Wall Street Has Started Lending Against the Chips Themselves

Q: What could go wrong with this structure?

Three things, in rough order of likelihood. Utilisation falls below the assumption. The loans assume a chip earns for a certain number of hours. If demand for inference capacity softens, or if efficiency gains mean the same work needs fewer chips, revenue per unit falls while the repayment schedule does not. Power becomes the binding constraint. A financed fleet that cannot be energised does not earn. As we wrote when looking at what these sites actually need, the scarce inputs are increasingly electricity, land and cooling water rather than silicon. The collateral turns out to be correlated. If one operator fails because GPU economics deteriorated, every other lender holding the same collateral is holding it in the same bad market at the same moment. Collateral that all moves together is the specific ingredient that turns a bad quarter into a credit event.

Nvidia and six financial institutions are mobilising more than 500 billion dollars for AI infrastructure, with the graphics processors as collateral. The question nobody has answered is what a used GPU is worth in 2031.

Outspoken Digest Technology Desk

Sunday, August 16, 2026/3 min read

Rows of server racks in a data centre corridor — Editorial illustration generated for Outspoken Digest

Nvidia and six major financial institutions, reported to include BlackRock and Goldman Sachs, are working to mobilise more than 500 billion dollars for AI data centre infrastructure, using the graphics processors themselves as the security behind the lending.

The mechanism deserves attention, because it is a different kind of claim on the future than the one the industry has been making until now.

What GPU-backed financing actually means

Until recently, building a data centre was financed the way any large property is financed: against the land, the building and a contracted stream of revenue from tenants. The computers inside were an operating expense, and lenders treated them roughly the way they treat office furniture.

GPU-backed financing inverts that. The accelerators become the asset. An operator borrows against a fleet of chips, buys more chips, and repays from the revenue those chips generate while running.

It works if two things hold. The chips have to keep earning for long enough to repay the loan, and if the borrower fails, the chips have to be worth enough on resale to cover what is left.

The depreciation question

This is where it gets interesting, and where the honest answer is that nobody knows.

A data centre building has a useful life measured in decades. Its value is underwritten by a long history of comparable transactions. A high-end AI accelerator has a useful life that is currently a matter of opinion. It is physically fine after five years. Whether anyone wants to rent it at a price that covers its power draw is a completely different question, because its successor will do more work per watt.

The residual value of a five-year-old GPU therefore depends less on the chip than on the electricity price and on what has replaced it. That is a genuinely unusual collateral profile: an asset whose worth is destroyed not by wear but by the arrival of a better one.

Lenders handle uncertainty by discounting it, which means the terms carry the doubt even when the headline number does not.

Why the money is moving this way now

Because the alternative sources have limits. The largest technology companies have been funding this buildout from cash flow, and even their cash flow has a ceiling relative to the scale being proposed. Bringing in asset managers and investment banks moves the financing from corporate balance sheets to capital markets, which is where genuinely enormous sums live.

The same week produced other signs of the scale involved. TSMC approved a capital budget of about 29.44 billion dollars for capacity expansion and announced a joint venture with Sony for an image sensor plant in Japan reported at 6.4 billion dollars. Intel priced a stock offering to raise roughly 19.7 billion dollars as it moves towards its 14A node.

These are not the numbers of an industry adding capacity at the margin. They are the numbers of an industry rebuilding its base, in the same season that a new fabrication plant is going up on a retired coal site.

What could go wrong with this structure?

Three things, in rough order of likelihood.

Utilisation falls below the assumption. The loans assume a chip earns for a certain number of hours. If demand for inference capacity softens, or if efficiency gains mean the same work needs fewer chips, revenue per unit falls while the repayment schedule does not.

Power becomes the binding constraint. A financed fleet that cannot be energised does not earn. As we wrote when looking at what these sites actually need, the scarce inputs are increasingly electricity, land and cooling water rather than silicon.

The collateral turns out to be correlated. If one operator fails because GPU economics deteriorated, every other lender holding the same collateral is holding it in the same bad market at the same moment. Collateral that all moves together is the specific ingredient that turns a bad quarter into a credit event.

Is this a bubble?

That word does more harm than good here. What can be said precisely is narrower and more useful.

The demand for AI compute is real and currently exceeds supply. The financing structure being built to satisfy it is new, is being underwritten against an asset class with no long history of resale values, and is arriving at very large scale quickly.

All three of those statements can be true together. Whether the result is an efficient way to fund genuine demand or an efficient way to spread a mistake depends entirely on what a five-year-old accelerator is worth, and that number will not be observable for five years.

Published in The Outspoken Digest

Editorial desk

Outspoken Digest Technology Desk

Software, hardware, artificial intelligence and what they change for everyone else.

gpu backed financing ai data centres nvidia blackrock ai infrastructure investment chip depreciation

Newsletter