I Owe My GPUs to the Company Store
Nvidia is starting to finance the AI economy that buys Nvidia. That doesn't mean the boom is fake. It does mean somebody outside the circle eventually has to pay.
There's an old song called "Sixteen Tons" about coal miners trapped in the economics of the company town. You load sixteen tons. You get another day older. You get deeper in debt. And, famously: you owe your soul to the company store.
Well, welcome to AI Economics 101. Except now you owe your GPUs to the company store.
Nvidia — currently sitting somewhere near the center of the known economic universe — is reportedly considering providing roughly $250 billion in financing guarantees to help OpenAI secure a massive 10-gigawatt data center project in southern Ohio.
Not million. Not billion with a cute little single digit in front of it. A quarter of a trillion dollars.
And before somebody comes screaming into the comments: no, Nvidia isn't simply handing Sam Altman a $250 billion check. The proposed arrangement is more sophisticated than that. Nvidia would reportedly backstop financing connected to OpenAI leasing the massive facility being developed by SoftBank subsidiary SB Energy. In effect, Nvidia would be standing behind OpenAI and telling lenders: relax. If the kid can't make rent, Dad's good for it.
That alone is extraordinary. But then it gets better. Or weirder. Depending on how much Nvidia stock you own.
Nvidia is also reportedly discussing financing as much as $350 billion worth of Nvidia GPU purchases connected with the project. And that's where my inner forensic accountant starts making that little Scooby-Doo noise.
Ruh-roh.
Because now we've got the company that sells the GPUs potentially helping finance the customer who needs the data center so the customer can fill the data center with GPUs purchased from the company financing the customer.
That doesn't automatically make anything fraudulent. It doesn't mean Nvidia's revenue is fake. It doesn't mean OpenAI doesn't need the compute. And it sure as hell doesn't mean AI is fake.
But it does mean we've reached the point in the AI infrastructure boom where you need a whiteboard to follow the money. And anytime you need a whiteboard to follow the money, I get interested.
WELCOME TO THE NVIDIA COMPANY STORE
The simplified version looks something like this: Nvidia financial strength → OpenAI → data centers → Nvidia GPUs → Nvidia revenue.
The money goes out one side of the ecosystem and some portion of it comes walking back through the front door wearing a purchase order.
That's not quite the perpetual-motion revenue machine some people are making it out to be. Nvidia can't hand somebody $100, have them hand the same $100 back, call it $100 of magical profit, and repeat until Jensen Huang owns Mars. Accounting doesn't work that way. The chips still have to exist. They still have to be manufactured. They still have costs. The financing creates assets, liabilities, and credit exposure somewhere in the system. A guarantee isn't the same thing as writing a check. And if the customer eventually can't pay, somebody eventually gets kicked squarely in the balance sheet.
Those distinctions matter.
But there's a much bigger question hiding underneath all of this. How much of AI demand is organic demand, and how much is increasingly being enabled by the companies benefiting from that demand?
That's the question. And it's a very different question than "Is AI a bubble?"
THIS ACTUALLY STARTED EARLIER
This isn't Nvidia suddenly waking up one morning and deciding it wanted to become JPMorgan with better leather jackets.
In September 2025, Nvidia and OpenAI announced an enormous strategic partnership involving at least 10 gigawatts of Nvidia systems. Nvidia said it intended to invest as much as $100 billion in OpenAI, progressively, as each gigawatt of infrastructure was deployed.
Read that again slowly. Nvidia invests in OpenAI. OpenAI builds AI infrastructure. That infrastructure contains millions of Nvidia GPUs. Nvidia sells the GPUs. OpenAI gets more compute. Nvidia gets more revenue. OpenAI potentially gets more valuable. Which makes Nvidia's investment potentially more valuable. Which strengthens everybody's ability to raise more capital. Which funds more infrastructure. Which requires more GPUs.
Around and around we go.
Again: none of this necessarily means anything nefarious is happening. There is a perfectly legitimate industrial logic here. If I'm Nvidia and I believe the world needs vastly more accelerated computing, it makes complete strategic sense for me to help remove the financing bottlenecks preventing customers from deploying accelerated computing.
In fact, Nvidia described itself in an investor presentation as providing the architecture, chips, systems, networking, data centers, software, operations and financing as an integrated AI infrastructure solution. That's remarkable. They're not just selling the shovel anymore. They're financing the mine. They're helping design the mine. They're providing the mining equipment. They're potentially investing in the mining company. And they're assuring the bank that the miner's good for the loan.
Okay. Cool. One tiny question. Who's buying the gold?
SOMEBODY OUTSIDE THE CIRCLE HAS TO PAY
This is the part I think gets lost in almost every conversation about the AI boom.
I've spent years looking at this from the infrastructure side. GPUs. Power. Cooling. Transformers. Generators. Fiber. Land. Data centers. Gigawatts. The physical scale of what's being built is difficult to comprehend unless you've actually spent time around this stuff. These aren't websites. They're industrial facilities. We're talking power infrastructure on the scale of cities.
Nvidia and OpenAI's existing 10-gigawatt arrangement alone represents millions of GPUs. Nvidia itself has said a single gigawatt buildout can require something like $50 billion to $60 billion in total spending. Now multiply that across OpenAI, Microsoft, Meta, Google, Amazon, Oracle, xAI, Anthropic, sovereign AI programs, and everyone else building AI factories like they're afraid the GPUs are going to run out Friday.
Reuters Breakingviews noted estimates putting hyperscaler investment around $875 billion this year, while cash flows at several major participants are under increasing pressure from the buildout.
This isn't an app-development boom anymore. It's an industrial mobilization. And industrial mobilizations eventually have to produce economic output.
That's the test. Not whether ChatGPT is cool. It obviously is. Not whether AI can generate productivity. It obviously can. Not whether businesses will adopt AI. They already are.
The trillion-dollar question is whether the economic value created downstream is ultimately large enough to support the staggering amount of capital being deployed upstream.
Because you can finance the GPUs. You can finance the data center. You can finance the power plant. You can finance the customer. You can finance the company financing the customer. Eventually someone who isn't sitting inside the AI capital carousel has to pull out an actual wallet.
THE DEMAND SIGNAL
This is why Nvidia financing its customers matters.
Normally, demand tells you something. If 10,000 companies show up waving their own money around screaming for Nvidia GPUs, that's a hell of a market signal. People want the product. Simple.
But imagine Nvidia increasingly does this: invests in AI company A. AI company A purchases Nvidia GPUs. Nvidia backs AI company B. AI company B leases infrastructure packed with Nvidia GPUs. Nvidia finances cloud company C. Cloud company C buys Nvidia GPUs and rents them to AI company D. AI company D raises money from investors partly because it now has enormous compute capacity. Everybody's valuation rises because everybody is growing. Everybody's growth requires more Nvidia GPUs. Nvidia sells more Nvidia GPUs.
The reported demand is still real in the accounting sense. The GPUs shipped. The invoices exist. The data centers exist. The electricity is definitely getting consumed.
But the quality of the demand signal becomes harder to read. That's the important distinction. Not fake demand. Subsidized, financed, enabled, or circularly reinforced demand. Those are very different things.
And the credit markets apparently noticed. After reports of the enormous OpenAI financing discussions emerged, Nvidia's credit-default swap pricing jumped as investors reassessed the risk of Nvidia putting its pristine balance sheet behind increasingly capital-hungry AI customers. Nvidia shares also dropped roughly 4% following news of the potential backstop.
The bond guys looked at the AI perpetual-motion machine and apparently said: hang on a fucking second.
Which is generally when I start paying attention.
WE'VE SEEN THIS MOVIE BEFORE
Technology history has a nasty habit of rhyming.
During the telecom and dot-com boom, equipment manufacturers discovered something wonderful. Their customers wanted enormous amounts of networking equipment. Unfortunately, some of those customers didn't possess the minor inconvenience known as money. No problem. The vendors helped finance them. Companies like Lucent and Nortel extended enormous amounts of vendor financing so telecommunications companies could keep building networks and, conveniently, buying Lucent and Nortel equipment.
And for a while it looked fantastic. More networks. More equipment. More revenue. More growth. More financing. More networks.
You can probably see where this is headed.
When the underlying telecom economics cracked, suddenly everybody discovered that a customer purchasing equipment with money effectively enabled by the equipment vendor isn't quite the same thing as a customer producing mountains of cash from a healthy underlying business.
The technology wasn't fake. That's important. The internet was real. Fiber optics were real. Routers were real. Broadband was real. The enormous economic transformation everybody predicted actually happened.
But investors still vaporized staggering amounts of money because they were right about the technology and wrong about the timing, capital structure, and economics.
Those are not mutually exclusive propositions. This is something people continually misunderstand about bubbles. A bubble doesn't require the underlying thing to be bullshit. Railroads changed civilization. Railroad companies still went bankrupt. The internet changed civilization. Dot-com companies still imploded. Telecommunications became the nervous system of the planet. Telecom investors still got annihilated.
You can be spectacularly right about the future and spectacularly wrong about what that future is worth today.
AI CAN CHANGE EVERYTHING AND STILL BE OVERBUILT
This is where the argument usually falls apart.
One side says: AI is transformative, therefore the investment boom is justified. The other says: the investment boom looks insane, therefore AI must be a bubble.
No. Both sides are committing the same intellectual sin. They're collapsing two separate questions into one.
AI can absolutely be one of the most important technologies developed during my lifetime. I happen to believe it is. And we can simultaneously be building too much infrastructure, too quickly, at the wrong prices, using increasingly creative financing structures based on economic assumptions that haven't yet been proven.
Both things can be true. In fact, historically, both things being true at the same time is almost normal. Transformative technologies attract capital. Capital attracts competition. Competition creates overbuilding. Overbuilding crushes margins. Crushed margins destroy weak balance sheets. The infrastructure survives. The technology keeps advancing. The survivors eventually inherit an enormous installed base built partially with somebody else's destroyed capital.
Welcome to capitalism. It's messy.
TOKENS AREN'T MONEY
And here's where this gets especially interesting for AI.
Everyone loves measuring usage. Tokens. Queries. Users. API calls. Model size. Compute deployed. Gigawatts commissioned. GPU shipments. Cool.
But none of those things are dollars of sustainable economic value.
The AI economy ultimately needs end users and businesses producing enough additional economic output that they're willing to transfer substantial amounts of their own money into the system.
An insurance company reduces claims-processing expense by $500 million. Great. A pharmaceutical company discovers a drug three years faster. Fantastic. A manufacturer improves yield by 8%. There's your money. A law firm increases attorney productivity by 30%. Money. A logistics company cuts fuel consumption. Money. A programmer using an AI coding agent produces twice as much software. Potentially money.
That's the part of this ecosystem that matters most. Because those dollars originate outside the AI financing loop. That's economic demand. That's the gold coming out of the mine.
If those productivity gains eventually become enormous — and I think there's a very good chance they will — then today's infrastructure buildout may look prescient. Maybe even cheap.
But if the revenue produced by AI applications can't eventually support the depreciation, energy, financing, land, networking, cooling, and replacement cycles of the machines underneath them? Then we've got ourselves one hell of a math problem.
Especially because GPUs aren't bridges. We're potentially financing enormous infrastructure over long periods using computing equipment with brutally fast technological obsolescence. Today's $50 billion AI factory isn't filled with assets expected to remain technologically dominant for 30 years. Nvidia is already working on the next architecture. Then the next. Then the next.
The machine that financed your loan may be obsolete before you've finished paying for the building it's sitting in. That's a very different infrastructure model.
JENSEN'S GENERAL STORE
None of this means I'd bet against Nvidia. God, no. If anything, Nvidia may have built one of the most strategically extraordinary positions I've ever seen in technology. They own the shovel. They own the shovel ecosystem. They've got CUDA. They've got networking. They've got systems. They've got software. They're increasingly involved in architecture. And now they're discovering that one of the biggest constraints preventing humanity from buying more shovels is financing.
So apparently they're getting into financing. That's what vertically integrated domination looks like when it reaches the capital layer.
At some point, I half expect Jensen to offer checking accounts. Walk into the Nvidia General Store:
"Morning. Need some Blackwells?"
"Can't afford them."
"That's okay. We finance."
"I also need somewhere to put them."
"We finance that too."
"Need about a gigawatt of electricity."
"Talk to the gentleman at counter three."
"What do I owe you?"
"Everything."
Sixteen tons of tensor cores. Another day older. Deeper in debt. Saint Peter don't you call me, 'cause I can't go. I owe my GPUs to the company store.
Funny line. Very funny line. But underneath the joke is probably one of the most important questions in the entire AI economy right now.
We've proven that investors will finance AI companies. We've proven that banks will finance AI infrastructure. We've proven that governments will finance power. We've proven that hyperscalers will spend hundreds of billions. We've proven Nvidia can sell damn near every accelerator it can manufacture.
Now comes the part that matters. Can AI produce enough real economic value outside its own ecosystem to pay everybody back?
I think it can. But at these numbers? With these capital structures? On these timelines?
That's no longer a technology question. That's economics.
And welcome to class.
Rich Washburn is a technologist and strategist working at the intersection of AI, infrastructure, and capital. He is Managing Partner and Chief AI Officer at Eliakim Capital.
Comments