chip designer, a power developer — are going to feel the squeeze.

Third, the “is this a bubble?” question is no longer rhetorical. A $105 billion guarantee on a 20-year lease is the kind of bet that only works if AI demand really does compound for two more decades. If it does, this is the most consequential infrastructure investment since the railroad. If it doesn’t, several of the largest balance sheets in tech are correlated on the same assumption. The Bubble camp points to: nine tech companies with roughly $3 trillion in off-balance-sheet AI commitments, an OpenAI that is now contractually locked into a 20-year lease, and a chipmaker that is functionally underwriting its own customers. The No-Bubble camp points to: real revenue, real users, real power bills being paid today. Both are looking at the same data.

I don’t know who’s right. But I know that the people who get to decide are no longer academics or commentators. They are the boards of Nvidia, OpenAI and SoftBank, and they have already decided.

The image that stays with me

What I keep coming back to is the site. A Cold War uranium enrichment plant in Appalachian Ohio — built to fuel nuclear weapons during a previous existential contest — is being torn up and replaced with a 20-year lease to train large language models. The same county. The same grid. The same labor force, retrained. The same ambition at industrial scale, redirected.

Whether you think the AI buildout is rational or reckless, the sheer willingness of the largest companies on Earth to plumb a gigawatt of compute into one of the poorest counties in Ohio is itself the story. The bet isn’t just on AI. It’s on AI happening here, in physical places, with real steel and concrete and union labor. For a moment, the abstract got very concrete.

I’ll be watching whether Phase 1 actually breaks ground on schedule, and whether the second gigawatt-scale OpenAI site (the 3.2 GW Georgia buildout) closes on similar terms. The shape of the next decade of AI is being set in those filings.