OpenAI Nvidia Ohio Data Center Deal: Real Demand?

The OpenAI Nvidia Ohio data center deal now being discussed is being pitched as the biggest computing project ever built: a 10-gigawatt, roughly $500 billion campus about 50 miles south of Columbus, developed by SBEnergy, SoftBank’s energy arm. That’s the equivalent of a small national grid poured into server racks. But buried in the same reporting is a detail that matters more than the megawatts: Nvidia has discussed fronting up to $250 billion of that total itself, so OpenAI can afford to lease the site.

I’ve spent years around systems that measure everything in petabytes and megawatts — first tracing particle collisions at CERN, later watching hyperscale data-center builds get financed — and here’s my blunt read: this isn’t primarily a power-and-cooling story, it’s a circular-financing story. When the chip vendor bankrolls its own customer’s rent for a building full of that vendor’s chips, the resulting $500 billion headline stops being a clean measurement of AI demand and starts looking like a mirror Nvidia built for itself. Demand for GPUs is real. How much of it is now self-generated is the question nobody in this deal seems eager to answer.

The OpenAI Nvidia Ohio Data Center Deal, By the Numbers

Strip away the financing structure and the physical numbers alone are startling. The proposed campus would run at roughly 10 gigawatts — enough continuous power to rival a mid-sized country’s grid — making it, in Forbes’ phrasing, the largest data center project in the world “by far” if it proceeds. It would sit in southern Ohio, built by SBEnergy, the SoftBank subsidiary that has quietly become one of the industry’s biggest land-and-power brokers. Neither OpenAI nor Nvidia has confirmed the talks publicly, per Forbes’ July 27 report.

Ohio wouldn’t even be the only quarter-trillion-dollar campus on the map. Meta’s Louisiana “Hyperion” site was just expanded to 5 gigawatts, with disclosed investment surpassing $50 billion and a total project cost now reported above $250 billion, according to Bloomberg. Two single-company campuses are each individually approaching or exceeding a quarter of a trillion dollars — a scale my earlier look at the trillion-dollar data center boom flagged as straining the grid alone, before financing questions even entered the picture.

How Circular AI Financing Turns Vendor Into Bank

Here’s the mechanism, stripped to its bones. Nvidia sells GPUs — the specialized chips that do the math behind AI models — to whoever builds a data center. In this deal, Nvidia would also hand OpenAI up to $250 billion to lease that same data center, which will be filled with Nvidia GPUs. That’s not really an investment in OpenAI’s future; it’s vendor financing, the decades-old trick of a supplier lending its customer the money to buy the supplier’s own product.

This is not a one-off. Analysts have already flagged the same circular AI financing pattern in Nvidia’s investment in CoreWeave, and in Microsoft’s compute-credit arrangements with OpenAI. Each time, cash that looks like fresh capital is actually round-tripping between a small handful of counterparties — chipmaker, cloud landlord, model developer — who report each other’s spending as their own revenue growth. Stack enough of these deals together and the aggregate figures start measuring the enthusiasm of three or four balance sheets more than demand from the rest of the economy.

What 10 Gigawatts Actually Buys

A gigawatt is a billion watts — roughly the output of a full-sized nuclear reactor. Ten of them, dedicated to one campus, is a genuinely new category of infrastructure, and I’ve argued before that the grid, not GPU supply, is the industry’s real bottleneck. Building the substations, transmission lines, and increasingly dedicated generation to feed a single site takes years regardless of how fast the financing closes. That mismatch between capital speed and physical build speed is exactly where circular financing does its most damage: it can inflate a demand signal instantly, while the actual electrons take a decade to show up.

It’s also worth asking whether ten gigawatts of GPU clusters is even the optimal way to buy this much compute. Rivals are already betting that wafer-scale chips could make sprawling GPU cluster campuses look like the mainframe era — expensive, power-hungry, and one architecture shift from obsolete. If that bet pays off even partially, a chunk of this 10 gigawatt data center could be over-built for the workloads it eventually runs. Nobody financing a $500 billion campus today can fully price in a chip architecture that doesn’t exist yet.

Zoom out and the scale gets stranger still. There are currently 11,826 data centers worldwide, with 4,467 of them in the United States, per the July 2026 industry count. A single campus discussed in a boardroom this month would, on power capacity alone, outmuscle a meaningful share of that entire installed base — built in one location, under one financing arrangement, for essentially one customer.

The Bubble Question the OpenAI Nvidia Ohio Data Center Deal Raises

Here’s the part that should unsettle anyone tracking the so-called AI infrastructure bubble: the bigger these headline numbers get, the less they may tell you about organic demand. Every dollar of Nvidia’s own money that flows into OpenAI’s lease is a dollar that will eventually come back to Nvidia as GPU revenue, reported as proof the market is growing. That’s not fraud — it’s disclosed, if opaque, financial engineering — but it means capex totals alone can no longer answer whether this boom is real.

Multiply this structure across CoreWeave, Microsoft, and now a potential Ohio megasite, and you get an industry where the same few hundred billion dollars appears to be counted several times over, in several companies’ growth charts. I don’t think that makes AI compute demand fake. I think it makes today’s $500 billion and trillion-dollar infrastructure estimates far less trustworthy as evidence of it.

What Happens Next

If the OpenAI Nvidia Ohio data center deal closes anywhere near the terms being discussed, expect every major AI lab to demand a similar arrangement from its chip suppliers — once one hyperscaler gets vendor financing at this scale, competitors can’t afford not to ask. I’d also expect regulators and credit-rating agencies to start scrutinizing these circular structures the way they eventually scrutinized vendor financing in the telecom bubble of the early 2000s. The GPUs are real, the power demand is real, and the Ohio site will likely get built regardless of how it’s financed.

What won’t be real, at least not in the way headlines suggest, is the idea that $500 billion of committed capital equals $500 billion of independent market demand. My bet: within two years, at least one of these circular deals unwinds publicly, and the industry has to start reporting compute demand the way particle physicists report a signal — net of the noise we already know is there.