The old GPUs get cheap first: H100 rents fall as the new chips arrive, the newest hold

A GPU earns its keep by the hour, so the price of a GPU-hour decides whether the companies renting AI chips out on borrowed money can pay for them. The claim here is that the price does not move as one: each new NVIDIA generation pushes the previous one down the price curve while the newest chips stay scarce. Blackwell is now most of NVIDIA's shipments, and its successor, Rubin, begins shipping in volume over the next year, so the chip most likely to get cheaper is the H100, which went into volume service in 2023. MenFem records rental prices every week. On vast.ai, a marketplace of observed offers, the H100 SXM averaged $2.38 an hour over the four weeks to 6 November, against $2.97 on 25 September, while the B200 averaged $9.69, up from $9.38. Microsoft's list prices on Azure did not move in six weeks, with a five-fold gap between spot ($2.27) and on-demand ($12.29) for the same H100. The renters that matter most are shielded: CoreWeave sells about 98% of its capacity on committed contracts. So a fall in H100 rents hits the uncontracted part of the market first, and the borrowers whose loans outlast their contracts, the split MenFem's debt-financed compute thesis turns on. Conviction is LOW. MenFem's series is only seven weeks old and swings 10–20% from week to week; outside indices disagree by as much as two times; and the most-quoted H100 rental index rose about 6% in the month to 5 November, with Nebius reported to have raised its H100 price twice this year. In 2023–24 H100 rents fell by about three quarters in a year, but that is a precedent, not a forecast. BULLISH on the thesis, conviction LOW, one year. This view was formed by MenFem's research process at Connor's request and adopted by him.