Nvidia Is Selling You a New Kind of Debt

Jensen Huang just convinced Goldman Sachs, BlackRock, Apollo, Blackstone, Brookfield, and KKR to treat GPUs like real estate. That is either the most elegant capital formation in tech history, or the setup for the worst-structured credit product since subprime CDOs. The distance between those two outcomes is thinner than Jensen’s jacket.

The pitch is clean: Nvidia and a consortium of the world’s largest asset managers are building $500 billion in financing to securitize compute. “These are revenue-generating assets now,” Huang told CNBC. “They’re productive, they’re long-lived, they’re fungible, they’re flexible.” Four adjectives. Each one doing serious work. Each one worth interrogating.

Start with “long-lived.” A GPU cluster has a useful economic life of roughly three to five years before the next architecture makes it look like a graphing calculator. Data centers built around H100s are already sweating their depreciation curves as Blackwell lands. The asset managers know this — they aren’t stupid — so either they’ve modeled the refresh cycle into the yield, or they’re pricing it wrong, or they’re planning to pass the rollover risk to someone further down the capital stack. History suggests the third option.

“Fungible” is doing even more load-bearing. The claim is that compute can be redeployed across use cases like barrels of oil can move between refineries. That’s approximately true at the infrastructure level and almost completely untrue at the workflow level. An H100 cluster optimized for LLM inference is not easily redeployed for genomics pipelines or autonomous-vehicle simulation without significant reengineering. Fungibility in commodities means the thing itself is interchangeable. Fungibility in compute means the thing plus the software stack plus the orchestration layer are interchangeable — and the software moats are exactly what makes AI valuable in the first place. You can’t have both.

“Revenue-generating” is the one that matters most, and it’s the one that depends entirely on a single assumption: that AI demand stays ahead of supply long enough for the securitization vehicle to mature. Right now, hyperscalers are spending hundreds of billions a year on capex and still can’t get enough compute. That’s a real signal. But the entire history of capital-intensive infrastructure cycles — railroads, fiber optic cable, commercial real estate, offshore drilling — is that the moment Wall Street packages the underlying asset into a structured product, supply catches up faster than the underwriters modeled. The packaging is the signal that you’re near the top of the demand curve, not the bottom.

None of this means Jensen is wrong. It means the trade has a specific shape. Nvidia wins regardless. It gets paid for the GPUs upfront, captures demand that would otherwise be constrained by hyperscaler balance sheets, and locks six of the largest asset managers in the world into a dependency on Nvidia’s continued architectural lead. The asset managers win if AI demand stays elevated and the yield clears their hurdle rates. The people who lose — if anyone loses — are whoever ends up holding the junior tranches when the first major GPU cluster gets marked down because the next Blackwell successor makes it uncompetitive.

The comparison to mortgage-backed securities is lazy and I’ll use it anyway: the MBS structure wasn’t wrong because houses are bad assets. It was wrong because the ratings models didn’t correctly price correlated default. The equivalent risk here isn’t that compute stops being valuable — it’s that the entire asset class gets marked down simultaneously when a new architecture lands, because every cluster in every securitization vehicle degrades at roughly the same time. That’s correlated depreciation. The collateral is fine; the correlation model is the question.

Here’s the steelman: maybe this time really is different because the demand signal is structurally unlike prior infrastructure booms. The railroad overbuild happened because anyone could lay track. The fiber glut happened because bandwidth cost-per-bit collapsed unexpectedly fast. Compute is different because the leading-edge fab capacity to produce the most capable chips is controlled by TSMC and two or three packaging vendors, and that supply constraint is structural, not cyclical. If the bottleneck is genuinely upstream and stays there, the securitization vehicle might actually work as advertised.

I don’t think that’s the bet the asset managers are making, though. I think the bet is simpler: they need yield, Nvidia needs demand, and $500 billion in structured compute financing lets both parties tell a story to their LPs that sounds better than “we’re lending money against depreciating chips.” Call it what it is — a confidence product. Confidence that AI demand is permanent. Confidence that Nvidia’s architecture stays dominant. Confidence that the correlation model is correct. All three of those might be warranted. But “confidence product” is not the same as “asset class.”

The most dangerous moment in any infrastructure cycle isn’t the buildout. It’s when the bankers give the infrastructure a ticker.