GPU Compute Price Index
Renting an H100 costs $3.84 an hour on demand, a trimmed average of 24 providers’ posted prices, and $2.31 on spot. One index per GPU, for every way of buying it, from interruptible spot capacity to five-year commitments, built from the prices 34 providers publish.
Per GPU-hour, global, all providers, as of Oct 28, 17:05 UTC. Each index is the trimmed mean of providers’ prices, each provider counted once.
On-demand and spot, by GPU
Per GPU-hour, log scale. The line runs from the cheapest provider’s on-demand price to the dearest.
What the prices say
Spot capacity, which the provider can take back at short notice, costs 57% of the on-demand price for a B300 and 67% for an L40S, with the other GPUs in between.
The hyperscalers (AWS, Azure, Google Cloud and Oracle) charge the most. On demand, an H100 costs $10.68 an hour from them against $3.72 from neoclouds, 2.9 times as much; the multiple is 2.6 for L40S, 2.2 for B300 and 2.2 for RTX PRO 6000.
Committing for a year brings the price down to about 69% of on-demand, comparing each provider with its own on-demand price (H100 69%, H200 69%, A100 76% and RTX PRO 6000 69%). Few providers post term prices, so the longer commitments rest on a handful of quotes; most large term deals are negotiated privately.
Daily fixing, on-demand
Per GPU-hour at 16:00 UTC each day, log scale, from the first fixing on October 26.
Every GPU and term
Per GPU-hour, global, all providers, with the number of providers in each index. Commitments are grouped by length; a figure from one or two providers is that provider’s price, not a market.
| GPU | Spot | On-demand | 1M | 3M | 6M | 1Y | 2Y | 3Y | 5Y |
|---|---|---|---|---|---|---|---|---|---|
| H100H100 SXM 80GB | $2.3110 | $3.8424 | $5.026 | $4.073 | $3.853 | $5.336 | $3.521 | $4.973 | $5.751 |
| H200H200 SXM 141GB | $3.089 | $5.2620 | $6.145 | $4.313 | $4.083 | $5.636 | $4.601 | $4.653 | — |
| B200B200 HGX 180GB | $5.007 | $7.7516 | $9.336 | $7.793 | $6.793 | $9.644 | $6.651 | $7.142 | — |
| B300B300 HGX 288GB | $5.316 | $9.2411 | $13.582 | $10.901 | $10.791 | $10.343 | $8.431 | $10.041 | — |
| GB200GB200 NVL72 | $6.221 | $16.003 | $10.581 | — | — | $21.461 | — | $14.751 | — |
| A100A100 SXM 80GB | $1.377 | $2.2820 | $2.553 | $2.652 | $2.562 | $3.005 | $1.651 | $2.343 | — |
| MI300XMI300X 192GB | $1.391 | $4.166 | $3.851 | $3.661 | $3.471 | $3.083 | — | $3.291 | — |
| L40SL40S 48GB | $0.964 | $1.4315 | $1.831 | $1.811 | $1.791 | $2.052 | $1.401 | $1.641 | — |
| RTX PRO 6000RTX PRO 6000 Blackwell (server) | $1.408 | $2.1719 | $2.132 | $2.072 | $2.012 | $2.754 | $1.941 | $2.292 | — |
Terms: Spot, spot / interruptible; On-demand, on-demand; 1M, commit ≤1 month; 3M, commit 1-3 months; 6M, commit 3-6 months; 1Y, commit 1 year; 2Y, commit 2 years; 3Y, commit 3 years; 5Y, commit 5 years. Download: gpu-index-latest.csv.
Price against on-demand
Each provider’s price for a term divided by its own on-demand price, the median across providers quoting both. This compares like with like: providers that post term prices tend to be cheaper on demand too, so the board above can show a term costing more than on-demand.
| GPU | Spot | 1M | 3M | 6M | 1Y | 2Y | 3Y | 5Y |
|---|---|---|---|---|---|---|---|---|
| H100 | 44%9 | 92%5 | 92%2 | 88%2 | 69%5 | 75%1 | 44%3 | 40%1 |
| H200 | 50%8 | 83%4 | 83%2 | 81%2 | 69%5 | 75%1 | 44%3 | — |
| B200 | 50%6 | 99%4 | 97%2 | 90%2 | 84%2 | 76%1 | 42%1 | — |
| B300 | 44%4 | 82%2 | 98%1 | 97%1 | 83%2 | 76%1 | 40%1 | — |
| GB200 | 19%1 | — | — | — | 64%1 | — | 44%1 | — |
| A100 | 51%7 | 79%2 | 97%1 | 96%1 | 76%4 | 75%1 | 47%3 | — |
| MI300X | 18%1 | — | — | — | 69%2 | — | 44%1 | — |
| L40S | 50%4 | 98%1 | 97%1 | 96%1 | 78%2 | 75%1 | 43%1 | — |
| RTX PRO 6000 | 51%8 | 98%1 | 97%1 | 96%1 | 69%3 | 75%1 | 44%2 | — |
By type of provider, on-demand
Per GPU-hour, with the number of providers. Hyperscalers: AWS, Azure, Google Cloud and Oracle. Neoclouds sell their own capacity at posted rates. Marketplaces: prices set by many independent hosts or by auction.
| GPU | All providers | Hyperscalers | Neoclouds | Marketplaces |
|---|---|---|---|---|
| H100 | $3.8424 | $10.684 | $3.7216 | $2.634 |
| H200 | $5.2620 | $10.304 | $4.8713 | $4.923 |
| B200 | $7.7516 | $15.542 | $7.7211 | $6.103 |
| B300 | $9.2411 | $20.052 | $9.197 | $7.282 |
| GB200 | $16.003 | $24.762 | $10.501 | — |
| A100 | $2.2820 | $4.644 | $2.1813 | $1.393 |
| MI300X | $4.166 | $6.752 | $3.324 | — |
| L40S | $1.4315 | $3.632 | $1.4011 | $0.782 |
| RTX PRO 6000 | $2.1719 | $4.503 | $2.0713 | $1.693 |
By region, on-demand
Per GPU-hour, from providers that say where their capacity is, with the like-for-like difference from North America below: each provider’s price in the region against its own North American price, the median across providers quoting both. Rate cards without a location count only in the global index.
| GPU | North America | Europe | Asia-Pacific | Middle East & Africa | Latin America |
|---|---|---|---|---|---|
| H100 | $6.0510 · base | $5.0411 · 0% | $5.276 · 0% | $10.003 · 0% | $10.782 · +34% |
| H200 | $6.4111 · base | $5.7912 · 0% | $7.028 · 0% | $9.754 · +22% | $13.212 · +17% |
| B200 | $7.8010 · base | $9.826 · 0% | $9.635 · 0% | $8.503 · 0% | $14.001 · 0% |
| B300 | $12.254 · base | $10.637 · 0% | $15.003 · 0% | $15.001 · 0% | $25.072 · +49% |
| GB200 | $16.003 · base | $16.003 · 0% | $26.252 · +17% | $33.152 · +43% | $35.042 · +50% |
| A100 | $2.7012 · base | $3.009 · 0% | $3.267 · +12% | $4.224 · 0% | $6.102 · +33% |
| MI300X | $6.003 · base | $6.832 · +3% | $6.972 · +5% | $7.292 · +10% | $9.002 · +33% |
| L40S | $1.677 · base | $1.939 · 0% | $3.503 · +16% | $3.832 · +11% | $3.501 · 0% |
| RTX PRO 6000 | $2.3513 · base | $2.769 · 0% | $3.046 · 0% | $4.501 · 0% | $4.501 · 0% |
Method
What is priced. US dollars per GPU-hour, before tax, for 9 reference GPUs (H100 SXM 80GB, H200 SXM 141GB, B200 HGX 180GB, B300 HGX 288GB, GB200 NVL72, A100 SXM 80GB, MI300X 192GB, L40S 48GB and RTX PRO 6000 Blackwell (server)). A price quoted per instance or node is divided by its GPUs. Each index covers one GPU, one way of buying it (spot or interruptible, on-demand, or a commitment grouped by length, from up to a month to five years), one region (global or one of five) and one type of provider.
Sources. 32 sources: providers’ own price lists and rate cards, floating spot markets and auctions, marketplace asks and catalogs of posted prices, collected six times a day (03:07 to 23:07 UTC, every four hours). 35 of 35 answered in the latest run. A source that fails keeps its last prices for up to 36 hours. These are offered prices, not transactions; large term contracts are negotiated privately and usually below posted rates.
Index. Each provider counts once, at the median of its prices in scope (its regions, instance sizes and configurations). The index is the 20% trimmed mean of the provider prices: with five or more providers the lowest and highest fifth are dropped and the rest averaged; with fewer, the median. Providers are equally weighted, since capacity isn’t public. Listing placeholders (under 15% of the GPU’s on-demand median) and outliers (more than four times away from the median of comparable providers, those of the same type when there are three or more) are set aside, and a spot price at or above the same provider’s on-demand price is dropped.
Fixing. One fixing a day at 16:00 UTC, from each price in force at that time; a final fixing stands as published. Between fixings this page shows the same calculation on the latest prices. Seven- and thirty-day moving averages of the fixings appear once they have five and twenty days. The first fixing is October 26.
Updates. Prices are collected and this page rebuilt six times a day; the time at the top is the latest collection. Downloads: gpu-index-latest.csv, the latest print of every GPU and term; gpu-index-daily.csv, every daily fixing, by GPU, term, region and type of provider.
Source: providers’ published prices. Indexes are 06.ai calculations. Not investment advice. See also GPU spot prices, Nebius’s spot market every 15 minutes.