GPU Securitization — What Could Go Wrong

Goldman Sachs, BlackRock, and four other firms are lining up $500 billion to finance AI infrastructure. The collateral behind the debt is GPUs. The resale value of those GPUs depends on what the manufacturer decides to ship next year.

Data centers rent GPUs from infrastructure providers. Those rental payments become a predictable cash flow. Wall Street packages that cash flow into debt instruments, and because the SEC ruled in August that data center bonds don't count as asset-backed securities, those instruments are now eligible for institutional investors with infrastructure mandates. Pension funds. Insurers. The same pools of capital that hold toll road and pipeline bonds.

Nvidia sets the standard for which hardware qualifies as collateral, offers residual-value guarantees on that hardware, and controls the release schedule that determines when the next generation makes the current one less valuable. One company sits on every side of the deal.

The $500 billion handshake

On August 10, Nvidia signed memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to create financing platforms that would channel over $500 billion of third-party capital into AI infrastructure.

The pitch is straightforward. A one-gigawatt data center costs $50 to $60 billion to build. The hyperscalers are running out of internal cash to fund the buildout at this pace. Microsoft's free cash flow is expected to go negative for the first time since 2001.

Somebody has to finance the gap. Goldman CEO David Solomon framed it directly: the goal is to "create a market for credit backed by Nvidia compute."

How you turn a GPU into collateral

Wall Street will lend against almost anything, but to securitize at scale, you need four things: a standard the market recognizes, predictable cash flows, a way to price the collateral, and a secondary market to transfer risk.

Nvidia claims all four. CUDA and its hardware ecosystem define the standard. Take-or-pay contracts, where the customer uses the compute or pays anyway, supply the cash flows. GPU rental markets provide the price signal. And securitization turns the whole thing into tradeable paper.

H100 GPUs currently rent for about $2.71 per hour, up from $1.96 six months ago. Collateral whose price is rising is exactly the kind of thing Wall Street loves to package.

Nvidia sweetens the structure by offering residual-value support of up to 25% per project. If the used GPUs are worth less than expected when the financing matures, Nvidia absorbs part of the loss.

The depreciation question

Jensen Huang says compute is infrastructure. Underwrite the income stream like a toll road. The deal terms say something different.

Nvidia is offering residual-value guarantees, and nobody insures the residual value of a toll road. They do for used cars.

So which is it? Infrastructure or equipment?

Nvidia shipped the A100 in 2020, and CoreWeave has contracted demand for those chips through 2029. Nine years of useful life looks more like infrastructure than a depreciating machine. Rental prices for older GPUs are rising, not falling.

But GPU residual values have never been tested in a demand downturn. Michael Burry, the investor who predicted the 2008 crisis, has argued that depreciating chips over six years is already too generous. The question isn't what a five-year-old GPU earns when demand is exponential. It's what it earns when demand flattens.

Think of it like a commercial oven in a restaurant. When the restaurant is packed every night, the oven pays for itself many times over. If the neighborhood changes and foot traffic drops, the oven still works but its resale value craters. GPUs face the same dynamic, except the "neighborhood" is AI demand, and the company selling new ovens also decides when to release a model that makes the old one half as efficient per watt.

One company, four roles

Fannie Mae created the conforming loan standard that the mortgage market ran on. Lenders who originated mortgages meeting Fannie's criteria could sell them into the secondary market, freeing up capital to originate more. The standard created liquidity. Liquidity created the modern mortgage market.

Nvidia occupies the same structural position for compute. Its CUDA ecosystem defines what counts as bankable hardware. If your data center runs Nvidia chips with CUDA software, it meets the conforming standard. If it doesn't, good luck finding financing at these rates.

But Fannie didn't build houses. Nvidia makes the chips. And Nvidia controls the release cadence that determines when today's collateral becomes last generation's. When Nvidia ships a new architecture that's twice as efficient per watt, the secondary market for the old one shifts. Not because the old chips stopped working, but because the economics changed.

In the mortgage market, no single entity set the conforming standard, manufactured the collateral, and decided when new supply would arrive. Nvidia holds all four roles in GPU-backed debt.

Role	Mortgage-Backed Securities	GPU-Backed Debt
Sets the standard	Fannie Mae / Freddie Mac Conforming loan criteria	Nvidia CUDA ecosystem + hardware specs
Manufactures the collateral	Homebuilders Lennar, D.R. Horton, etc.	Nvidia Designs and ships the GPUs
Guarantees residual value	Fannie Mae / Freddie Mac Implicit government backstop	Nvidia Up to 25% per project
Controls new supply timing	Market-driven Thousands of independent builders	Nvidia Sets the chip release cadence

The regulators cleared the way

The same day Nvidia announced the $500 billion partnerships, the SEC published guidance exempting a major category of data center bonds from post-2008 securitization rules.

The reasoning: data centers are physical infrastructure that keeps operating and earning revenue, unlike a mortgage that pays down to zero. So bonds backed by data center assets don't need the risk retention rules that force issuers to keep skin in the game, or the disclosure requirements added after 2008.

The logic is defensible. A data center is not a mortgage. But the post-2008 rules exist because securitization markets, left without alignment-of-interest requirements, have a track record of mispricing risk and distributing it to investors who don't fully understand what they hold.

Removing those guardrails for a brand-new asset class, on the same day that $500 billion of capital is being lined up to fill it, is worth paying attention to.

The financing needs the curve to keep climbing

The deals are still memorandums of understanding, subject to final agreements. The A100 is still earning. The GPUs are not dark.

But roughly half the contracted backlog in AI infrastructure sits with labs that have never turned a profit. Their compute bills get paid by raising new capital at higher valuations. That financing works because the demand curve keeps climbing.

Anthropic went from $9 billion of annualized revenue in December to about $74 billion by late July. Possibly the fastest private company growth story in history.

Monthly growth cooled from roughly 51% in May to about 8% by July. Not a crisis, not a slowdown. Just the exponent softening.

A financing machine built for exponential curves does not require failure to get into trouble. It requires the exponent to slow down.

The mortgage market didn't break because Americans stopped needing houses. It broke when house price appreciation cooled from 15% to 8%, and the refinancing math that the whole structure depended on stopped working.

GPU compute is an asset class now. The question is how much paper has already been written against demand that must keep climbing.