Compute
What a GPU costs per hour.
The raw material behind every token: renting one GPU for one hour across clouds, GPU clouds and marketplaces. Same rules as the token index: every number comes from a stored snapshot, and records are added, never edited.
GPU-hour reference pricestrialMedian on-demand price for one GPU for one hour across providers. How it is built
| GPU | Memory | On demand, from | On demand, up to | Lowest of any kind | Marketplace median | Providers |
|---|---|---|---|---|---|---|
| NVIDIA B300Blackwell Ultra generation, data centre | 268 GB | 9.99Together AI | 17.80AWS | 4.588CoreWeave · Spot | — | 3 |
| NVIDIA B200Blackwell generation, data centre | 180 GB | 5.98RunPod Community Cloud | 16.11Google Cloud | 4.09Together AI · Spot | 10.6318 asks | 8 |
| NVIDIA H200Hopper generation, data centre | 141 GB | 3.59RunPod Community Cloud | 10.60Microsoft Azure | 2.58CoreWeave · Spot | 4.84118 asks | 8 |
| NVIDIA H100Hopper generation, data centre | 80 GB | 1.99RunPod Community Cloud | 12.29Microsoft Azure | 1.418Microsoft Azure · Spot | 1.86821 asks | 9 |
| NVIDIA A100 80GBAmpere generation, data centre | 80 GB | 1.19RunPod Community Cloud | 5.028Google Cloud | 0.679Microsoft Azure · Spot | 0.84435 asks | 7 |
| NVIDIA L40SData centre, inference and graphics | 48 GB | 0.79RunPod Community Cloud | 3.766AWS | 0.774Vast.ai · Market (median ask) | 0.77434 asks | 5 |
| NVIDIA RTX 5090Consumer card, rented on GPU clouds and marketplaces | 32 GB | 0.69RunPod Community Cloud | 1.39RunPod Secure Cloud | 0.584Vast.ai · Market (median ask) | 0.58464 asks | 3 |
| NVIDIA RTX 4090Consumer card, rented on GPU clouds and marketplaces | 24 GB | 0.34RunPod Community Cloud | 0.89RunPod Secure Cloud | 0.34RunPod Community Cloud · On demand | 0.50164 asks | 3 |
| NVIDIA V100 32GBVolta generation, data centre | 32 GB | 3.901AWS | — | 0.39AWS · Spot | — | 1 |
| NVIDIA T4Turing generation, inference | 16 GB | 0.526AWS | 0.978AWS | 0.209AWS · Spot | — | 1 |
In mainland China, in CNY
China’s clouds sell a different shelf: the newest NVIDIA cards cannot be exported there, so the same money buys older models. These are list prices in yuan, shown as published and never converted, so they stand next to the dollar prices rather than among them.
| GPU | Memory | On demand, from | On demand, up to | Stock at the lowest price |
|---|---|---|---|---|
| NVIDIA L20Data centre, a model made for the China market | 48 GB | ¥14.41Alibaba Cloud · ebmgn8is | ¥23.94 | none in stock |
| NVIDIA A10Ampere generation, inference and graphics | 24 GB | ¥9.53Alibaba Cloud · gn7i | ¥13.31 | 2 of 4 zones |
| NVIDIA V100 32GBVolta generation, data centre | 32 GB | ¥19.74Alibaba Cloud · gn6e | — | 5 of 5 zones |
| NVIDIA V100Volta generation, data centre | 16 GB | ¥26.46Alibaba Cloud · gn6v | — | none in stock |
| NVIDIA T4Turing generation, inference | 16 GB | ¥11.63Alibaba Cloud · gn6i | ¥17.19 | 2 of 6 zones |
Same GPU, different providers
How to read this
Each dot is one provider at its cheapest on-demand price for one GPU-hour. The line runs from zero to the dearest provider for that GPU, so the gaps between dots are to scale within a row.
We are building an exchange for compute. What it will be, and the waitlist →
How to read these prices
- One GPU, one hour
- Every price is in US dollars per GPU-hour. When a cloud sells whole machines, we divide the price by the number of GPUs, so that figure includes the bundled CPUs, memory and disk. A GPU cloud or marketplace usually pairs the GPU with a smaller machine. The two are comparable, not identical.
- Kinds of price
- On demand: start now, stop any time. Spot: cheaper, but the provider can take the machine back. Reserved: a commitment of one or three years, shown per hour. Market: the median of what hosts on a marketplace were asking when we looked; it moves all the time and hosts differ in reliability.
- Where it comes from
- Public price lists and APIs, fetched once a day and stored with a hash. We do not buy or measure anything: these are asking prices, not invoices, and we do not check availability.
- What is not here yet
- Google Cloud, Together and the Chinese clouds are being added. Contract prices and negotiated discounts are private and not shown.
The whole dataset as JSON. Corrections: hello@auramatrix.ai