Meta Index History
The class-balanced meta index over time, rebased to 100 at the start of the series. One vote per GPU class, normalized by MLPerf throughput, sourced from ~7 months of Vast.ai daily medians.
Daily USD per H100 SXM-equivalent compute-hour: each GPU's price divided by its MLPerf-derived throughput, so every card type lands on one comparable axis.
The class-balanced meta index over time, rebased to 100 at the start of the series. One vote per GPU class, normalized by MLPerf throughput, sourced from ~7 months of Vast.ai daily medians.
The public dataset is one MLPerf-normalized, cross-GPU price series with daily distribution bands.
Daily raw median values by GPU class.
The latest daily index statistics.
| Date | Observations | Min | P25 | Median | P75 | Max |
|---|---|---|---|---|---|---|
| 2026-08-04 | 101 | 1.0436 | 1.5833 | 2.1538 | 2.8625 | 4.5455 |
| 2026-06-02 | 24 | 1.8524 | 2.4000 | 2.9900 | 3.6400 | 4.8058 |
Raw variant prices remain visible below the single normalized headline index.
| Class | Observations | Min | P25 | Median | P75 | Max |
|---|---|---|---|---|---|---|
| RTX 3090 24GB | 12 | 0.1148 | 0.1274 | 0.1356 | 0.1386 | 0.5000 |
| RTX A5000 24GB | 2 | 0.1600 | 0.1875 | 0.2150 | 0.2425 | 0.2700 |
| V100 32GB | 2 | 0.1900 | 0.2000 | 0.2100 | 0.2200 | 0.2300 |
| RTX 4080 16GB | 2 | 0.2700 | 0.2725 | 0.2750 | 0.2775 | 0.2800 |
| RTX 4090 24GB | 12 | 0.2830 | 0.2939 | 0.3085 | 0.3156 | 0.6900 |
| RTX A6000 48GB | 4 | 0.3300 | 0.4042 | 0.4489 | 0.4842 | 0.5300 |
| L4 24GB | 1 | 0.3900 | 0.3900 | 0.3900 | 0.3900 | 0.3900 |
| A100 PCIe 80GB | 14 | 0.4289 | 0.6296 | 0.7489 | 1.1237 | 1.3900 |
| A100 SXM 80GB | 12 | 0.5615 | 0.6131 | 0.7019 | 0.8422 | 1.4900 |
| L40S 48GB | 7 | 0.6015 | 0.7635 | 0.8015 | 0.8958 | 1.0741 |
| A100 40GB | 2 | 1.0000 | 1.0250 | 1.0500 | 1.0750 | 1.1000 |
| H100 NVL 94GB | 4 | 1.4696 | 2.3099 | 2.6161 | 2.7792 | 3.1900 |
| H100 SXM 80GB | 12 | 1.7356 | 2.2647 | 2.5977 | 3.0105 | 4.2900 |
| H100 PCIe 80GB | 4 | 1.9000 | 1.9675 | 2.4400 | 2.9900 | 3.2900 |
| H100 NVLink 80GB | 1 | 1.9500 | 1.9500 | 1.9500 | 1.9500 | 1.9500 |
| MI300X 192GB | 1 | 2.3900 | 2.3900 | 2.3900 | 2.3900 | 2.3900 |
| GH200 144GB | 1 | 3.1900 | 3.1900 | 3.1900 | 3.1900 | 3.1900 |
| H200 SXM 141GB | 5 | 3.5900 | 3.7900 | 3.7900 | 4.0811 | 4.3900 |
| H100 HGX 80GB | 1 | 3.9000 | 3.9000 | 3.9000 | 3.9000 | 3.9000 |
| B200 180GB | 2 | 5.8900 | 5.9125 | 5.9350 | 5.9575 | 5.9800 |
MLPerf-derived throughput per GPU class (H100 SXM = 1.00×) converts each raw price into a comparable H100-equivalent compute-hour.
| Class | Perf vs H100 | Source | Raw median $/hr | Norm. $/H100-eq hr | Obs |
|---|---|---|---|---|---|
| RTX 3090 24GB | 0.11× | spec est. | 0.1356 | 1.2327 | 12 |
| RTX 4090 24GB | 0.20× | spec est. | 0.3085 | 1.5425 | 12 |
| V100 32GB | 0.12× | spec est. | 0.2100 | 1.7500 | 2 |
| A100 SXM 80GB | 0.40× | MLPerf | 0.7019 | 1.7546 | 12 |
| H100 NVLink 80GB | 1.00× | MLPerf | 1.9500 | 1.9500 | 1 |
| B200 180GB | 3.00× | MLPerf | 5.9350 | 1.9783 | 2 |
| A100 PCIe 80GB | 0.36× | MLPerf | 0.7489 | 2.0801 | 14 |
| RTX 4080 16GB | 0.13× | spec est. | 0.2750 | 2.1154 | 2 |
| RTX A5000 24GB | 0.10× | spec est. | 0.2150 | 2.1500 | 2 |
| MI300X 192GB | 1.00× | MLPerf | 2.3900 | 2.3900 | 1 |
| H100 NVL 94GB | 1.05× | MLPerf | 2.6161 | 2.4915 | 4 |
| H100 SXM 80GB | 1.00× | MLPerf | 2.5977 | 2.5977 | 12 |
| GH200 144GB | 1.20× | MLPerf | 3.1900 | 2.6583 | 1 |
| H200 SXM 141GB | 1.40× | MLPerf | 3.7900 | 2.7071 | 5 |
| RTX A6000 48GB | 0.16× | spec est. | 0.4489 | 2.8056 | 4 |
| L40S 48GB | 0.28× | MLPerf | 0.8015 | 2.8625 | 7 |
| A100 40GB | 0.34× | MLPerf | 1.0500 | 3.0882 | 2 |
| H100 PCIe 80GB | 0.78× | MLPerf | 2.4400 | 3.1282 | 4 |
| L4 24GB | 0.10× | spec est. | 0.3900 | 3.9000 | 1 |
| H100 HGX 80GB | 1.00× | MLPerf | 3.9000 | 3.9000 | 1 |
Each daily snapshot stores aggregate index statistics.
| Metric | Definition |
|---|---|
| Normalized index | Median USD per H100 SXM-equivalent compute-hour: each offer's price ÷ its MLPerf-derived perf factor, across all GPU classes. |
| Perf factor | Per-GPU throughput relative to H100 SXM (=1.00×), from MLPerf Inference where available, else estimated from FP16/BF16 specs. |
| Variant prices | Raw median USD per GPU-hour within each GPU class, before normalization. |
| Spread | Min, p25, p75, and max show the daily market range. |