GPU Rental Price Index: September 2026.
What renting an H100, H200, A100, L40S or RTX 4090 actually cost in September 2026: market and cheapest hourly prices from 292,295 listings, and what moved.
By Vi Nguyen · Published · Prices on this page are live · Editorial policy
The numbers
The market price is what a typical provider charged for one GPU per hour — the median across providers, day by day, over the month. The cheapest column is the lowest-priced provider over the same month, which is what you'd pay if you shopped around and were flexible about where you rent. Both are medians of daily values, so a single promotional or mispriced listing doesn't move them.
| GPU | VRAM | Market $/hr | vs prev. month | Cheapest $/hr | Cheapest at | Providers |
|---|---|---|---|---|---|---|
| Nvidia B200 | 192GB | $5.60/hr | +2.1% | $3.69/hr | DeepInfra | 8 |
| Nvidia H200 | 141GB | $3.14/hr | -3.0% | $2.30/hr | Vast.ai | 7 |
| Nvidia H100 SXM | 80GB | $2.39/hr | +14.4% | $1.32/hr | Lium | 13 |
| Nvidia H100 PCIe | 80GB | $2.12/hr | +41.5% | $1.00/hr | io.net | 7 |
| Nvidia H100 | 80GB | $3.00/hr | +6.9% | $2.29/hr | Theta EdgeCloud | 3 |
| Nvidia A100 80GB PCIe | 80GB | $1.44/hr | — | $0.30/hr | Lium | 8 |
| Nvidia A100 SXM | 40GB | $1.39/hr | +27.5% | $1.39/hr | io.net | 7 |
| AMD MI300X | 192GB | $0.50/hr | +0.0% | $0.50/hr | RunPod | 3 |
| Nvidia L40S | 48GB | $0.79/hr | +0.0% | $0.38/hr | Lium | 11 |
| Nvidia RTX 6000 Ada | 48GB | $0.65/hr | +0.0% | $0.54/hr | Vast.ai | 5 |
| Nvidia RTX A6000 | 48GB | $0.38/hr | +13.9% | $0.32/hr | Nosana | 10 |
| Nvidia L4 | 24GB | $0.70/hr | +58.0% | $0.083/hr | Clore.ai | 8 |
| Nvidia GeForce RTX 5090 | 32GB | $0.64/hr | +58.6% | $0.32/hr | Nosana | 10 |
| Nvidia GeForce RTX 4090 | 24GB | $0.36/hr | +20.7% | $0.29/hr | Nosana | 11 |
| Nvidia GeForce RTX 3090 | 24GB | $0.19/hr | +27.2% | $0.12/hr | Vast.ai | 8 |
What moved in September 2026
- Biggest drop: the Nvidia H200, down 3.0% month on month.
- Biggest rise: the Nvidia GeForce RTX 5090, up 58.6%.
- 4 of 14 GPUs moved by less than 3% — a month where most prices shifted noticeably.
Datacenter vs consumer cards
The median datacenter card in the basket rented for $2.12/hr at market price, against $0.38/hr for consumer and workstation cards — a 5.5× gap. That gap is what you pay for more memory, HBM bandwidth, and (on SXM parts) NVLink. If your model fits in 24–48 GB and you don't need multi-GPU scaling, the consumer side of the table is where the savings are; our guide to how much VRAM an LLM needs shows where that line falls.
Shopping around mattered most for the Nvidia L4: the cheapest provider (Clore.ai) came in 88% under the market price. Cheapest listings on marketplaces often come from individual hosts, so weigh reliability as well as price — see marketplaces vs managed clouds.
How to use this report
Use the market price to budget, and the cheapest price as the floor you can reach with some shopping. Neither is a quote: marketplace prices move by the hour, which is why every GPU name above links to its live page with today's prices by provider. For a model-first view, the VRAM fit calculator lists every card that can run a given model with its current cheapest price.
The basket is fixed so months stay comparable. A GPU missing from a month simply had no single-GPU listing that month. Full details are in the methodology.