GPU price index
median across GPU configurations$0.26 / GPU-hr+$0.05
Live chart · updates with every published snapshot
H100 GPU rental price
The live median hourly rental price for an NVIDIA H100 80 GB SXM5 on the Akash open compute marketplace.
median price per GPU-hour$2.55 / GPU-hr
Behind the number
Where this sits in the field
- NVIDIA GTX1050 2 GB PCIe$0.05
- NVIDIA GTX1070TI 8 GB PCIe$0.08
- NVIDIA RTX 3060 12 GB PCIe$0.14
- NVIDIA RTX 3090 24 GB PCIe$0.15
- NVIDIA RTX 4080 16 GB PCIe$0.15
- NVIDIA T4 16 GB PCIe$0.15
- NVIDIA RTX 6000 24 GB PCIe$0.17
- NVIDIA RTX 8000 48 GB PCIe$0.21
- NVIDIA P2000 5 GB PCIe$0.30
- NVIDIA RTX 4090 24 GB PCIe$0.34
- NVIDIA RTX 4000ADA 20 GB PCIe$0.41
- NVIDIA PRO 6000WE 96 GB PCIe$1.76
- NVIDIA H100 80 GB SXM5$2.55
The board in rank order, with the tracked subject of this chart highlighted. Showing the top 13 of 16 ranked entries.
| Rank | Entry | median price per GPU-hour |
|---|---|---|
| 1 | NVIDIA GTX1050 2 GB PCIe | $0.05 / GPU-hr |
| 2 | NVIDIA GTX1070TI 8 GB PCIe | $0.08 / GPU-hr |
| 3 | NVIDIA RTX 3060 12 GB PCIe | $0.14 / GPU-hr |
| 4 | NVIDIA RTX 3090 24 GB PCIe | $0.15 / GPU-hr |
| 5 | NVIDIA RTX 4080 16 GB PCIe | $0.15 / GPU-hr |
| 6 | NVIDIA T4 16 GB PCIe | $0.15 / GPU-hr |
| 7 | NVIDIA RTX 6000 24 GB PCIe | $0.17 / GPU-hr |
| 8 | NVIDIA RTX 8000 48 GB PCIe | $0.21 / GPU-hr |
| 9 | NVIDIA P2000 5 GB PCIe | $0.30 / GPU-hr |
| 10 | NVIDIA RTX 4090 24 GB PCIe | $0.34 / GPU-hr |
| 11 | NVIDIA RTX 4000ADA 20 GB PCIe | $0.41 / GPU-hr |
| 12 | NVIDIA PRO 6000WE 96 GB PCIe | $1.76 / GPU-hr |
| 15 | NVIDIA H100 80 GB SXM5 | $2.55 / GPU-hr |
$2.55 / GPU-hr on every publication so far. A line appears here as soon as the number moves.
About this metric
The H100 is the accelerator most AI training and inference economics are quoted against. This chart tracks the median advertised price to rent one H100 80 GB SXM5 by the hour on Akash, an open GPU marketplace whose full order book is public. Because the venue publishes every listing, the median is computed from real offers rather than a vendor's rate card.
Each point is the median across active positive-price H100 SXM5 listings at publication time, normalized to price per GPU-hour. Zero-price and unavailable listings are excluded, and the same aggregate tape feeds the availability figures on the underlying board. Open-marketplace prices sit below big-cloud list prices and move faster, which is exactly why they are worth watching.
Rental price is the purest signal of AI compute supply and demand. When H100 hours get cheaper, training runs, fine-tunes, and inference-heavy products all get cheaper to operate, and marginal ideas become viable. When the line rises, capacity is tight and the advantage shifts to whoever already owns accelerators.
Related charts