by Inference Hub

H100 Rental Prices in 2026: The Cheapest Clouds per GPU-Hour

What it costs to rent an NVIDIA H100 in November 2026 on marketplaces, specialist GPU clouds and AWS, Google Cloud, Azure and Oracle, with monthly and 8-GPU node costs and A100, H200 and B200 comparisons.

gpuh100a100pricingcomparison

The cheapest H100 rental we track is $1.67 per GPU-hour for an H100 SXM on the Vast.ai marketplace. If you want a fixed price from a provider rather than a marketplace listing, Latitude.sh charges $1.68 and Voltage Park starts at $1.99. Most specialist GPU clouds charge $2.50–$4.90 on demand, and the hyperscalers charge $6.88–$12.29.

All prices below are on-demand USD per single GPU-hour, from our GPU price comparison and the providers’ own pricing pages, as of 17 November 2026. Marketplace prices move by the hour, so check the live listing before you commit.

How much does it cost to rent an H100?

Marketplaces

On a marketplace, independent hosts set their own prices. These are the cheapest H100s you can get, but quality, network and uptime depend on the host.

ProviderH100 variantPrice per GPU-hourNotes
Vast.aiSXMfrom $1.67Lowest listing was on an unverified host; verified hosts started around $2.33
TensorDockSXMfrom $2.25GPU only; vCPU, RAM and storage are extra
FluenceSXMfrom $2.71Container plan; VMs from $3.30
Prime Intellect80GB$3.00Cheapest offer was an 8-GPU node ($24/hr)
Runcrate80GBfrom $3.91Cheapest of 9 listings

Specialist GPU clouds

These providers run their own data centers or lease dedicated capacity, and the price is fixed rather than set by the host.

ProviderH100 variantPrice per GPU-hour
Latitude.sh80GB, bare metal, 1 GPU$1.68
Voltage ParkSXM, Ethernet, no contractfrom $1.99
HyperstackPCIe / NVLink / SXM$2.50 / $2.60 / $3.20
RunPodPCIe / NVL / SXM$2.89 / $3.19 / $4.39
Thunder ComputeSXM$3.20
JarvislabsSXM$3.49
LambdaPCIe / SXM$3.79 / $4.59–$4.89
CrusoeSXM$3.90 (rises to $4.50 on 1 December)
ModalSXM, serverless$3.95
DigitalOceanSXM$4.41
NebiusSXM$4.50
CoreWeaveHGX, 8 GPUs$6.16

Hyperscalers

AWS, Google Cloud, Azure and Oracle rent H100s mainly as 8-GPU instances. The per-GPU figure is the instance price divided by eight.

ProviderInstanceInstance per hourPer GPU-hour
AWSp5.48xlarge (us-east-1)$55.04$6.88
Oracle CloudBM.GPU.H100.8$80.00$10.00
Google Clouda3-highgpu-8g (us-central1)$88.49$11.06
Microsoft AzureND96isr H100 v5 (East US)$98.32$12.29

AWS also sells a single-GPU p5.4xlarge at the same $6.88. Even so, the cheapest hyperscaler costs about four times the cheapest marketplace listing, and 1.5–2.5 times what most specialist clouds charge.

H100 SXM vs PCIe: which one are you renting?

The SXM H100 is the version in 8-GPU HGX servers. Its GPUs connect over NVLink at up to 900 GB/s and it runs at a higher power limit, so it is faster for multi-GPU training. The PCIe card runs at lower power and talks to other GPUs over the PCIe bus. The H100 NVL is a PCIe card with 94GB of memory, sold in NVLink-bridged pairs.

Where a provider sells both, PCIe is cheaper: Hyperstack charges $2.50 vs $3.20, RunPod $2.89 vs $4.39, Lambda $3.79 vs $4.59. For single-GPU inference or fine-tuning, PCIe is usually the better deal. For training across several GPUs, pay for SXM. See the H100 price page for every variant we track.

What does an H100 cost per month?

Running 24/7 for a month (730 hours):

ProviderPer GPU-hourOne GPU per month8-GPU node per hour8-GPU node per month
Vast.ai (lowest listing)$1.67$1,219$13.36$9,753
Latitude.sh$1.68$1,226n/a (1-GPU server)n/a
Voltage Park$1.99$1,453$15.92$11,622
Hyperstack (SXM)$3.20$2,336$25.60$18,688
Lambda (8x HGX)$4.59$3,351$36.72$26,806
CoreWeave$6.16$4,497$49.24$35,945
AWS$6.88$5,022$55.04$40,179
Oracle Cloud$10.00$7,300$80.00$58,400
Google Cloud$11.06$8,075$88.49$64,598
Azure$12.29$8,972$98.32$71,774

On a marketplace, a full 8-GPU node at the lowest single-GPU price may not be available, so treat the Vast.ai node figure as a floor. Some hosts list only one or two GPUs per machine.

How do A100, H200 and B200 rental prices compare?

A100 80GB cloud pricing

The A100 is still the cheapest way to get 80GB of data-center GPU memory. At most providers it costs about half as much as their H100, sometimes less.

ProviderA100 80GB variantPer GPU-hour
Vast.aiSXM4 / PCIefrom $0.52 / $0.84
Thunder Compute80GB$1.09
HyperstackPCIe / NVLink / SXM$1.35 / $1.40 / $1.60
Jarvislabs80GB$1.49
RunPodPCIe / SXM$2.09 / $2.19
Lambda8x HGX$3.19
AWSp4de.24xlarge$3.43
Oracle CloudBM.GPU.A100-v2.8$4.00
AzureND96amsr A100 v4$4.10
Google Clouda2-ultragpu-1g$5.07

The 40GB A100 is cheaper again: $0.89 at Jarvislabs and $1.99 at Lambda. More on the A100 price page.

H200 and B200

The H200 has 141GB of memory, which lets one GPU serve larger models than an H100. The B200 has 180GB and is the current Blackwell option most clouds sell.

Provider typeH200 per GPU-hourB200 per GPU-hour
Cheapest specialist cloud$3.99 (Hyperstack, Oblivus)$6.00 (Hyperstack)
Typical specialist cloud$4.29–$5.99 (Crusoe, DigitalOcean, Modal, RunPod, Together AI)$6.25–$8.60 (Modal, Lambda, RunPod, Nebius, CoreWeave)
AWS$7.91 (p5en.48xlarge)$14.24 (p6-b200.48xlarge)
Google Cloud$10.60 (a3-ultragpu-8g)$8.06 (a4-highgpu-8g, Flex-start only)
Azure$10.60 (ND96isr H200 v5)n/a
Oracle Cloud$10.00$14.00

At specialist clouds, an H200 costs only $0.06–$1.10 per hour more than the same provider’s H100 SXM. If an H100’s 80GB forces you to split a model across two GPUs, one H200 is usually cheaper. Crusoe’s H200 rises from $4.29 to $5.40 on 1 December. Compare every chip side by side on the GPU comparison page.

Why do H100 prices vary so much?

On-demand vs spot vs reserved. Every price above is on-demand: you can start and stop at any time. Spot or preemptible capacity is much cheaper but can be taken away. Google Cloud’s H100 spot price works out at $6.59 per GPU-hour against $11.06 on demand, Azure’s at $2.27, and Nebius lists preemptible H100s from $0.79. Reserved capacity cuts the rate if you commit for a long term: AWS’s 3-year Savings Plan brings the H100 to $2.59, and Google’s 3-year commitment brings it to $4.86. Hyperstack’s H100 PCIe drops from $2.50 to $1.75 reserved.

Interconnect. Multi-node training needs InfiniBand or an equivalent fabric between servers. Low “from” prices are often for Ethernet-only machines. Voltage Park’s $1.99, for example, is Ethernet; its InfiniBand clusters start at 8 GPUs and are priced through sales. Hyperscaler 8-GPU instances include high-bandwidth networking, which accounts for part of their premium.

What’s included. Some prices cover the GPU only. TensorDock and Modal bill CPU, RAM and storage separately. Bare-metal and hyperscaler instances include CPUs, memory and local NVMe. Compare total instance cost, not just the GPU line.

Region and availability. Hyperscaler prices differ by region (the figures above are for us-east-1, us-central1 and East US), and the cheapest marketplace hosts may be anywhere. If data residency or latency matters, filter by location first.

Verified vs community hosts. On Vast.ai, the lowest H100 listing was on an unverified host. Verified data-center hosts cost more, but they are a better fit for production work.

Where should you rent an H100?

  • Short experiments and fine-tuning: a marketplace (Vast.ai, TensorDock) or a low-cost cloud (Latitude.sh, Voltage Park, Hyperstack) at $1.67–$3.20.
  • Multi-GPU training on 8-GPU nodes: specialist clouds such as Lambda, Nebius or CoreWeave, where 8x SXM nodes with fast networking cost about $27,000–$36,000 a month.
  • Already committed to a hyperscaler: AWS is the cheapest of the four on demand at $6.88. A 1- or 3-year commitment brings any of them much closer to specialist-cloud pricing.

For every provider and GPU we track, see the GPU price comparison.

Subscribe to our newsletter

News and updates, 1–2 emails per month. No spam.