October 20, 2026 · Timothy O’Brien
Old Chips Are Supposed to Get Cheaper
The H100 rental price peaked near $8 an hour, fell to $2, then rose again — on a chip that is now three years old. What that says about reading a price curve.
Yesterday I wrote that a SKU is not a machine. This one is about the price of the machine’s time, and about the one rule everyone applies to hardware prices: things get cheaper as they get older. Yesterday’s flagship is this year’s bargain. The H100 spent three years obeying that rule, and then it broke it.
The whole history fits in one table
| When | H100 per GPU-hour | What was happening |
|---|---|---|
| 2023, launch era | $4.70 to over $8 | Scarcity pricing, if you could get one at all |
| August 2024 | $1 to $2 on marketplaces | Supply caught up in about eighteen months |
| December 2025 | $2.85 to $3.50 | AWS cut P5 pricing about 44 percent that June |
| Oct 2025 to Mar 2026 | $1.70 to $2.35, one-year contracts | Up almost 40 percent; on-demand sold out |
| Aug to Sep 2026 | $3.38 to $3.84, tracker medians | Spot up 33 percent year over year |
Read the middle rows and the story is the one everyone priced in: the rental bubble burst, supply caught demand, the oldest chips got cheap. Financial analysts spent 2025 treating any GPU provider that used a six-year depreciation schedule as a walking write-off.
Then read the last row. A three-year-old chip got more expensive. Not “stayed firm” — the spot price rose 33 percent year over year, and one-year rental contracts rose 40 percent off their October 2025 low while on-demand capacity simply sold out. Even the A100, one generation older and nominally obsolete, is up 17 percent year over year. 56
Why the rule broke
The price of a GPU-hour is not a property of the chip. It is the price of the seconds nobody else has booked. While the fleet was young and oversized, those seconds were nearly free, and the chip looked like it was depreciating on schedule. In early 2026 the slack disappeared: inference demand ate the surplus, memory prices went parabolic enough that server makers repriced and some operators withheld supply, and Blackwell lead times pushed buyers back onto Hopper hardware that was supposed to be fading. 7
So the price re-inflated around the old chip. Depreciation schedules assume demand fades with the hardware; demand did the opposite.
The second thing the table hides is that the H100 never has one price. Right now the same silicon rents for about $1.50 an hour on a marketplace and about $12 on a hyperscaler, depending on who is selling it and what comes attached. The trackers below show that spread every day, which is why I am not going to reproduce it here. 8
Where to look
- Silicon Data — a daily rental price index for the H100, and separate indexes for the H200, A100, B200, B300, and MI300X. H100 ticker: SDH100RT.
- SemiAnalysis — the H100 one-year rental index, public, updated monthly.
- GetDeploying — weekly medians since July 2024, and the weekly dataset behind the charts is free.
- AIMultiple — free weekly medians across more than 60 providers, history back to July 2024.
- Vast.ai — live marketplace supply, demand, and pricing, updated every five minutes.
What the curve is worth
If you priced a three-year commitment off the 2025 trend line, you would be signing it into a market that no longer exists. The current prices are on the trackers’ pages. The history — the part that tells you whether today is cheap or expensive — is the part nobody keeps for you, and the part worth keeping yourself. 9
Eugene Cheah, “$2 H100s: How the GPU Rental Bubble Burst,” Latent.Space, October 11, 2024: link — launch-era rates around $4.70/hr rising over $8, $1–2/hr on marketplaces by August 2024.↩︎
Silicon Data, “H100 Rental Price Over Time (2023–2025),” December 21, 2025: H100 Rental Price Over Time — monthly hyperscaler medians $7.62–$8.24 through May 2024, marketplace/neocloud $2.50–$3.20 from mid-2024.↩︎↩︎
SemiAnalysis, “The Great GPU Shortage — Rental Capacity,” April 2, 2026: The Great GPU Shortage — H100 one-year rental up almost 40 percent from $1.70/hr (October 2025) to $2.35/hr (March 2026), on-demand sold out, memory prices “going completely parabolic,” server repricing withheld supply.↩︎↩︎↩︎
Shattered, “H100 Rental Prices Halve to $3.38/Hr in 2026,” September 22, 2026: H100 Rental Prices Halve — AIMultiple index median $3.38/hr (August 25, 2026); Mercatus $3.84/hr across 27 providers (September 13, 2026); H100 spot $2.77–$2.80 up 33 percent year over year and A100 $1.65 up 17 percent per Bank of America notes as reported by TheStreet and TechFlowPost. Introl’s December 2025 collapse figures: GPU Cloud Prices Collapse.↩︎↩︎
Akash, “H100 Rental Price in September 2026,” September 2026: H100 Rental Price — rates checked September 12, 2026: Vast.ai marketplace about $1.73 and up, RunPod $1.99–$2.89, Lambda $3.29–$3.99, AWS P5 about $6.88, Azure ND about $12.29 per GPU-hour.↩︎
