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 |
|---|---|---|---|---|---|---|
| 2026-09-16 | 82 | 1.0573 | 2.0208 | 2.6039 | 3.2988 | 5.4156 |
| 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 | 0.1163 | 0.1430 | 0.1756 | 0.2061 | 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 A6000 48GB | 2 | 0.3300 | 0.3800 | 0.4300 | 0.4800 | 0.5300 |
| RTX 4090 24GB | 12 | 0.3400 | 0.3977 | 0.4575 | 0.4880 | 0.7400 |
| A100 SXM 80GB | 7 | 0.4711 | 0.7216 | 0.9259 | 1.2659 | 1.5900 |
| L4 24GB | 1 | 0.4900 | 0.4900 | 0.4900 | 0.4900 | 0.4900 |
| L40S 48GB | 4 | 0.7900 | 0.7986 | 0.8023 | 0.8748 | 1.0900 |
| A100 40GB | 2 | 1.0000 | 1.0250 | 1.0500 | 1.0750 | 1.1000 |
| A100 PCIe 80GB | 5 | 1.0624 | 1.1900 | 1.2900 | 1.3500 | 1.5900 |
| 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 |
| H100 SXM 80GB | 8 | 2.2500 | 2.6175 | 3.1154 | 3.8272 | 5.4156 |
| MI300X 192GB | 1 | 2.3900 | 2.3900 | 2.3900 | 2.3900 | 2.3900 |
| H100 NVL 94GB | 5 | 2.5900 | 2.7756 | 2.9089 | 3.0332 | 3.1900 |
| GH200 144GB | 1 | 3.1900 | 3.1900 | 3.1900 | 3.1900 | 3.1900 |
| H200 SXM 141GB | 8 | 3.2953 | 3.7400 | 4.1900 | 4.7398 | 5.0022 |
| H100 HGX 80GB | 1 | 3.9000 | 3.9000 | 3.9000 | 3.9000 | 3.9000 |
| B200 180GB | 2 | 5.9800 | 6.1825 | 6.3850 | 6.5875 | 6.7900 |
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. | 0.1756 | 1.5964 | 12 |
| V100 32GB | 0.12× | spec est. | 0.2100 | 1.7500 | 2 |
| H100 NVLink 80GB | 1.00× | MLPerf | 1.9500 | 1.9500 | 1 |
| RTX 4080 16GB | 0.13× | spec est. | 0.2750 | 2.1154 | 2 |
| B200 180GB | 3.00× | MLPerf | 6.3850 | 2.1283 | 2 |
| RTX A5000 24GB | 0.10× | spec est. | 0.2150 | 2.1500 | 2 |
| RTX 4090 24GB | 0.20× | spec est. | 0.4575 | 2.2875 | 12 |
| A100 SXM 80GB | 0.40× | MLPerf | 0.9259 | 2.3147 | 7 |
| MI300X 192GB | 1.00× | MLPerf | 2.3900 | 2.3900 | 1 |
| GH200 144GB | 1.20× | MLPerf | 3.1900 | 2.6583 | 1 |
| RTX A6000 48GB | 0.16× | spec est. | 0.4300 | 2.6875 | 2 |
| H100 NVL 94GB | 1.05× | MLPerf | 2.9089 | 2.7704 | 5 |
| L40S 48GB | 0.28× | MLPerf | 0.8023 | 2.8652 | 4 |
| H200 SXM 141GB | 1.40× | MLPerf | 4.1900 | 2.9929 | 8 |
| A100 40GB | 0.34× | MLPerf | 1.0500 | 3.0882 | 2 |
| H100 SXM 80GB | 1.00× | MLPerf | 3.1154 | 3.1154 | 8 |
| H100 PCIe 80GB | 0.78× | MLPerf | 2.4400 | 3.1282 | 4 |
| A100 PCIe 80GB | 0.36× | MLPerf | 1.2900 | 3.5833 | 5 |
| H100 HGX 80GB | 1.00× | MLPerf | 3.9000 | 3.9000 | 1 |
| L4 24GB | 0.10× | spec est. | 0.4900 | 4.9000 | 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. |