The H100 rental price depends on the provider and on how you rent. On the providers’ own pricing pages, checked October 2, 2026, one Nvidia H100 ranged from about $2.46 per GPU-hour on CoreWeave’s spot market to $11.06 per GPU-hour on demand on Google Cloud, with Lambda and Nebius listing on-demand H100s at $3.99 and $4.50. These prices change often; the full, regularly re-checked list is on our cloud GPU price tracker.

On this page

Per instance or per GPU

Clouds publish H100 prices in two ways. Some list a price for a single GPU: Lambda shows its H100 SXM at $3.99 “/GPU/HR”, and Nebius lists the HGX H100 per GPU-hour. Others sell whole machines, usually eight H100s with CPUs, memory and local storage, and quote the machine’s hourly price. AWS, Google Cloud and CoreWeave do this. To compare the two, divide the instance price by the number of GPUs in it. CoreWeave’s 8-GPU HGX H100 instance lists at $49.24 an hour in North America, which works out to $6.155 per GPU-hour.

The division hides what else is in the box, and that differs by provider:

  • Google Cloud’s a3-highgpu-8g (8 H100s) comes with 208 vCPUs, 1,871 GB of memory and 6,000 GiB of Local SSD, all included in its hourly price.
  • CoreWeave’s HGX H100 instance includes 128 vCPUs, 2,048 GB of system RAM and 61.44 TB of local storage.
  • Lambda lists its on-demand H100 SXM configuration with 208 vCPUs, 1,800 GiB of RAM and 22 TiB of SSD.

A higher per-GPU figure can include more CPU, memory and disk. When you compare offers, compare the whole configuration, not the GPU line alone.

It also matters which H100 you get. Nvidia sells the H100 SXM with 80 GB of memory and 3.35 TB/s of memory bandwidth, and the H100 NVL with 94 GB and 3.9 TB/s. Rental listings usually name the form factor (SXM, PCIe, HGX); our AI chip catalog has the specs.

On-demand, spot and reserved capacity

The same GPU costs different amounts depending on the commitment.

On-demand is pay-as-you-go: you start and stop instances and pay by the hour at the list price. All of the per-GPU figures from Lambda, Nebius and Google Cloud in the table below are on-demand.

Spot is spare capacity sold at a discount, which the provider can take back when it needs it. CoreWeave lists its HGX H100 instance at $19.71 an hour on spot against $49.24 on demand, or $2.46 versus $6.155 per GPU-hour.

Reserved and committed capacity trades flexibility for a lower or guaranteed rate. CoreWeave says committed usage brings “up to 60% discounts” on its on-demand prices. Google Cloud offers committed use discounts on its A3 machines; it says these machine types are not eligible for sustained use discounts or flexible committed use discounts.

AWS Capacity Blocks are a reservation for a fixed future window. AWS says the price “consists of a reservation fee and an operating system fee,” and the reservation fee “is charged up front at the time you schedule the reservation.” For the p5.48xlarge (8 H100s) in US East (N. Virginia), AWS lists $41.528 per instance-hour, or $5.191 per accelerator-hour. AWS updates these rates periodically based on supply and demand, and its pricing page states that effective October 7, 2026, the per-accelerator hourly rate for P5 across all available regions will be $5.970.

H100 prices on the providers’ pages

All figures below were read on each provider’s own pricing page on October 2, 2026. Per-GPU-hour values for instance prices are the instance price divided by eight.

Provider Offer Billing Listed price Per GPU-hour
CoreWeave HGX H100, 8 GPUs, North America Spot $19.71 per instance-hour $2.46
Lambda H100 SXM 80 GB On-demand $3.99 per GPU-hour $3.99
Nebius HGX H100 On-demand $4.50 per GPU-hour (effective October 1, 2026) $4.50
AWS p5.48xlarge Capacity Block, US East (N. Virginia) Reserved, paid up front $41.528 per instance-hour $5.191
CoreWeave HGX H100, 8 GPUs, North America On-demand $49.24 per instance-hour $6.155
Google Cloud a3-highgpu-8g, Iowa (us-central1) On-demand $88.49 per instance-hour $11.06

Prices move in both directions and sometimes within days. Nebius’s table shows the HGX H100 at $3.85 per GPU-hour in its previous on-demand column and $4.50 in the column effective October 1, 2026. AWS’s P5 Capacity Block rate changes on October 7. Treat the table as a dated snapshot and check /gpu-prices for the current figures, which also cover Azure, RunPod, DigitalOcean, Crusoe, Hyperstack and other providers.

What the hourly price leaves out

A GPU-hour price is rarely the whole bill.

  • Tax. Lambda states its prices are “plus applicable sales tax/VAT/GST.” Most list prices are quoted before tax.
  • Data transfer. CoreWeave says it charges “no ingress, egress, or transfer fees,” and Lambda advertises no egress fees. On clouds that do charge for data transfer, the charge is billed separately from the GPU-hour.
  • Storage. Local SSD is part of several of the instance prices above, but persistent or network storage for datasets and checkpoints is priced separately.
  • Operating system and software. AWS bills the Capacity Block operating system fee separately from the reservation fee, and lists different OS fees for Linux, RHEL, SLES and Ubuntu Pro.
  • Region. Hyperscaler prices are per region. The AWS and Google Cloud figures above are for US East (N. Virginia) and Iowa respectively; other regions list different rates.

How to compare H100 rental offers

Start with the per-GPU-hour figure, then check four things: what is bundled with the GPU (vCPUs, RAM, local disk), which H100 variant it is, how long you must commit and whether you pay in advance, and what transfer, storage and tax charges apply on top. A spot price is the cheapest line in the table but suits only jobs that can stop and resume. A Capacity Block guarantees GPUs for a set window but is paid up front.

The GPU price tracker lists each offer with its billing type, region, how the per-GPU figure was calculated, its source page and the date it was checked. Purchase prices for H100s are not covered, because Nvidia does not publish them.

Sources