GPU rental price index: what did a GPU-hour cost in October 2026?

What are the key figures of the October 2026 index?

One sentence per GPU, each true on 29 October 2026, according to GPUwithCrypto's price check:

• The median on-demand H100 80GB rented for $4.16 per GPU-hour across 16 providers, from $2.73 to $12.29.
• The median on-demand H200 141GB rented for $5.29 per GPU-hour across 13 providers, from $3.62 to $10.60.
• The median on-demand B200 rented for $8.50 per GPU-hour across 9 providers, from $5.43 to $14.24.
• The median on-demand A100 80GB rented for $2.23 per GPU-hour across 13 providers, from $1.09 to $5.07.
• The median on-demand L40S rented for $1.67 per GPU-hour across 11 providers, from $0.97 to $3.15.
• The median on-demand RTX 4090 rented for $0.61 per GPU-hour across 2 providers, from $0.33 to $0.89.
• The median on-demand RTX 5090 rented for $0.84 per GPU-hour across 3 providers, from $0.50 to $1.19.

The spread is wide. The dearest on-demand H100, an 8-GPU Azure VM, cost 4.5 times the cheapest, Massed Compute's single H100.

How has the GPU rental price changed over time?

This index cannot say yet. October 2026 is its first edition, and it does not borrow history from other sources, because their rules differ. The November 2026 edition will show the first month-on-month change, with the same method.

One change is already announced. Salad raises its High Priority price on 1 November 2026: the RTX 4090 from $0.33 to $0.44 per hour, and the RTX 5090 from $0.50 to $0.60. Both were the lowest on-demand prices for those GPUs in this edition.

How do AWS, Google Cloud and Azure change the median?

They pull it up. Their GPU prices include the whole VM (vCPUs, RAM and local SSD), and their H100, H200 and B200 machines are mostly 8-GPU VMs. Without the three hyperscalers, the medians on 29 October 2026 were:

• H100 80GB: $3.49 across 13 providers, against $4.16 with them.
• H200 141GB: $4.80 across 11, against $5.29.
• B200: $8.25 across 8, against $8.50.
• A100 80GB: $2.00 across 11, against $2.23.
• L40S: $1.62 across 10, against $1.67.

The index keeps them in: they are self-serve on-demand list prices like the others.

How is the index calculated?

The same rule for every GPU and every provider:

1. Sources. Each price was read on 29 October 2026 on the provider's own pricing page, or on the public price feed or API that page loads. No aggregator is used.
2. One price per provider. The provider's lowest self-serve on-demand price for that GPU, across all its configurations. The H100 groups SXM, PCIe, NVL and unstated form factors; the H200 groups SXM and NVL.
3. Per GPU. A price listed only for a multi-GPU instance is divided by its GPU count: CoreWeave, Google Cloud, AWS and Azure 8-GPU machines, OVHcloud's 8× H200 and Massed Compute's 8× B200.
4. Hyperscalers keep the whole VM in the price, as they sell it.
5. Statistics. The median of an even count is the mean of the two middle prices. The 25th percentile uses linear interpolation between ranks, the method of a spreadsheet's PERCENTILE.INC.
6. Exclusions. Spot, preemptible, community and marketplace prices; contact-sales prices; stale pages (TensorDock's page dated July 2024, Spheron's RTX 4090 marked "Last known Sep 17"); Voltage Park's "starting at $1.99" H100, because its pricing cards say "Contact for pricing".
7. Vast.ai is excluded. Its public price feed allows display with attribution, but its licence asks for a data licence before use in an index or benchmark.

Fluidstack, Genesis Cloud, io.net and Prime Intellect had no readable public price list on 29 October 2026, and Hetzner listed no GPU from the index. Paperspace and DigitalOcean count as one provider.

When is the next edition?

The index is planned monthly. Each edition keeps its own check date, so a figure quoted from this page stays tied to 29 October 2026. To quote it, give the date and link to this page. Per-GPU detail, with spot and marketplace prices listed apart, is on the GPU price pages and the cheapest GPU cloud page.