Every month we check the public on-demand prices of specialist GPU clouds to set our own: CoreGPU prices every GPU at least 25% below the market median. This is the data behind that number, as listed on November 4, 2026.
Method
- One on-demand price per provider and GPU, taken from public pricing pages or price APIs. Prices for 8-GPU nodes are divided by eight.
- Specialist GPU clouds only for the median. Hyperscalers (AWS, Azure, Google Cloud, Oracle) are listed separately because their prices are 2.5 to 4.5 times higher.
- Monthly cost = hourly price × 730 hours. Prices exclude taxes.
Median prices by GPU
| GPU | Prices found | Lowest $/h | Median $/h | Median / month |
|---|---|---|---|---|
| NVIDIA B300 | 8 | 6.60 | 8.98 | $6,552 |
| NVIDIA B200 | 11 | 5.43 | 8.19 | $5,979 |
| NVIDIA H200 | 12 | 2.90 | 4.94 | $3,606 |
| NVIDIA H100 SXM | 16 | 1.99 | 3.72 | $2,716 |
| NVIDIA RTX PRO 6000 | 9 | 1.40 | 2.19 | $1,599 |
| NVIDIA L40S | 11 | 0.73 | 1.57 | $1,146 |
| AMD MI325X | 2 | 3.80 | 4.21 | $3,071 |
| AMD MI300X | 4 | 2.59 | 3.22 | $2,351 |
The lowest prices often come from marketplaces with variable host quality or from “starting at” prices that are rarely available. The median is a better picture of what a team actually pays for reliable capacity.
AMD CDNA 4 has thin public pricing
MI355X and MI350X are listed by very few providers, often without an on-demand price. For MI355X we use a floor of $4.50/h (a published spot price), consistent with cross-checks against MI300X pricing. For MI350X we scale the MI355X reference by relative FP8 throughput (0.92), which sits between published spot ($3.00/h) and 12-month reserved ($4.76/h) prices.
Hyperscalers
On-demand prices per GPU-hour at the largest clouds remain far above specialist clouds: H100 from $6.88 (8-GPU instance divided by eight) to more than $12 depending on the provider, and B300 from $15 to $17.80. Commitments of one to three years cut those prices by 30 to 56%.
What moved in 2026
- Prices are rising again for flagships. At least one large specialist cloud announced increases of 10 to 26% on H100, H200 and B300 from December 1.
- Clusters cost more than single instances. Multi-node offers are commonly 20 to 50% more per GPU than on-demand single servers.
- Listed is not available. Several providers list GPUs with no deployable region, or “contact sales” for every flagship.
- Fees moved to storage and idle time. Egress is mostly free; stopped-instance storage, public IPs and minimum billing periods are where extra cost appears.
Where CoreGPU sits
Our monthly price per GPU is at least 25% below the medians above, with no setup fee and no commitment beyond the month. See live prices on our pricing page or compare models in the GPU comparisons.
This index is updated monthly. Figures are public list prices observed on the date above and can change without notice.