What Nvidia earnings probably won’t show us
Nvidia reports this afternoon, and the whole market will likely treat it as a referendum on the AI bubble. Analysts expect another enormous quarter, around $108.7 billion in revenue, and people will read the stock move as a verdict on whether all this spending was worth it.
I think the number tells us less than people expect. Nvidia's revenue counts chips sold. We already know demand for those chips is insanely high and won’t be slowing down. We spend our days running the fleets those chips end up in, and from where we sit, the interesting stuff happens after the chip ships: whether it gets plugged in, whether it gets used, and who's on the hook if it doesn't.
Here's what I'm actually paying attention to.
- Chips shipped and chips running are two very different numbers. GPUs ship in months. The power gear they need takes years: Wood Mackenzie found lead times on the big grid transformers topped 160 weeks earlier this year. Even Microsoft is feeling it. Its CFO said 2026 spending will land around $170 billion because third-party data centers are running late. So a big Nvidia quarter can sit right next to a lot of hardware that isn't earning anything yet. Nobody reports how many GPUs are actually plugged in, and that's the number I'd most want to see.
- If something breaks, I'd bet it starts with the debt. Demand honestly looks fine to me. Open-weight models are a good gauge, since anyone can run them on rented GPUs. On OpenRouter they went from about 27% of tokens a year ago to about 72% by September, and their volume grew more than 60x. Vercel sees the same thing in production: open-weight models ran about 72% of the tokens on its AI gateway in September but only 23% of the spend. The money is starting to follow, since that spend share has quadrupled since June. Still, the fastest-growing slice of demand is also the cheapest per token, and that matters if you borrowed to buy the GPUs serving it. The shaky part is how the GPUs got paid for. A lot of them were bought with loans backed by the GPUs themselves, on the assumption the chips hold their value for five or six years. Credit markets are already nervous: Bloomberg reported CoreWeave's debt is pricing in better-than-60% odds of default within five years, even with a $104 billion backlog. And a handful of AI labs sit under most of that demand. If one slows down, an operator can lose its biggest customer and watch its collateral get cheaper in the same month.
- The weirdest thing in the market right now is idle GPUs sitting next to rising prices. Rental prices went up this year. Nebius raised its on-demand H100 price from $3.85 to $4.50 an hour in October. Meanwhile a Cast AI study of about 23,000 clusters found average GPU utilization of 5%. Both are true because capacity gets stuck with whoever reserved it, used or not. That's why we've started working on compute recycling: taking underused capacity and putting it to work in spot markets and batch jobs. It matters on the way down too. When a borrower gets squeezed, the lender seizes the GPUs and dumps them, and that drags down every GPU backing every other loan. Capacity that's already earning in spot and batch is a lot less likely to end up in a fire sale.
- Some of the best deals in AI right now might be GPUs that are already written down. Every new Nvidia generation pushes the last one into the resale market, and with Blackwell ramping and Rubin on the way, that's happening to H100s now. Used H100s were trading around $15,000 to $20,000 this summer, well under half of what they cost at the peak. Rental rates haven't fallen nearly as hard: the H100 on-demand median is $3.92 an hour this month. Napkin math, a used H100 running about 70% of the time at around $3 an hour earns back its price in roughly a year, before power and hosting. That only works if you can keep it busy, which is exactly why I think this phase rewards whoever runs older hardware best. There's real risk here too. If Nvidia cuts prices on new chips, the used market can reprice overnight. IBM did exactly that to its mainframes in 1979, and the people holding used machines took the hit.
My bet is the boom continues and some people financing it still get burned. If there’s a reckoning, it’ll look like a lot of GPUs changing owners, and they’ll end up with whoever can keep them busy.