Nvidia Is Quietly Underwriting Its Own Boom

The chips that refuse to depreciate

Here's the thing nobody saw coming: old GPUs are getting more expensive, not cheaper.

Everyone assumed AI chips would follow the normal tech curve — buy the shiny new thing, watch last year's model lose value fast. As investor Gavin Baker put it, even the bulls in 2024 and 2025 expected prices to decline slowly; the bears thought they'd fall off a cliff. Nobody thought old GPUs would go 'vertical.'

But they are. Renting an Nvidia H100 for an hour now costs $2.71, up from $1.96 last November. One customer, Baseten, said its Blackwell B200 rental will jump from $2.63 to $5.10 an hour when its contract renews in October — nearly double. And CoreWeave, which rents out GPU infrastructure, said it's leasing 2020-era A100 chips out to 2029 and remains 'largely sold out' of older generations. Chips from five years ago. Still commanding a premium.

Why does this matter? Because a fast-depreciating asset can't be borrowed against. A durable one can. Rising, stable prices are what turn a pile of silicon into collateral — and that's the whole game here.

When a supplier starts underwriting the demand for its own product, it's telling you something: natural demand might not be enough on its own.

How a chip becomes a mortgage

Once you believe the asset holds its value, you can build a financing machine on top of it. And that's exactly what's happening.

Nvidia CEO Jensen Huang is making the case out loud. His pitch: because Nvidia's CUDA software (the coding platform that lets developers run and upgrade Nvidia chips) keeps old chips useful for years, they're 'fungible' — interchangeable, always in demand, always productive. In his words, that makes them 'rentable, durable and financeable.' Read that as a sales pitch to lenders, not just engineers. He's telling Wall Street these things are safe to lend against.

And Wall Street is listening. Nvidia is anchoring a deal with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize over $500 billion in outside capital for AI infrastructure. Blackstone's Jon Gray compared it to a mortgage: the bank underwrites you, but it also looks at the value of your house. Here the 'house' is the compute — valuable no matter who's renting it. BlackRock's Larry Fink went further, saying this is like the birth of the mortgage-backed securities market in the 1970s — where home loans got bundled up and sold to investors as bonds. He called it 'the next future for financial engineering.'

That's a thrilling line if you're bullish. It should also give you pause. The MBS market did become enormous — and it also blew up spectacularly in 2008 when the collateral turned out to be worth far less than everyone assumed.

Why 'guarantor of last resort' should worry you

Here's the uncomfortable loop. Nvidia sells the chips. Nvidia's software makes the chips durable. Nvidia is helping arrange the financing that lets buyers afford the chips. And Nvidia benefits when demand for the chips stays high. That's a lot of hats for one company to wear.

When a supplier starts helping underwrite the demand for its own product, it's telling you something: that natural, unassisted demand might not be enough to justify the buildout on its own. So the seller steps in to grease the wheels. That can be perfectly rational in a genuine boom — or it can be the mechanism that keeps a bubble inflating past the point where anyone would buy at unsubsidized prices.

The entire structure rests on one load-bearing assumption: that GPU prices stay high and stable. If a cheaper chip, a more efficient AI model, or simply a demand air-pocket knocks those rental rates down, the collateral loses value, the loans built on it look shaky, and the financing that funded the buildout gets a lot harder to roll over. Concentration is the risk here — a huge share of the AI economy's financing plumbing runs through one company's chips and one company's assurances about them.

Worth noting: this appetite for turning a hot asset into something you can borrow against and trade isn't unique to chips. Over in crypto, StablecoinX — a treasury company built around the ENA token — saw its shares jump 12% just for revealing it holds 20% of that token's supply. Different asset, same instinct: when something's going up, the market races to build leverage and financial structures on top of it. That's how you get scale. It's also how you get fragility.

What to watch next

The single most important number for anyone betting on the AI infrastructure trade is the GPU rental rate. As long as it keeps climbing — and Silicon Data's forward rates currently curve upward into 2027 and 2028 — the financeable-asset thesis holds and the capital keeps flowing.

Watch the forward curve, not just today's price. The forward curve is what the market expects to pay to rent these chips years from now, and right now it's betting they'll still be worth renting. If those forward rates start flattening or inverting, that's the early signal the collateral story is softening — well before today's prices actually fall.

And keep an eye on how much of the buildout leans on Nvidia-arranged or vendor-linked financing versus genuinely independent demand. The more the boom depends on the seller to backstop its own market, the more it looks like a machine that needs to keep running just to justify itself. For retail investors, the takeaway isn't 'sell everything' — it's 'know what you actually own.' If you're in this trade, you're not just betting on AI. You're betting that a five-year-old chip is still good collateral in 2029.