Blog

Funny Money?

Sept 9, 2026

Calling out what many are getting slightly wrong

You've seen the circular-deal charts by now. Nvidia invests in OpenAI, OpenAI buys Nvidia chips, Nvidia books the revenue, everybody's numbers go up, and somewhere a diagram with arrows in a circle floats across channels where this is reported.

The usual take is that this is fake money. Round-tripping. Nobody's actually paying anybody.

That's the part I had wrong in my initial assessment. It took a bit of research to get to a more accurate read.

The money is real. I know, that's a bit of a shocker. But here's how that works, it just goes in one direction and comes back as a purchase order. Nvidia agreed to guarantee up to $105 billion on a twenty-year lease for an OpenAI data center campus in Pike County, Ohio. Real dollars, real concrete, real obligations that will outlast most marriages, statistically speaking these days. Sorry, but true. Separately Nvidia has been exploring financing to help OpenAI buy something in the neighborhood of $350 billion of its own chips.

So what does this actually mean? In simple terms, the marginal buyer of compute is being financed by the seller of compute. When the seller stops, the buying stops, and the concrete is still standing there in Pike County.

Here's my favorite detail, and it takes two sources to extrapolate because the story changed underneath it. That Ohio guarantee was reported at $250 billion on July 27. It got signed at $105 billion on August 17.

A $145 billion haircut in three weeks. That's a bubble negotiating with itself.

At some point the story pivots away from AI

As I followed the money trail, eventually I got to a particular loan.

Roughly half of the $2.9 trillion data center buildout projected through 2028 isn't coming from Google and Microsoft's operating cash.

Isn't that interesting?

About $1.15 trillion of it is borrowed: private credit, corporate bonds, and securitized debt. Another $350 billion arrives as private equity, venture, and sovereign money. Which means a great deal of the risk isn't sitting on bank balance sheets where regulators are looking. It's sitting in insurance books and pension allocations, which is precisely the concern four senators put to the Financial Stability Oversight Council back in January.

And it's being priced, right now. CoreWeave's five-year credit default swaps hit roughly 855 basis points in late July. A credit default swap is just insurance on a company's debt, so that's the market charging 8.55% a year to cover the risk of default, which works out to somewhere around a coin-flip chance of one over five years. Its most recent term loan cleared at 10.4% all-in, and lenders demanded maintenance covenants, which the leveraged loan market has largely not bothered with in over a decade.

I want to translate that, because "credit spreads widened" makes eyes glaze over.

It means professional lenders looked at an AI infrastructure company and said:

"We'll lend you the money, but we want the old rules back. The ones we stopped using when we felt safe."

That's the correction. And it's already underway. It just isn't happening on CNBC, because it's happening in loan documents.

Two curves moving the wrong way at the same time

Something I keep circling back to, as someone who builds things with these models rather than trading around them.

Between late May and early August, blended token prices fell about 43%. Ten weeks. OpenAI cut pricing on one model line by 80%, and Chinese open-weight models set a ceiling nobody can charge above.

Earlier in the same year, DRAM contract prices were running up 90 to 95% quarter over quarter, with memory expected to stay elevated through 2028.

Input costs up, output prices down, across twelve months.

That's not a market finding its level. That's a land grab funded by somebody else's balance sheet, and it works beautifully right up until the balance sheet develops an opinion.

Time to panic?

Perhaps surprisingly, no.

Everything above describes a capital structure problem rather than a technology problem. The chips will exist. The data centers will exist. The models will keep getting cheaper, which is the trend and not the bug 🪲. 

What's actually in question is whether the people who paid for all of it get paid back, and that's a completely different question from whether any of it works.

We've seen this shape before. Telecom companies laid over $100 billion of fiber between 1998 and 2000. Most of it sat dark for a decade. Then streaming showed up and used nearly all of it.

The infrastructure was never the mistake. The capital structure was.

The uncomfortable version of that, for anyone standing inside the industry: the infrastructure surviving and your employer surviving are unrelated events. 

Both can be true. The personal aspect of this is that only one of them pays your mortgage.

2027: Somebody is going to have to answer the depreciation question.

Whether any of this is profitable depends entirely on how long a GPU earns. Microsoft, Google and Oracle say four to six years. Meta says eleven or twelve. Nvidia ships a new architecture roughly every year.

Somebody is wrong. And 2027 is when the 2023 and 2024 hardware finally gets old enough that its useful life stops being a modeling assumption and may become an observable fact. Watch for any hyperscaler shortening its depreciation schedule. That isn't an accounting update, it's a confession, and it comes with earnings consequences for the whole sector at once.

2027 into 2028: The maturity wall. Debt raised for the 2023 buildout comes due into whatever credit conditions happen to exist by then. That's the mechanical stress point, and the most likely place for something to actually break rather than just grind.

After that: Absorption. Whatever survives owns compute at a written-down basis, and compute will progressively get less expensive. Which is good news if you build applications and survive the volatility of this projection. The fiber didn't stop existing when the telecoms went under. It just got cheaper. And then streaming happened.

The indicator I'd actually watch is the one nobody puts in a headline: aggregate productivity. The Fed can find individual people getting measurably more done with AI and cannot find it in the national numbers. If that's still true in 2030, this was a transfer rather than an investment. If it turns, everything above becomes a footnote about a bumpy financing period.

I don't know which one it'll be. 

Neither does anyone quoting a number at you with total confidence.

So yes, AI has affected my industry and we're still figuring out what that will look like. 

Disclaimer: Figures in this post are as reported through early September 2026, and several are projections or estimates rather than settled fact. I am not a financial advisor and this is not investment advice. It's my read on first hand obersvations and logical conclusions given personal research.