The GPU rush: why an older H100 rents for more this fall

Computer hardware usually gets cheaper as it gets older. A new chip ships, a faster one follows, and the old one slides down the price list.

The H100 is going the other way. NVIDIA announced it in March 2022, and two generations have followed it: Blackwell and Vera Rubin. Yet in the fall of 2026, an H100 rents for more than it did a year ago.

That is the GPU rush in one chip. Here are the numbers behind it, and what they mean if you rent GPUs.

The numbers

Three different measures point the same way:

• Contracts. SemiAnalysis tracks the rate of one-year H100 rentals. It climbed from $1.70 per GPU-hour in October 2025 to about $2.35 in March 2026, a rise of 38%.
• Market indexes. Silicon Data's H100 index went from about $2.53 per GPU-hour in mid-2026 to $2.81 on October 7, and Ornn's index of executed trades settled at $2.86 on October 6.
• List prices. Nebius raised its on-demand H100 from $3.85 to $4.50 per GPU-hour on October 1, and its H200 from $4.50 to $5.40.

Why an older chip is in demand

Demand keeps climbing. NVIDIA's Data Center revenue reached $89.0 billion in the quarter it reported on August 26, up 117% from a year earlier, and NVIDIA guided to $108.0 billion of total revenue for the next quarter. Jensen Huang put it this way: "The AI infrastructure buildout is at full steam."

Capacity is tight. SemiAnalysis reported this year that on-demand capacity for the major models was effectively sold out across the clouds it surveyed, that eight-node H100 and H200 clusters were scarce, and that Blackwell lead times had stretched into June and July.

The H100 still does the work. Its 80 GB of HBM3 memory holds a lot of today's models. OpenAI's gpt-oss-120b fits on a single 80 GB GPU, and Black Forest Labs names an "H100-equivalent GPU" for FLUX.2 [dev]. Running a model asks for memory and steady speed, and an H100 has both. Coloprice's history of H100 rates sums it up: inference growth "re-priced 'obsolete' hardware."

One GPU, many prices

The market is wide as well as rising. Uvik read the on-demand list prices of up to 16 clouds on October 7, 2026. All figures are in USD per GPU-hour.

The highest list price for an H100 SXM is nearly four times the lowest. AWS, for one, lists its eight-GPU p5.48xlarge at $55.04 an hour, which is $6.88 per GPU, and that is after a 44% cut in June 2025.

So the gap between a good rate and a poor one for the same GPU is far larger than the rise of the whole market this year. For anyone renting, that gap is the number to watch.

What this means if you rent

• Start from memory, not from the name. The cheapest GPU is the one that fits your model, and it may be smaller than you expect. We wrote a guide to how much VRAM a model needs.
• Compare per GPU-hour. Clouds list prices per instance, and an eight-GPU price reads very differently from a one-GPU price.
• Pay by the second. Work that takes 20 minutes should bill 20 minutes. When rates climb, a GPU sitting idle is the first thing to stop.

Where Texel fits

Texel is made for this market. We are building the onchain GPU cloud on Robinhood Chain, where you pay with any token and Texel routes every rental and output to the cheapest GPU that fits. When one GPU has this many prices, the route you take is worth real money, for you and for your agent.

Sources

• NVIDIA, NVIDIA Announces Hopper Architecture, the Next Generation of Accelerated Computing, March 22, 2022.
• SemiAnalysis, post on its H100 rental index, June 12, 2026.
• Nebius, prices.
• Uvik, GPU Price Index 2026, measured October 7, 2026.
• NVIDIA, NVIDIA Announces Financial Results for Second Quarter Fiscal 2027, August 26, 2026.
• OpenAI, gpt-oss-120b and gpt-oss-20b model card.
• Black Forest Labs, FLUX.2.
• Coloprice, H100 rental prices in 2026, August 2, 2026.
• Vantage, p5.48xlarge.