A price per GPU-hour says nothing about what the hour delivers. The only public measurements these sellers stand behind are their own MLPerf Training submissions, so this puts them beside the prices published here. Most of what the table shows is what is missing, and that is the result rather than a caveat.
Two sellers can only be divided by one another where they ran the same benchmark, on the same accelerator, at the same GPU count, in the same round. MLPerf revises its suite between rounds, so the same benchmark name in two rounds is not the same workload. Across the 3 rounds collected that leaves 1 cell in which both sellers also have a price here.
MLPerf Training v5.0 · each seller's own submission
| Seller | Time to train min | Price USD/GPU‑hr | Cost of the run USD | Priced in |
|---|---|---|---|---|
| Lambda | 10.9 | 6.79 | 9.89 | DE |
| Oracle | 11.0 | 14.00 | 20.49 | DE, ES, FR, IT, NL, SE |
Delivered performance differs by 0.5%. The cost of the same run differs by 2.07×. On this workload the difference between these two sellers is price rather than speed.
Time to train is the median of that seller's own successful runs, computed from the logs it published. It is not MLCommons' official score: MLCommons publishes those itself and the scoring rule varies by benchmark. Cost is that time multiplied by the accelerator count and by this site's EU/EEA price for that seller's matching accelerator on 2026-09-13.
| Seller | Round | Benchmark | Accelerator | GPUs | Minutes | USD/GPU‑hr | Run cost USD |
|---|---|---|---|---|---|---|---|
| CoreWeave | v6.0 | gpt_oss_20b | NVIDIA Blackwell GPU (B200-SXM-180GB) | 64 | 27.0 | no EU price | — |
| CoreWeave | v6.0 | llama31_8b | NVIDIA Blackwell GPU (B200-SXM-180GB) | 64 | 16.3 | no EU price | — |
| v5.0 | stable_diffusion | NVIDIA H200-SXM5-141GB | 8 | 27.1 | no EU price | — | |
| v5.0 | stable_diffusion | NVIDIA H200-SXM5-141GB | 64 | 5.4 | no EU price | — | |
| Lambda | v6.0 | gpt_oss_20b | NVIDIA Blackwell GPU (B200-SXM-180GB) | 8 | 95.7 | 6.79 | 87 |
| Lambda | v5.0 | llama2_70b_lora | NVIDIA Blackwell GPU (B200-SXM-180GB) | 8 | 10.9 | 6.79 | 10 |
| Lambda | v6.0 | llama31_8b | NVIDIA Blackwell GPU (B200-SXM-180GB) | 8 | 83.8 | 6.79 | 76 |
| Nebius | v5.1 | flux1 | NVIDIA Blackwell GPU (B200-SXM-180GB) | 32 | 93.8 | no EU price | — |
| Nebius | v6.0 | flux1 | NVIDIA Blackwell Ultra GPU (B300-SXM-270GB) | 32 | 76.4 | no EU price | — |
| Nebius | v6.0 | flux1 | NVIDIA Blackwell Ultra GPU (B300-SXM-270GB) | 64 | 45.8 | no EU price | — |
| Nebius | v6.0 | gpt_oss_20b | NVIDIA Blackwell Ultra GPU (B300-SXM-270GB) | 8 | 82.3 | no EU price | — |
| Nebius | v6.0 | gpt_oss_20b | NVIDIA Blackwell Ultra GPU (B300-SXM-270GB) | 64 | 23.2 | no EU price | — |
| Nebius | v5.1 | llama2_70b_lora | NVIDIA Blackwell GPU (B200-SXM-180GB) | 8 | 9.6 | no EU price | — |
| Nebius | v5.1 | llama2_70b_lora | NVIDIA Blackwell Ultra GPU (B300-SXM-270GB) | 8 | 8.8 | no EU price | — |
| Nebius | v5.1 | llama2_70b_lora | NVIDIA Blackwell GPU (B200-SXM-180GB) | 16 | 5.8 | no EU price | — |
| Nebius | v5.1 | llama2_70b_lora | NVIDIA Blackwell GPU (B200-SXM-180GB) | 32 | 3.0 | no EU price | — |
| Nebius | v5.0 | llama31_405b | NVIDIA H200-SXM5-141GB | 512 | 209.7 | 4.50 | 8,051 |
| Nebius | v5.1 | llama31_405b | NVIDIA H200-SXM5-141GB | 512 | 209.7 | 4.50 | 8,051 |
| Nebius | v5.0 | llama31_405b | NVIDIA H200-SXM5-141GB | 1024 | 106.7 | 4.50 | 8,195 |
| Nebius | v5.1 | llama31_405b | NVIDIA H200-SXM5-141GB | 1024 | 106.7 | 4.50 | 8,195 |
| Nebius | v5.1 | llama31_8b | NVIDIA Blackwell GPU (B200-SXM-180GB) | 8 | 84.6 | no EU price | — |
| Nebius | v5.1 | llama31_8b | NVIDIA Blackwell Ultra GPU (B300-SXM-270GB) | 8 | 75.8 | no EU price | — |
| Nebius | v6.0 | llama31_8b | NVIDIA Blackwell Ultra GPU (B300-SXM-270GB) | 8 | 72.0 | no EU price | — |
| Nebius | v5.1 | llama31_8b | NVIDIA Blackwell GPU (B200-SXM-180GB) | 16 | 52.2 | no EU price | — |
| Nebius | v5.1 | llama31_8b | NVIDIA Blackwell GPU (B200-SXM-180GB) | 32 | 28.3 | no EU price | — |
| Nebius | v6.0 | llama31_8b | NVIDIA Blackwell Ultra GPU (B300-SXM-270GB) | 64 | 14.4 | no EU price | — |
| Oracle | v5.0 | bert | NVIDIA H200-SXM5-141GB | 64 | 0.9 | 10.00 | 9 |
| Oracle | v5.1 | flux1 | NVIDIA Blackwell GPU (B200-SXM-180GB) | 16 | 173.9 | 14.00 | 649 |
| Oracle | v5.1 | flux1 | NVIDIA Blackwell GPU (B200-SXM-180GB) | 72 | 55.4 | 14.00 | 931 |
| Oracle | v5.0 | llama2_70b_lora | NVIDIA Blackwell GPU (B200-SXM-180GB) | 8 | 11.0 | 14.00 | 20 |
| Oracle | v5.0 | llama2_70b_lora | NVIDIA H200-SXM5-141GB | 8 | 22.9 | 10.00 | 31 |
| Oracle | v5.0 | llama2_70b_lora | NVIDIA Blackwell GPU (B200-SXM-180GB) | 64 | 2.0 | 14.00 | 29 |
| Oracle | v5.0 | llama2_70b_lora | NVIDIA H200-SXM5-141GB | 512 | 1.9 | 10.00 | 166 |
| Oracle | v5.1 | llama31_8b | NVIDIA Blackwell GPU (B200-SXM-180GB) | 8 | 84.1 | 14.00 | 157 |
| Oracle | v5.1 | llama31_8b | NVIDIA Blackwell GPU (B200-SXM-180GB) | 64 | 18.7 | 14.00 | 279 |
| Oracle | v5.0 | retinanet | NVIDIA Blackwell GPU (B200-SXM-180GB) | 8 | 22.9 | 14.00 | 43 |
| Oracle | v5.0 | retinanet | NVIDIA H200-SXM5-141GB | 8 | 34.5 | 10.00 | 46 |
| Oracle | v5.0 | retinanet | NVIDIA Blackwell GPU (B200-SXM-180GB) | 64 | 4.4 | 14.00 | 66 |
| Oracle | v5.0 | retinanet | NVIDIA H200-SXM5-141GB | 64 | 7.4 | 10.00 | 79 |
| Oracle | v5.0 | stable_diffusion | NVIDIA Blackwell GPU (B200-SXM-180GB) | 8 | 13.9 | 14.00 | 26 |
| Oracle | v5.0 | stable_diffusion | NVIDIA Blackwell GPU (B200-SXM-180GB) | 64 | 2.8 | 14.00 | 42 |
These sellers are priced here and have never submitted to MLPerf Training. They appear as what they are. Nothing is estimated for them, and the absence is not evidence about their machines.
aws, datacrunch, genesis_cloud, hetzner, leaseweb, ovhcloud, runpod, scaleway, seeweb, vast.ai
No seller priced here has ever submitted an H100 system, which is the accelerator the headline index prices. Every figure above is therefore about a different chip from the one the headline measures.
A submitted system is a configuration its vendor
tuned for the benchmark, at a node count it chose, in a region it did not have to
name. It is not what a customer gets by default and it is a point in time. The price
it is multiplied by is an on-demand EU/EEA price for a matching accelerator on one
date, which is a different thing bought in a different shape. A single cell is a data
point and not a conclusion about either seller. Source:
MLPerf Training v5.0 at d7e703fccc · MLPerf Training v5.1 at 441d65f444 · MLPerf Training v6.0 at eabf23a07b.