An H100 hour now costs 44% more than it did ten months ago.

GPU market · H100

The market prices the hour.
Not the card behind it.

H100 rental prices are up 44% since last December, for a generation that would normally be getting cheaper. Rental markets quote a model and an hour. As Hopper cards move into the resale market, the price will also have to reflect what each card has been through.

Silicon Data’s H100 Rental Index reached $2.88 per GPU-hour on 24 October 2026, up from $2.00 on 9 December 2025, a rise of 44%. Nebius raised its on-demand H100 price from $3.85 to $4.50 on 1 October. Older chips are usually discounted as buyers move to newer silicon. This time, they are not.

$2.88

H100 Rental Index per GPU-hour, 24 Oct (+44%)

$4.50

Nebius on-demand H100, from $3.85

$3.99

Runpod Secure Cloud H100 SXM, from $3.29

Figures as reported by Startup Fortune, 25 October 2026, citing Silicon Data, Nebius and Runpod.

The Blind Spot in Compute Pricing

A rental price quotes a model and an hour. It does not quote the card. An hour on a GPU with clean memory and an hour on one that has remapped memory rows, logged Xid 48 errors and spent weeks throttling under heat appear on the same price list.

For rentals, that is tolerable: a bad card can be swapped. For a purchase, it is not. When the same cards are sold rather than rented, the buyer keeps whatever history comes with them.

Condition Lives in the Logs

A card’s real condition is not on the spec sheet. It is in ECC counters, row-remap records, kernel logs and telemetry. Some of it survives on the card itself; most of it stays with the operator who ran it.

The same article cites Gartner’s forecast that 80% of neocloud providers will cease independent operations by 2030. If that happens, many Hopper cards will change hands. The useful question will not be only “how much per hour?” but “what does this card’s record show, and can I check it?”

Rental markets price time. The resale market will have to price condition, and condition is only worth what the seller can prove.

Source: Startup Fortune: H100 rental prices are spiking even as Wall Street worries about an AI bubble (25 Oct 2026)