What’s a used H100 worth in four years? If you can’t answer that with a straight face, you’ve found the exact question Wall Street is now asking Nvidia, and the reason a financing plan announced in August is running into friction.
Here’s the setup. Nvidia wanted its chips to work like collateral. Not just products you buy, but assets you borrow against. The company lined up financiers including Blackstone, Apollo, and KKR with the goal of building a lending market modeled on aircraft leasing. The logic is clean: airlines don’t buy planes outright, they finance them, because a Boeing holds predictable value for decades and lenders know roughly what it’ll fetch at resale. Apply the same structure to GPUs and suddenly AI companies can tap capital pools far deeper than venture funding or cash reserves.
According to Reuters reporting, some lenders are now asking for higher guarantees than the plan originally laid out. They’re more cautious than Nvidia about how long these chips keep generating revenue. That’s the whole story, and it’s more interesting than its sentence count suggests.
Why the aircraft comparison breaks
I spend my working hours testing AI tools and watching how fast the underlying infrastructure shifts beneath them. So the lender skepticism reads as entirely reasonable to me.
A 737 built in 2015 does the same job in 2025. Physics hasn’t changed. Air is still air. A GPU built in 2015 is close to irrelevant for current model training, not because it stopped working, but because the work changed around it. Nvidia itself keeps shipping new architectures that make the previous generation look slow. The company’s own product cadence is the strongest argument against treating its hardware as a long-duration asset.
That’s the awkward position. To sell chips, you promise each generation is a meaningful leap. To finance chips, you need to promise the old ones hold value. Those two pitches sit uncomfortably next to each other, and lenders are the ones holding the calculator.
What this means if you’re building on top of all this
Most readers here aren’t structuring debt facilities. You’re picking tools, paying for API credits, deciding whether to commit to a platform for the next two years. Chip financing feels several layers removed from that. It isn’t.
The costs of acquiring compute flow directly into what you pay for inference, fine-tuning, and hosted models. A few things follow from cheaper or more expensive capital:
- If financing gets easier, more companies can afford large GPU clusters, which means more competition among providers and more pressure on prices.
- If lenders demand steeper guarantees, the cost of compute rises for everyone downstream, including the startups whose tools you’re evaluating.
- Companies that depend on cheap capital to subsidize generous free tiers become less reliable bets for anything you plan to run long-term.
That last point is the one I’d actually act on. I’ve reviewed plenty of tools priced below what they cost to operate. That gap gets funded somehow. When funding gets more expensive, the gap closes, usually through price increases or rate limits that arrive without much warning.
Caution as a signal, not a verdict
I want to be careful not to overread this. Lenders negotiating for better terms is lenders doing their job. Asking for higher guarantees isn’t the same as walking away, and the reporting describes friction in a deal structure, not collapse. Nvidia is still selling everything it makes. The financing idea may well get built with terms that satisfy both sides.
What’s notable is who is applying the brakes. For two years, the loudest voices on AI infrastructure have been the people selling it and the people buying it, both motivated to describe an expanding future. Credit markets have a different motivation. They get paid for being right about downside scenarios. When that group looks at GPU resale value and asks for more cushion, it’s a data point from a party with no reason to inflate the story.
The mismatch Reuters describes, between how Nvidia values its chips over time and how lenders do, could make it harder for AI companies to reach those new capital pools. That’s a real constraint on how fast this gets built out.
What I’d watch
Not stock prices. Watch pricing pages. Watch whether free tiers on the tools you rely on quietly shrink, whether enterprise contracts start requiring longer commitments, whether providers push harder toward smaller models that cost less to serve. Those are the places where capital costs show up in work you can actually feel.
The question of what a used GPU is worth in 2029 sounds like a problem for structured finance desks. It’s going to end up on your invoice.
đź•’ Published: