- Nvidia may backstop $250B for OpenAI’s 10 GW data-center build.
- Critics see circular financing; OpenAI sees scarce compute.
- The real bet: capacity on time beats avoiding overbuild risk.
The reported financing structure around OpenAI’s next mega-data-center project is almost designed to make investors nervous.
According to a recent Reuters report, Nvidia is in talks to provide roughly $250 billion in financing guarantees to help OpenAI lease a 10-gigawatt data-center project in southern Ohio. The project itself could cost more than $500 billion, and Nvidia is reportedly also discussing financing as much as $350 billion in chip purchases.
That is a lot of numbers, and all of them have commas.
The bearish case writes itself: Nvidia may be helping finance a customer’s ability to buy Nvidia hardware. The arrangement could create contingent liabilities, blur the line between organic demand and vendor-supported demand, and make the AI boom look uncomfortably dependent on ever-more-elaborate financial engineering.
Michael Burry has spent months making versions of that argument. He has called the AI industry’s financing plumbing “fugazi,” questioning how enormous GPU commitments move through a maze of cloud providers, special-purpose vehicles, private-credit firms, and leases. The reported Ohio structure will not make him feel better.
But there is a more bullish read, as well.
OpenAI is not primarily making a bet on whether people will want more AI. That's as close to a sure thing as exists right now. The ChatGPT making is betting the companies with compute available at the right moment will own the next stage of the market.
And that bet may be far more rational than it looks.
The expensive part is being late
The instinctive response to a 10 GW announcement is to ask whether OpenAI could possibly need that much capacity.
It is a fair question. It is also, in my view, the less important one.
OpenAI already reported more than 700 million weekly active users in its September 2025 Nvidia partnership announcement, with unverified reports north of 1 billion in 2026. Its products are no longer a science experiment looking for an audience. ChatGPT is a consumer product, an enterprise software layer, a developer platform, and increasingly a distribution channel for agents that consume substantially more compute than a conventional chatbot session.
Demand may fluctuate. Revenue, margins, and model efficiency will all matter. But the basic idea that the world will want more capable, more reliable AI is the part of this story I worry about least.
The real risk is that OpenAI gets the timing wrong.
As I wrote in my analysis of Anthropic’s compute gap, the most expensive compute in AI may be the compute you realize you need two years too late.
When a model company runs short of capacity, it does not merely suffer an accounting inconvenience. It has to throttle users, impose rate limits, slow down launches, defer enterprise workloads, or pay a huge premium for capacity that is available right now. In a market moving this fast, that is not a temporary operational headache. It is a competitive event.
Developers remember whose API was available. Enterprises remember whose systems stayed reliable at peak usage. Users build habits around the products that can actually serve them.
Compute availability is a product feature.
Nvidia is not giving OpenAI a handout
The proposed structure deserves scrutiny, but it is not hard to see Nvidia’s incentive.
A financing guarantee could lower OpenAI’s cost of capital and make an enormous project easier to fund. In exchange, Nvidia gains something exceptionally valuable: a clearer path to years of demand for its systems, networking, software, and services.
That is different from blindly propping up a weak customer.
Nvidia is positioning itself less like a component supplier and more like an industrial partner. It sells the accelerators, helps shape the data-center architecture, supports the financing structure, and locks in a customer whose infrastructure requirements can justify an entire generation of hardware.
The company’s prior agreement with OpenAI already contemplated at least 10 GW of Nvidia systems, with Nvidia intending to invest up to $100 billion progressively as capacity came online. The latest reported discussions suggest the relationship is evolving from strategic partnership into something closer to infrastructure co-development.
That can still be risky. But strategic financing is not automatically fake demand just because it helps unlock the purchase.
Airlines finance planes. Utilities finance power plants. Telecom companies financed networks long before all of the traffic existed. The question is not whether the financing is complicated. At this scale, it will be. The question is whether the asset being financed will be useful when it arrives.
For OpenAI, the answer is likely yes.
The 10 GW number is not tomorrow’s compute
There is an important caveat hiding beneath the headlines: the reported 10 GW is an endpoint, not a switch that flips next quarter.
Reuters reported that the first phase could be finished in 2028 with around 800 MW of power. The full project involves power generation, land, transmission, construction, financing, cooling, accelerators, networking, and the small matter of putting all those things together without blowing through the budget or schedule.
That means this is a huge execution bet.
It also explains why OpenAI would want to make the commitment now. A company cannot decide in 2028 that it needs 10 GW in 2028. By then, the critical decisions, like site selection, power contracts, financing, hardware roadmaps, construction, and political approvals, were supposed to have been made years earlier.
The AI race is increasingly a race to secure future delivery dates.
The bear case is real, but incomplete
Burry’s critique matters because circular structures can hide fragility. If the ultimate end-customer demand disappoints, vendor guarantees and financing vehicles do not make the cash flows appear by magic. Nvidia could be left with more credit exposure than a conventional hardware supplier. OpenAI could be locked into a capital structure built for a demand curve that never arrives.
There are also genuine reasons to be cautious:
- The reported arrangement is still under discussion, not a finalized transaction.
- A 10 GW project is exposed to construction, energy, regulatory, and supply-chain delays.
- Nvidia’s incentives are intertwined with OpenAI’s ability to finance Nvidia hardware.
- Model efficiency improvements could change how much power is needed per unit of useful intelligence.
But “model efficiency will improve” is not a complete bear thesis. Efficiency usually lowers the price of intelligence, which tends to expand usage. The history of computing is full of technologies that became cheaper per unit and more widely used overall. As computing, cloud servers, storage, and bandwidth became cheaper, they did not shrink the market, they put PCs on desks, apps in the cloud, photos in people’s pockets, and video on every screen.
AI may be heading in the same direction: cheaper inference, more agents, more workloads, more users, and a far larger amount of total compute consumed.
If that is the future, underbuilding is not prudent. It is expensive.
OpenAI is buying optionality
The best case for OpenAI’s 10 GW push is not that every megawatt will be needed immediately. It is that capacity certainty gives the company options its competitors may not have.
It can launch a more compute-intensive model without wondering whether usage caps will become the headline. It can support agents that run for minutes or hours instead of seconds. It can negotiate with cloud partners from a stronger position. It can serve enterprise customers that want dedicated capacity, predictable performance, and long-term commitments.
Most importantly, it can build around an operating base rather than shop for emergency capacity after demand has already arrived.
That is the lesson from the current AI infrastructure scramble. The companies that look reckless during the buildout may look merely early once the capacity becomes the constraint.
Nvidia’s reported $250 billion backstop could become an exhibit in a future cautionary tale about AI finance. The risk is real, and investors should not confuse a financing guarantee with cash revenue.
But the competitive logic is clear.
OpenAI is trying to avoid being the company that discovers, too late, that it needed the compute two years ago. So far, that strategy from recent years continues to pay off.
