Nvidia Is Reportedly Backstopping OpenAI to Buy Nvidia

Nvidia is reportedly in talks to guarantee $250 billion of financing so OpenAI can lease a 10-gigawatt data centre — a facility whose entire purpose is to buy Nvidia chips, on separate financing talks worth another $350 billion. Michael Burry, who shorted the 2008 housing bubble, summed it up in five words: 'Around and around we go.' Whether he is right or wrong, the number tells you something urgent about the ground your automation stands on.

This is the story that lit up every timeline this weekend, and it deserves the attention, because the number is genuinely difficult to hold in your head. A quarter of a trillion dollars, guaranteed by a chipmaker, to help a customer without an investment-grade credit rating buy a data centre — built on a decommissioned uranium-enrichment site — whose reason for existing is to fill up with that same chipmaker's hardware.

Let me walk through what was actually reported, why serious investors are calling it a bubble and why serious people are calling that wrong, and then the part that matters most: what a number like this tells you about the compute your business quietly depends on.

A quick word first on why this story, of all the week's news, is the one that spread. It has the two ingredients that make a financial story go viral: a number so large it stops meaning anything, and a shape so simple anyone can see it. "Half a trillion dollars" is abstract. "The chip company is lending its customer the money to buy its chips" is a diagram a child could draw. When a complex situation collapses into a shape that obvious, it travels — and the shape it collapsed into this weekend was a circle.

What Was Actually Reported

On July 26, the Wall Street Journal reported — and Reuters relayed while noting it could not immediately verify the figures — that Nvidia is in talks to provide a roughly $250 billion financial backstop to help OpenAI lease a 10-gigawatt data centre that SoftBank's energy subsidiary is developing in southern Ohio.

The scale is hard to overstate. Per Outlook Business, including the AI chips that would fill it, the project could cost more than $500 billion — the largest data-centre project announced to date. It would draw around 10 gigawatts, enough to power several million homes, built over multiple years, with a first phase of roughly 800 megawatts expected by 2028.

The details that make it a story

• The backstop exists because OpenAI lacks an investment-grade credit rating — per Benzinga, Nvidia would guarantee financing tied to the data-centre lease and construction debt, lowering lenders' worry about being repaid
• Separately, per the same reporting, Nvidia is discussing financing for the purchase of Nvidia chips that could total another $350 billion
• The site is a decommissioned uranium-enrichment plant in Piketon, Ohio, about 50 miles south of Columbus, redeveloped by SoftBank's energy unit
• The power comes from a natural-gas facility backed by a $33 billion Japanese investment under a recent US-Japan trade deal, with the power allocation effectively controlled by the US government
• Microsoft, Google and Anthropic have also reportedly expressed interest in the site, per Benzinga

Treat the figures as reported rather than confirmed — Reuters was explicit that it could not verify them, and neither Nvidia nor OpenAI had commented. But the direction is not in dispute, and the direction is the point.

A chipmaker guaranteeing a quarter-trillion in debt so a customer can buy a data centre to fill with its chips, plus another $350 billion in chip financing on the side. You do not need to know if it's a bubble to know it's a circle.

It is worth pausing on how new this posture is for Nvidia. For most of the last decade, Nvidia sold chips and let customers worry about how to pay for them. This deal is a different company: one that finds the land, guarantees the debt, and lines up the power, because selling the chips is no longer enough when the customers cannot individually afford the scale the chipmaker needs them to buy. That is a supplier underwriting its own market — and it happens when a supplier concludes the market cannot grow fast enough on its customers' own balance sheets. Whether you read that as confidence or as a warning sign is exactly the debate.

Why the Bears Are Circling

The reaction was immediate and it came from people with track records worth respecting.

Michael Burry — the investor who famously shorted the 2008 housing bubble — reacted on X with the line that became the weekend's headline. Per Benzinga, he wrote: "Around and around we go. Nvidia to guarantee $200 billion of ChatGPT's spending on $NVDA chips." The same reporting noted that Burry had increased his short position against Nvidia the previous Friday.

His objection is the one everyone can see once it is pointed out. When a supplier finances its customer's purchases of the supplier's own products, the revenue that results is not quite the same thing as revenue from an independent buyer spending its own money. Some of the demand is, in a sense, the seller's balance sheet talking to itself. If that loop unwinds, it can unwind fast, because the same entity is on both sides of it.

The structural worry is concentration. As critics quoted in the coverage put it, the arrangement is striking partly because the campus is being built by SoftBank — an entity with its own history of enormous, leveraged technology bets — and partly because the question of where all the money ultimately comes from does not have a clean answer. Layered guarantees concentrate risk rather than dispersing it.

There is also a timing argument the bears make. These structures tend to appear late in a cycle, when the easy capital has already been deployed and the remaining growth has to be manufactured through financial engineering rather than drawn from fresh, independent demand. A supplier guaranteeing a customer's debt is, on this reading, a sign that the ordinary channels of funding have been exhausted and the growth story now needs scaffolding to stand up. Whether that reading is correct depends entirely on the one thing nobody can see from the outside: how much of the demand would exist without the guarantees. If most of it would, this is prudent financing of a real boom. If much of it would not, it is the sound a cycle makes near its top.

When the seller guarantees the loan the buyer uses to buy from the seller, a chunk of the demand is the seller's own balance sheet wearing a customer's name. That is the part that keeps short-sellers up at night.

Why the Bulls Say That's the Wrong Read

Now the other side, because a popular story that only tells you the scary half is doing you a disservice, and the bull case is not stupid.

The strongest argument for the defence is simple: the demand is real. This is not fibre-optic cable laid in 1999 for internet traffic that would not exist for a decade. The data centres being financed are running at capacity the moment they come online, serving models that millions of people and businesses use every day and pay for. A circular financing structure wrapped around real, present, paid demand is a different animal from one wrapped around a hope.

And vendor financing is not inherently a scandal. As the more measured coverage noted, it is a normal feature of capital-intensive industries — aircraft manufacturers finance airlines, telecom-equipment makers finance carriers, and have for decades. A supplier helping a creditworthy-enough customer afford a long-lived asset is ordinary. The question is only whether the customer and the demand are real enough to carry it.

There is a genuine physical constraint underneath all of it that the bubble framing misses. As Finimize observed, the AI race has shifted from "who has the best chips?" to "who can fund and power the factories that run them?" — and the binding limit is increasingly electricity, not capital. When power allocation for a project is effectively controlled by the government and financed by a foreign trade deal, you are not looking at a purely speculative bubble. You are looking at an industrial build-out with a national-security dimension.

So hold both. The financing structure genuinely concentrates risk, and the underlying demand is genuinely real. Both are true at once, and anyone selling you only one half is selling you something. The honest position is that this is real demand financed in a fragile way — which is exactly the combination that is hardest to reason about, and easiest to argue about forever.

The detail that tilts me toward taking the build-out seriously rather than dismissing it as mania is the power. Speculative bubbles are made of paper — promises, valuations, financing engineering. This one is made partly of a gas plant, a grid connection, and a government allocation of electricity on federal land. You cannot inflate a gigawatt. The fact that the binding constraint has moved from money to megawatts is the strongest evidence that something physical and real is being built, even if the capital wrapper around it is precarious. Bubbles do not usually run into power-grid limits. Industrial revolutions do.

The Part That Actually Matters For You

Here is why this belongs on your desk and not just in your feed. You do not have to resolve the bubble debate. You have to notice what the debate is telling you about the ground under your own automation.

Every AI workflow you run rests on an assumption you probably have never written down: that compute will be abundant and roughly affordable indefinitely. This story is a $250 billion neon sign pointing at that assumption. The reason these structures exist at all is that compute is scarce, expensive to build, and getting more so — scarce enough that a chipmaker will guarantee a quarter-trillion in debt to bring more of it online, and that a data centre's power now depends on a government allocation and a foreign trade deal.

You have been treating cheap, abundant compute as a law of nature. This deal is the reminder that it is a construction project — financed, powered, and politically allocated, any of which can change the price you pay.

Consider how much of the modern AI business model quietly assumes the opposite. Free tiers, flat-rate subscriptions, agents that burn thousands of tokens to save a human thirty seconds — all of it is underwritten by an unspoken belief that the marginal cost of intelligence trends to zero. This deal is what that belief looks like from the supply side: not a smooth glide to zero, but a half-trillion-dollar scramble to build enough capacity to keep the price from rising. The trend toward cheaper intelligence is real, but it is being manufactured at enormous cost and risk, not delivered by nature. Business models that assumed it was free are the ones most exposed if the manufacturing falters.

This is the same physical reality I wrote about during the memory and compute crisis reshaping automation economics, now expressed at the level of national infrastructure. The cost floor under your token bill is not set by software. It is set by silicon, power, and the financing that brings both online — and all three of those are under visible strain.

Two futures, and you should be ready for either

The deal points at a fork. In one future, the build-out works: the capacity comes online, the demand keeps pace, and compute stays roughly affordable because supply finally catches up. In the other, some link in the chain of guarantees strains, financing tightens, and the price of compute jumps for everyone downstream — including you, who never signed any of these deals but pays the per-token consequences of them.

You cannot control which future arrives. You can build a business that survives both, and the way you do that is the same regardless of which one shows up: own your process rather than any specific compute source, and keep your model layer portable enough that a price shock at one provider is a config change, not a crisis.

Notice that the disciplined response is identical in both futures. If compute gets cheaper, a lean, portable, well-measured stack captures the savings without effort. If compute gets more expensive, the same stack absorbs the shock because it was never wasteful and was never locked to one provider. You do not have to predict the outcome to prepare for it — which is the whole point of building for resilience rather than for a forecast. The teams that will be hurt are the ones that quietly bet everything on the cheap-compute future by wiring themselves permanently into one provider's economics.

What This Means If You Run AI In A Business

1. Treat cheap compute as a weather forecast, not a climate

Plan your economics for the possibility that per-token costs rise, not just fall. The efficiency headlines of recent weeks are real, but so is a $250 billion backstop that only makes sense because compute is scarce. The protection is structural: use fewer tokens per unit of work through task-model matching and the advisor pattern — cheap default model, frontier escalation as the exception — so a price shock hits a lean stack rather than a wasteful one.

2. Portability is now a financial hedge, not just an engineering nicety

If one provider's economics are propped up by a fragile financing structure, you want to be able to leave without a rebuild. An abstracted model layer — where switching provider is a configuration change — turns a counterparty's balance-sheet problem into someone else's problem. The teams locked into a single provider are the ones exposed to that provider's financing risk, whether they realise it or not.

Most teams never think of provider lock-in as a financial exposure, but that is exactly what it is. When you hardcode a single model provider across your stack, you have taken an unhedged position on that provider's balance sheet — its funding, its pricing power, its survival. You would never let a single supplier become an unhedged risk in any other part of your business without at least a backup. Compute deserves the same treatment, and this deal is the reason why: the provider you depend on may be depending, in turn, on a chain of guarantees you cannot see and cannot influence.

3. The open-weight tier just became more strategically important

The same weekend this deal broke, the largest open-weight model in history went free to download. That is not a coincidence of the calendar so much as two faces of the same pressure. With a third or more of US enterprise tokens already routing to open models, a capable model you can self-host or run through competing inference providers is a hedge against exactly the concentration risk this deal embodies. You do not have to bet against the build-out. You just have to not be entirely dependent on it.

4. Rules cost nothing to run

The cheapest possible response to expensive compute is to use less of it, and the cheapest workflow is the one that never calls a model at all. Most tasks need a rule, not an agent, and a rule is immune to every twist in this story — no tokens, no provider, no financing, no exposure. Every workflow you can demote from an agent to a rule is a workflow that does not care whether this $250 billion bet pays off.

This is the point most easily lost in a week of ten-figure headlines: the frontier of your own efficiency is usually not a better model or a cheaper provider. It is the honest question of whether a given task needed AI at all. A regex, a lookup table, a deterministic transform — these do not appear in any funding round, but they are the parts of your stack most insulated from every risk this article describes. The most expensive-compute-proof architecture is the one that reaches for a model only when a model is genuinely required, and reaches for the biggest model only when the task genuinely earns it.

The Metric That Doesn't Care About the Bubble

Whether this deal is genius or madness, it does not touch the number that decides whether your AI spending returns anything: your Automation Ratio — the share of AI-assisted outputs that ship without human correction.

A $500 billion data centre does not raise your ratio. Cheaper compute, if it arrives, does not raise your ratio. More expensive compute, if that arrives instead, does not lower it. The ratio is set by how well your processes are designed and measured, and it is entirely within your control while everything in this story is entirely outside it. That is precisely why it is the number worth watching: it is the one variable in the whole AI economy that belongs to you.

There is a strange comfort in that. The headlines are about numbers with twelve digits, decided in rooms you will never enter, by people whose incentives have nothing to do with yours. It is easy to feel like a passenger. But the number that actually determines whether AI pays off for your business is small, local, and yours — measured on your own workflows, moved by your own decisions, indifferent to whether a chipmaker guarantees a customer's debt in Ohio. Spend less energy forecasting the bubble and more measuring the ratio. One of those is a spectator sport; the other is your actual job.

Eighty percent of executives still report no measurable AI ROI, and 74 percent of agent deployments still get rolled back. Not one of those failures will be fixed by a bigger data centre or cheaper chips. They are process and measurement failures, and they would persist unchanged whether compute cost half as much or twice as much.

The businesses that will be fine whichever way this bet breaks are the ones built like Bending Spoons — $2.57 million of revenue per employee, humans owning judgment, machines executing lean. That model is efficient enough to absorb a compute price rise and disciplined enough to keep returning either way. It is not exposed to the $250 billion question, because it was never betting on cheap compute in the first place.

The Bottom Line

The number is real even if the deal is not yet signed. A chipmaker is reportedly willing to guarantee a quarter-trillion dollars of debt so a customer can build the thing that buys its chips, on a former uranium site, powered by a government-allocated gas plant financed by a foreign trade deal. Whatever else it is, it is the clearest possible signal of how scarce, expensive, and strategically contested compute has become.

Michael Burry sees 2008. The bulls see aircraft financing and a national industrial build-out. They might both be partly right — real demand, financed in a fragile way, is exactly the combination that produces genuine value and genuine risk in the same structure. What you should not do is treat the argument as entertainment. It is a status report on the foundation your automation is built on.

You are not a spectator to this. Every token you spend runs on the compute this $250 billion is meant to build. You do not get to vote on the deal — but you do get to decide how exposed you are to whether it works.

So do the boring, durable things. Assume compute prices can go up as well as down. Keep your model layer portable so a provider's financing problem is not your problem. Hold an open-weight hedge. Demote every workflow you can from an agent to a rule. And measure the one number in this entire story that is actually yours to move. That is what systems-first thinking buys you: a business that does not need the bubble to inflate or the bet to pay off, because it was built to return value at any compute price.

Keep the near-term calendar in view. Kimi K3's full open weights went live today, July 27, the largest open-weight release in history. Sonnet 5 introductory pricing expires August 31, with standard pricing and the tokeniser multiplier landing September 1, and Fable 5's grace period ends September 30. And the IPO calendar looms — Anthropic's October filing and OpenAI's November target will put every one of these financing structures in front of public-market scrutiny for the first time. Boring, disciplined moves compound. A quarter-trillion-dollar circle may or may not — and you do not want your business to be the part of it that finds out.

So watch the deal the way you would watch the weather before a long trip: not with panic, and not with indifference, but with a bag packed for either forecast. The build-out may deliver the cheapest compute in history or the sharpest price shock in the sector's short life. Your job is not to guess which. Your job is to have built something that wins either way — lean, portable, measured, and reaching for expensive compute only when the task truly demands it. Do that, and the biggest number in the history of the industry becomes, for you, just another headline you read and set down.

Frequently Asked Questions

What exactly is Nvidia reportedly doing for OpenAI?

Per the Wall Street Journal via Reuters, Nvidia is in talks to provide a roughly $250 billion financial backstop — guaranteeing debt tied to a data-centre lease and construction — so OpenAI can lease a 10-gigawatt campus SoftBank's energy unit is building in Piketon, Ohio. Per Benzinga, Nvidia is separately discussing up to another $350 billion in financing for chip purchases. Reuters could not independently verify the figures, and neither company has commented, so treat them as reported rather than confirmed.

Why are people calling it circular or a bubble?

Because the supplier is financing the customer's purchase of the supplier's own products. Michael Burry, who shorted the 2008 housing crisis, wrote "around and around we go" and increased his Nvidia short. The concern is that revenue generated by a seller's own guarantees is not fully equivalent to independent demand, and that layered guarantees concentrate risk. It is a serious critique from serious investors — but it is a critique, not a consensus.

What's the counterargument?

That the demand is real and present, not speculative, which distinguishes it from historical bubbles built on future hope; and that vendor financing is normal in capital-intensive industries like aircraft and telecom. There is also a physical constraint the bubble framing misses: the binding limit is increasingly electricity, and this project's power is government-allocated and financed via a US-Japan trade deal, making it an industrial build-out with a national-security dimension rather than pure speculation.

Why is it being built on a former uranium site?

The campus sits on a decommissioned uranium-enrichment plant in Piketon, Ohio, on federally owned land. Per reporting on the SoftBank plan, the first phase is expected to deliver about 800 megawatts by early 2028 at a cost of $30–$40 billion. Such sites are attractive because they come with existing power infrastructure, grid connections, and federal involvement — exactly what a 10-gigawatt facility needs and what is hardest to permit from scratch.

Should this change how I run AI in my business?

Not by panicking — by hedging. Assume compute prices can rise, use fewer tokens through task-model matching and rule-based workflows, keep your model layer portable so a provider's financing trouble is not yours, and hold an open-weight option. The goal is a business that returns value whether compute gets cheaper or more expensive, because you cannot control which happens.

Does this mean OpenAI is in financial trouble?

Not necessarily — it means OpenAI is capital-constrained relative to its ambitions, which is different. The backstop exists specifically because OpenAI lacks an investment-grade credit rating and needs help securing favourable financing for an enormous asset. That is a statement about the mismatch between its scale of ambition and its balance sheet, not a verdict on its business. But it does mean that anyone depending entirely on OpenAI's infrastructure is, indirectly, depending on these financing structures holding — which is a reason to keep alternatives ready, not a reason to flee.

What's the single most important thing to do about this?

Make sure switching AI providers is a configuration change rather than a rebuild, and keep at least one capable open-weight option you could run independently. That single piece of architecture converts every financing wobble, price shock, and provider-specific crisis from a threat into an inconvenience. You cannot predict whether the $250 billion bet pays off, but you can make sure your business does not depend on the answer.

Related Reading

Go deeper:

1. RAMageddon: The Memory Crisis Reshaping AI Costs — the compute-scarcity reality this $250 billion deal is a response to.

2. The AI Agent Platform War — why a portable model layer is now a financial hedge, not just an engineering one.

3. You Don't Have an AI Problem, You Have a Systems Problem — owning your process instead of any specific compute source.

4. CNBC Confirms 30–46% of US Enterprise Tokens Flowing to Chinese Models — the open-weight hedge against provider concentration.

5. You Don't Need an Agent, You Need a Rule — the workflow that runs on no compute at all.

6. Bending Spoons: $2.57M Revenue Per Employee — the business that returns value at any compute price.

7. Anthropic: Profitable, $47B ARR, October IPO — the public-market scrutiny these financing structures are about to face.

About the Author

Hamza Baig is the founder of Hexona Systems, an AI automation agency operating across six continents, and the AI Automation Institute, where he has trained more than 40,000 entrepreneurs in practical AI systems design.

He has been featured in the GHL Top 50, Yahoo Finance and Brainz Magazine. His work focuses on the gap between what AI can do in a demo and what it reliably ships in production — the Automation Ratio, task-model matching, and governance-first architecture.

Read more analysis on the Hamza Automates blog, get in touch about your automation stack, or follow @hamza_automates on Instagram for daily breakdowns.

Sources: Wall Street Journal reporting as relayed by Reuters, Outlook Business, Benzinga, Finimize, AOL, DigiTimes and Breaking The News, plus Bloomberg reporting on the SoftBank Ohio project. The $250 billion backstop and $350 billion chip-financing figures are from WSJ reporting that Reuters could not independently verify and that neither Nvidia nor OpenAI has confirmed; treat them as reported rather than confirmed. Figures current as of July 27, 2026.