GPUPriceBook

Cheapest H100 rental in 2026: where to get an H100 for under $3/hour

By GPUPriceBook editorial · 2026-06-19

In short: In June 2026 the cheapest on-demand H100 is about $1.90/hr (PCIe) on Hyperstack and ~$2.30/hr (SXM) on Vultr, with RunPod Community and Vast.ai interruptible dipping under $2. Hyperscalers (AWS ~$6.88, GCP ~$10.98, Azure ~$12.29 per GPU) cost 3-5x more because they sell H100s as whole 8-GPU instances bundling vCPU, RAM and support.

The NVIDIA H100 is still the workhorse of LLM training in 2026 — and renting one has never been cheaper. Prices fell sharply through 2025-2026 as supply caught up with demand. Here is where to get one cheapest today.

Cheapest H100 rentals, ranked (June 2026)

ProviderVariantOn-demand /hrSpot / community
HyperstackH100 PCIe~$1.90$1.52 (spot)
VultrH100 SXM~$2.30
RunPodH100 PCIe$2.39$1.99 (Community)
Verda (ex-DataCrunch)H100 SXM$3.25spot lower
LambdaH100 SXM$3.29
RunPodH100 SXM$3.29$1.99 (Community)
CrusoeH100 SXM$3.90
Together AIH100 SXM (cluster)$4.79
CoreWeaveHGX H100 SXM$6.16
AWSH100 SXM (p5/8)~$6.88~$3.37 (spot)
GCPH100 SXM (a3/8)~$10.98~$4.98 (spot)
AzureH100 SXM (ND/8)~$12.29~$2.27 (spot)

Snapshot June 2026 — prices change weekly; verify on each provider’s pricing page. See the live cheapest H100 ranking and the full H100 SXM and H100 PCIe price tables.

The cheapest tier: community and interruptible

If your training job checkpoints and can survive a restart, the discounted tiers beat everything:

The trade-off is interruption: your instance can be reclaimed. For checkpointed training and batch jobs that is usually fine. See on-demand vs spot.

Why hyperscalers cost 3-5x more

AWS, GCP and Azure sell flagship H100s only as whole 8-GPU instances (p5, a3, ND). The instance bundles dozens of vCPUs, hundreds of GB of RAM and enterprise support, so the per-GPU figure (instance price / 8) lands at $7-12/hr. You pay that premium for the ecosystem, regions and reliability — not for the silicon. For most independent ML work, a neocloud is far cheaper.

How to pick

  1. Need NVLink / multi-GPU? Choose SXM (Vultr, Lambda, RunPod, Verda).
  2. Single-GPU job, tightest budget? H100 PCIe on Hyperstack (~$1.90) or RunPod.
  3. Checkpointed / batch? RunPod Community or Vast.ai interruptible.
  4. Need a full node for big distributed training? See the cheapest 8xH100 node.

Price your exact job — GPUs x hours across every provider — in the cost calculator.

Frequently asked questions

What is the cheapest H100 per hour right now?

Around $1.90/hr for an H100 PCIe on Hyperstack and ~$2.30/hr for an H100 SXM on Vultr (on-demand). RunPod Community Cloud and Vast.ai interruptible can go below $2/hr if you can tolerate interruption. Prices are a June 2026 snapshot and change weekly.

Why is an H100 so much cheaper on RunPod or Vast than on AWS?

Specialist neoclouds and marketplaces sell raw GPU compute with little bundled overhead, so an H100 runs ~$2-4/hr. Hyperscalers sell flagship GPUs only as whole 8-GPU instances that bundle large amounts of vCPU, RAM and enterprise support, pushing the derived per-GPU price to $7-12/hr.

SXM or PCIe H100 — which is cheaper to rent?

PCIe is sometimes a little cheaper and is fine for single-GPU training or inference. SXM costs a touch more but has NVLink and higher bandwidth, which you want for multi-GPU distributed training. Most neoclouds list SXM; only a few publish PCIe.

Related articles

Last updated: 2026-06-19