GPU rental price index · USD per GPU-hour
The price of compute,
measured every hour.
- Lowest ask
- —
- Middle half
- —
- RunPod list
- —
- Per GB VRAM
- —
The board
Tap a row to chart it. Market = median asking price across available Vast.ai listings. List = RunPod Secure Cloud on-demand price.
| GPU | Market | Low | Spread | Listings | List | List vs market | 24h | 7d | 7 days |
|---|---|---|---|---|---|---|---|---|---|
| Loading prices… | |||||||||
Power
Compute is electricity with a markup. Hourly wholesale prices and load in the grids where AI data centres are being built, from each grid operator's own real-time data.
Methodology
What is measured. Every hour, two minutes past, the collector reads every available on-demand listing for each tracked GPU on the Vast.ai marketplace and the published on-demand prices on RunPod Secure Cloud.
Per-GPU pricing. Marketplace listings are often multi-GPU machines. Each listing's total hourly price (GPU plus the marketplace's default storage charge) is divided by its GPU count, then we take the lowest, the 25th percentile, the median and the 75th percentile across listings. Each listing counts once regardless of size.
What it is not. These are asking prices for capacity that is free right now. Rented machines drop out of the sample, so in a tight market the median can rise simply because the cheap GPUs are taken. Reserved contracts and hyperscaler pricing are not included.
Known limits. Anonymous marketplace searches return at most 64 listings, so for the most common consumer cards (RTX 4090, RTX 5090) we query by machine size and combine the results; in busy hours some size buckets can still hit the cap. 80GB A100 series exclude 40GB cards that share a marketplace name. MI300X currently has list prices only.
No backfill. The series starts the day the collector started. Nothing before that is estimated or reconstructed. Gaps are logged and left as gaps.
Power. Each hour (UTC, labelled by the hour's start) is the average of the grid operator's own intervals in that hour. Texas: ERCOT real-time Hub Average settlement point price (15-minute) and system demand (5-minute). Ontario: IESO's real-time Ontario Zonal Price in Canadian dollars and its total market load, which includes exports. Québec: Hydro-Québec total demand (15-minute); Québec has no wholesale spot market, so there is no price to track. PJM: hourly demand from the U.S. EIA-930 survey, which arrives a few hours late; PJM prices are not tracked yet. Each run re-reads the last few hours so late data fills in, and an hour is published once at least half its intervals have arrived. The power series starts on October 2, 2026.
Free API & CSV
The data is free to use under CC BY 4.0. Please credit “GridCompute GPU Price Index” with a link to gridcompute.ai.
GET /api/latestEvery GPU, current snapshot, 24h and 7d change
GET /api/history?gpu=h100-sxm&days=30Hourly series (daily averages beyond 21 days)
GET /api/export.csv?gpu=all&days=90Raw hourly rows as CSV
GET /api/gpusTracked GPU keys
GET /api/power/latestEach grid's latest hour, 24h average and range, 7-day series
GET /api/power/history?region=ercot&days=30Hourly price and load (ercot, pjm, ontario, quebec)
GET /api/power/export.csv?region=all&days=90Raw hourly power rows as CSV