As of the week of October 12, 2026, the median on-demand price across providers is $3.50 per GPU-hour for an H100 and $4.88 for an H200 — up from $2.99 and $3.50 a year earlier: +17% and +39% in twelve months, over a period in which chip performance per dollar has been improving ~49% a year (Epoch AI). Lambda's own list price for an H100 SXM is $3.99/GPU-hour.
| week | h100 usd per gpu hr median | h200 usd per gpu hr median |
|---|---|---|
| 2025-10-13 | 2.985 | 3.5 |
| 2026-01-05 | 2.99 | 3.495 |
| 2026-04-06 | 2.926 | 3.5 |
| 2026-07-06 | 3.317 | 4.29 |
| 2026-10-12 | 3.5 | 4.875 |
Download: CSV · JSON — compiled data CC BY 4.0; underlying sources keep their own terms.
Sampled weekly on-demand provider medians (five sampled weeks across the year) from GetDeploying's GPU price dataset (CC BY 4.0, credit GetDeploying); the series starts Oct 13, 2025. A provider median can move when the provider mix changes, and on-demand is the most expensive way to buy compute — reserved and contract prices are lower and not covered. The Epoch performance-per-dollar trend is a multi-year hardware trend, not a measurement over this exact window; the contrast between the two is the point, with that caveat attached.
Part of BRAIAIN — the AI buildout, measured. Every figure carries its source and date; the full rules are in the methodology. Corrections: see the contact note there. © 2026 BRAIAIN, a dnilgis project.