Price spread: the same GPU, different providers

The Price Index says what an H100-hour costs. This page says how much that depends on where you rent it. For every card listed by at least 3 providers: the cheapest, median and dearest listed on-demand rate per GPU-hour, each provider counted once at its own cheapest region. Marketplace floors are excluded — a floor is one machine’s ask, not a rate card.

As of September 12, 2026, across the 15 GPUs listed by 3 or more providers, the dearest listed on-demand rate is a median 3.4× the cheapest one for the same card. The widest is the NVIDIA V100 at 6.8×: $0.17 on Verda against $1.15 on Paperspace. Where both kinds of provider list a card, the hyperscaler median runs a median 128% above the specialist-cloud median.

Card by card

• NVIDIA V100: $0.17 on Verda to $1.15 on Paperspace (Paperspace) per GPU-hour, a factor of 6.8 across 4 providers; the median is $0.51, and the cheapest rate sits 67% below it.
• NVIDIA H100 SXM: $1.79 on Cudo Compute (no-kristiansand-1, GPU only) to $11.06 on Microsoft Azure (eastus) per GPU-hour, a factor of 6.2 across 17 providers; the median is $3.95, and the cheapest rate sits 55% below it, with hyperscalers a median 170% above specialist clouds.
• NVIDIA L40S: $0.76 on Sesterce (from) to $3.50 on Oracle Cloud (OCI) per GPU-hour, a factor of 4.6 across 12 providers; the median is $1.69, and the cheapest rate sits 55% below it, with hyperscalers a median 66% above specialist clouds.
• NVIDIA H200: $2.48 on Sesterce (from) to $10.60 on Microsoft Azure (westus2) per GPU-hour, a factor of 4.3 across 16 providers; the median is $4.52, and the cheapest rate sits 45% below it, with hyperscalers a median 128% above specialist clouds.
• NVIDIA B200: $4.00 on GMI Cloud (from) to $16.11 on Google Cloud (us-west1, GPU only) per GPU-hour, a factor of 4.0 across 13 providers; the median is $6.69, and the cheapest rate sits 40% below it, with hyperscalers a median 130% above specialist clouds.
• NVIDIA A100 40GB: $0.89 on Jarvislabs to $3.40 on Microsoft Azure (westus2) per GPU-hour, a factor of 3.8 across 9 providers; the median is $2.10, and the cheapest rate sits 58% below it, with hyperscalers a median 79% above specialist clouds.
• NVIDIA GB200 NVL72: $8.00 on GMI Cloud (from) to $27.04 on Microsoft Azure (westus3) per GPU-hour, a factor of 3.4 across 4 providers; the median is $13.25, and the cheapest rate sits 40% below it, with hyperscalers a median 133% above specialist clouds.
• NVIDIA A100 80GB: $1.19 on RunPod (Community Cloud) to $4.00 on Oracle Cloud (OCI) per GPU-hour, a factor of 3.4 across 14 providers; the median is $2.75, and the cheapest rate sits 57% below it, with hyperscalers a median 77% above specialist clouds.
• NVIDIA RTX A6000: $0.33 on RunPod (Community Cloud) to $1.09 on Lambda per GPU-hour, a factor of 3.3 across 5 providers; the median is $0.61, and the cheapest rate sits 46% below it.
• NVIDIA L4: $0.44 on Jarvislabs to $1.00 on OVHcloud (US) per GPU-hour, a factor of 2.3 across 5 providers; the median is $0.80, and the cheapest rate sits 45% below it.
• NVIDIA B300: $6.94 on RunPod (Community Cloud) to $15.00 on Oracle Cloud (OCI) per GPU-hour, a factor of 2.2 across 9 providers; the median is $7.85, and the cheapest rate sits 12% below it.
• NVIDIA RTX 6000 Ada: $0.74 on RunPod (Community Cloud) to $1.57 on Paperspace (GPU Droplets) per GPU-hour, a factor of 2.1 across 3 providers; the median is $1.04, and the cheapest rate sits 29% below it.
• NVIDIA L40: $0.69 on RunPod (Community Cloud) to $1.25 on CoreWeave per GPU-hour, a factor of 1.8 across 3 providers; the median is $1.00, and the cheapest rate sits 31% below it.
• NVIDIA RTX PRO 6000 Blackwell: $1.69 on RunPod (Community Cloud) to $3.03 on Modal (GPU only) per GPU-hour, a factor of 1.8 across 7 providers; the median is $1.89, and the cheapest rate sits 11% below it.
• NVIDIA H100 PCIe: $1.99 on RunPod (Community Cloud) to $3.32 on Scaleway (fr-par-2) per GPU-hour, a factor of 1.7 across 5 providers; the median is $3.25, and the cheapest rate sits 39% below it.

How to read it

A wide spread is not automatically an arbitrage: the cheapest row is often a community tier, a GPU-only line item where CPU and memory are billed separately, or a single region, and each of those is tagged in the table. What the spread does show is that “the price of an H100-hour” is not one number, and that a quoted market rate means little without the provider next to it. The median in this table counts each provider once, at its own cheapest region, so a provider with many regions cannot weigh it down.

Cheapest rate per card by provider type: /cheapest · the level over time: Price Index · every quote behind this page: prices.json.

Cite this page

Published under CC BY 4.0: use and redistribute with attribution and a link. Cite the linked source for a price itself, and cite us for the normalization, the index and the comparison.

Data updated: September 12, 2026