#How the data was collected
We read the public pricing pages and APIs of 40 GPU clouds and dedicated hosts on 23 September 2026 and kept one on-demand price per provider and per GPU, expressed per GPU-hour. Where a provider publishes an 8-GPU node, we used its per-GPU price. For GeForce cards, which are mostly rented by the month, we used each host’s cheapest monthly server.
Three groups stay out of the medians and are shown separately: hyperscalers (AWS, Google Cloud and Azure), marketplaces together with spot and interruptible offers, and prices that require a commitment or a sales quote. Including the hyperscalers would have pulled every median up.
Keeping one price per provider stops a provider with many listings from weighing more than the others. Using the median rather than the average stops a single very cheap or very expensive listing from moving the reference.
#Median prices per GPU
| GPU | Providers | Lowest | Median | Highest | Highest ÷ lowest |
|---|---|---|---|---|---|
| B300 | 8 | $6.60 | $7.87 | $8.60 | 1.30 |
| B200 | 10 | $5.43 | $6.74 | $8.60 | 1.58 |
| H200 | 12 | $3.49 | $4.465 | $6.31 | 1.81 |
| H100 SXM | 15 | $2.69 | $3.49 | $6.16 | 2.29 |
| RTX PRO 6000 | 9 | $1.335 | $1.89 | $3.00 | 2.25 |
| L40S | 10 | $0.646 | $1.56 | $2.25 | 3.48 |
| RTX 5090, per month | 15 | $384 | $481 | $999 | 2.60 |
| RTX 4090, per month | 16 | $283 | $387 | $888 | 3.14 |
Prices are per GPU-hour, except the two GeForce rows, which are per GPU per month.
Two GPUs are missing on purpose. No provider publishes an on-demand price for the MI355X; the only public figures are two starting prices, $2.59 and $2.95 per GPU-hour. The RTX 5080 has only two public prices, $285 and $313 a month. Both are too thin for a real median, so we treat them as indicative.
#How far apart providers are
Within the HBM line, the spread grows with the age of the GPU: the most expensive listing is 1.30 times the cheapest for the B300, 1.58 times for the B200, 1.81 times for the H200 and 2.29 times for the H100. The median sits 19% above the cheapest listing for the B300, 24% for the B200, 28% for the H200 and 30% for the H100.
PCIe cards spread wider still: 2.25 times for the RTX PRO 6000 and 3.48 times for the L40S, where the cheapest offer comes from a dedicated host and the median is 2.4 times that price. The GeForce market spreads 2.60 times for the RTX 5090 and 3.14 times for the RTX 4090.
Over a month the gaps become large. An 8-GPU H100 node running for 730 hours costs $15,709.60 at the cheapest listing, $20,381.60 at the median and $35,974.40 at the most expensive. For an 8-GPU B300 node the same month ranges from $38,544 to $50,224, with the median at $45,960.80.
In practice, providers of the newest HBM GPUs are priced close together, while for older and PCIe cards comparing offers pays more.
#Hyperscalers and marketplaces
Against the medians above, the hyperscalers’ on-demand prices run from 1.8 to 3.5 times higher per GPU-hour:
| Instance | GPU | $/GPU-hour | Times the median |
|---|---|---|---|
| AWS p5 | H100 | $6.88 | 1.97 |
| Google Cloud a3-highgpu-8g | H100 | $11.07 | 3.17 |
| Azure ND H100 v5 | H100 | $12.29 | 3.52 |
| AWS p5en | H200 | $7.91 | 1.77 |
| AWS p6-b200 | B200 | $14.24 | 2.11 |
| AWS p6-b300 | B300 | $17.80 | 2.26 |
Marketplaces sit on both sides of the median. On Vast.ai the median H100 was $2.27 per GPU-hour, 35% below the on-demand median and below the cheapest on-demand listing. The median H200, at $4.34, was within 3% of the on-demand median. The median B300, at $11.24, was 43% above it and higher than any on-demand listing: for the newest GPU, the marketplace was the expensive option. Marketplace capacity comes from independent hosts, and interruptible instances can be stopped when outbid, which is why we keep it separate.
#Where prices are heading
The latest moves are upward. Nebius raises its on-demand prices for the H100, H200, B200 and B300 by 17 to 21% on 1 October 2026. That is after our survey date, so the increase is not in these medians yet.
GeForce supply is tight. When we checked, one host had most of its RTX 4090 and RTX 5090 servers rented out, and another sold several configurations on pre-order only.
These are list prices on a single day: a snapshot, not a forecast, which is why every figure here carries its date.
#What you cannot rent yet
- NVIDIA Rubin (VR200) is not publicly rentable. Providers take reservations, with deliveries planned across 2026 and 2027. The B300 remains the most powerful GPU you can rent on demand, from eight providers.
- AMD MI455X is not rentable on demand either: capacity is sold through pre-orders and waitlists, and AMD says the Helios racks built on it are on target for the second half of 2026.
- AMD MI355X is available, but without a public on-demand price: offers come as quotes or starting prices.
That is why the B300 and the MI355X are the flagships of our catalogue.
#How the data sets our prices
Our rule is mechanical. For each GPU we take the median per-GPU hourly price, multiply it by 730 hours to get a monthly market median, and price 30% below it, rounded down. The same per-GPU price applies at every server size. For 9 of our 10 GPUs the result is below the lowest on-demand list price we found; the exception is the L40S, where one dedicated host is cheaper.
| GPU | Market median per month | Our price per month | Hourly equivalent (÷ 730) |
|---|---|---|---|
| B300 | $5,745 | $4,019 | $5.51 |
| B200 | $4,920 | $3,439 | $4.71 |
| H200 | $3,259 | $2,279 | $3.12 |
| H100 | $2,548 | $1,779 | $2.44 |
| MI355X | $2,022 (indicative) | $1,409 | $1.93 |
| RTX PRO 6000 | $1,380 | $959 | $1.31 |
| L40S | $1,139 | $789 | $1.08 |
| RTX 5090 | $481 | $329 | $0.45 |
| RTX 4090 | $387 | $269 | $0.37 |
| RTX 5080 | $299 (indicative) | $199 | $0.27 |
The same figures are available as JSON from our public pricing API, GET https://cryptgpu.com/api/v1/gpus, with no key needed. The break-even against hourly billing is worked out in monthly vs hourly GPU rental, and the comparison page puts every GPU side by side. When you know what you need, configure a server.