Let’s talk about a dollar that gets counted as revenue three times before it stops moving.
There’s a number worth sitting with. As of the 26th of July this year, Nvidia had ninety-nine billion dollars of equity sitting on its books. It isn’t a portfolio of stocks. It’s its own customers.
A year earlier that number was about seven billion. In January it was around forty. It grew four-and-a-half times in six months.
For scale: Nvidia’s total revenue last quarter was ninety-six billion. So the chipmaker now holds more value in the companies buying its chips than it takes in from selling to anyone in a full quarter.
Nothing about that is illegal. Nothing about it is even necessarily wrong.
But it changes what the price on your AI bill actually means.
Think about the old way this worked. General Motors had a financing arm, and GM lent people the money to buy GM cars. Legal, normal, a century old. It was fine because the car was real, the customer needed it, and the loan got repaid out of the customer’s own income over five years.
Now turn it up.
Follow one loop. It’s documented. Nvidia puts three billion dollars into SB Energy, the SoftBank unit building a data centre campus in Piketon, Ohio. SoftBank puts sixty-four point six billion into OpenAI. OpenAI commits to fill its data centres with ten gigawatts of Nvidia systems. And then Nvidia signs a guarantee backing up to a hundred and five billion dollars of OpenAI’s lease obligations at that same Ohio campus.
The supplier is underwriting the customer’s rent. The customer committed to buy the supplier’s product. The supplier’s money is inside the landlord.
Then look at Nscale, the neocloud that filed its IPO prospectus on the 18th of September. Nscale booked a hundred and forty-one million dollars of revenue in six months. Anthropic and Microsoft have signed up to pay it somewhere around eighty-eight billion. Nvidia sits on four sides of that company at once — shareholder, customer, guarantor of its Texas lease, and the supplier of every chip it runs.
And Nscale’s own prospectus says it has no binding financing to build what it has now promised to deliver. Anthropic can cancel its forty-four point six billion tranche by tranche if Nscale turns up late.
Or take Broadcom, whose latest filing shows it backstopping up to twenty-nine billion dollars of one unnamed customer’s rack leases, while potentially lending that same customer another forty-two billion so it can pay rent on racks Broadcom is buying.
You see the shape. A dollar goes out from a supplier, comes back as revenue, goes out again as investment, comes back as revenue, and gets counted each pass. Michael Burry’s number for it was eight hundred and seventy-nine billion dollars of hyperscaler commitments circling back through Nvidia. His line is the cleanest: the money is circling, not multiplying.
What people are missing is that this is not a story about lying. It’s a story about correlation — and about what your price is made of.
Everyone in that loop is real. The chips are real, the campuses are real, the demand is real. Nvidia generated close to forty-eight billion dollars of free cash flow in a recent quarter. That’s cash, not accounting. The analysts calling this overstated aren’t fools: JPMorgan’s point is that this buildout is funded from margins and free cash flow rather than from the credit markets, which genuinely separates it from 1999, and Bank of America reckons vendor financing will be five to ten percent of the roughly five trillion dollars spent on AI by 2030. Jensen Huang stood up on the 10th of September and said it plainly — it’s not circular, because we put a little bit of money in and a lot of money comes back.
Steel-man him. He’s not wrong about the direction the money flows.
But nobody pricing a business has absorbed what happens next.
The first stress test already arrived, and it wasn’t a product failure. It was the cost of money. On the 16th of September the Fed raised rates and the 10-year yield touched five percent. The Fed chair, Kevin Warsh, said out loud that hyperscalers competing for capital were part of what pushed it there. SoftBank is paying junk-bond rates — guided up to nine point eight seven five percent — to fund its last OpenAI instalment. Oracle sold twenty billion dollars of new shares, and nearly half its operating cash flow that quarter was customers paying in advance. Morgan Stanley put the off-balance-sheet purchase and lease commitments across six of these companies at one point three trillion dollars, and rating agencies have already moved on at least one of them. The Bank for International Settlements said it in its annual report: the debt tied to this buildout is growing faster than the balance sheets carrying it.
And Bain’s figure is the one to keep on the wall. The AI ecosystem needs about two trillion dollars of annual revenue by 2030 to justify the infrastructure being built. On the current trajectory it falls roughly eight hundred billion short.
So here’s the operator question.
If the price of your AI is being held down by a supplier that has bought equity in the company selling it to you, then the price you’re paying is not a market price. It’s a subsidy. A subsidy is a decision somebody else made, on a balance sheet you can’t see, for reasons that have nothing to do with your business.
Three things to do with that, and they’re boring on purpose.
One. Find out who paid for your discount. When your cost per inference drops seventy percent in eighteen months, ask what’s carrying it. If the answer is cheap capital and a supplier’s equity book, that price has a clock on it. Rebuild your unit economics at two to three times today’s number and check whether the thing still works. If it only works at the subsidised price, you don’t have a business. You have a bet on somebody else’s balance sheet.
Two. Treat your vendor’s debt as part of your supply chain. Not the roadmap — the balance sheet. When one company is shareholder, customer, lease guarantor and sole chip supplier at the same time, those four roles fail together, on the same morning. Ask for the fallback in writing, in a form you can run. A second provider you’ve never routed real traffic through is not a fallback.
Three. Own your exit while it’s cheap. Keep your state, your definitions, your prompts, your data, and the ability to recompute your own costs inside your own systems. The cheapest month you’ll ever get from a vendor is the month before the subsidy ends.
Now the part that outlives the money.
This century taught all of us to accept a price without asking what’s holding it up. Free search, free social, free storage. Never free — paid for by something you weren’t shown. The AI buildout is the same bargain at industrial scale, except the currency is borrowed money and the collateral is the customer.
A machine that’s cheap because someone financed the person selling it to you is not a cheap machine. It’s a machine running on a term sheet you were never handed.
You can’t consent to a cost you were never told about.
So ask the question every single time. Who paid for this discount, and what happens to me when they stop.
Because the answer tells you whether you’re standing on infrastructure — or inside somebody else’s loop.