Who rents the H200 SXM cheapest?

RunPod, at $3.59/GPU-hr on H200 · Community Cloud, read on 17 Sept 2026. That is the on-demand rate for one GPU for one hour; the whole machine bills $3.59/hr.

What is the NVIDIA H200 SXM?

An H100 with HBM3e: same compute, nearly twice the memory, which is what decides how big a model fits. It is a Hopper board from 2024 with 141 GB of memory at 4.80 TB/s, in the SXM package. Those figures come from the vendor, not from any provider's price page.

Can I rent a H200 SXM on spot?

Yes. 6 of the 9 configurations here publish an interruptible rate, the cheapest being $2.16/GPU-hr at Verda (DataCrunch), 50% under its own on-demand rate. The machine can be taken back mid-run, so it suits work that checkpoints.

Why do H200 SXM prices differ so much?

The spread here runs from $3.59/GPU-hr to $6.31/GPU-hr for the same silicon, because what else comes with it differs: the vCPU and RAM around the GPU, whether a disk is included, how many regions it runs in, and whether the machine belongs to the provider or to a third party on a marketplace. Each row says which.

How current are these prices?

They are read every morning from each provider's own public page or API, and this page recomputes with them. Last refresh: 17 Sept 2026. Reserved and committed rates are not indexed. Prices exclude tax.

NVIDIA H200 SXM pricing.

prices as of 17 Sept 2026 · 5 providers · 10 configurations

An H100 with HBM3e: same compute, nearly twice the memory, which is what decides how big a model fits.

The cheapest H200 SXM today is H200 · Community Cloud at RunPod, $3.59/GPU-hr, read 17 Sept 2026; the cheapest interruptible rate is $2.16/GPU-hr at Verda (DataCrunch).

141 GB memory4.80 TB/sHopperSXMlaunched 2024

Head to headCrusoe vs Verda (DataCrunch)Crusoe vs Modal Crusoe vs RunPod Verda (DataCrunch) vs RunPod or any two machines →

Cheapest on demand

$3.59

per GPU-hour at RunPod

Median of the spread

$4.33

across 9 configurations

Cheapest interruptible

$2.16

6 of 9 publish one

Per GB of memory

$0.0255/GB-hr

market median $0.0300/GB-hr

What the board is

published by NVIDIA, not by a provider

Memory: 141 GB
Memory bandwidth: 4.80 TB/s
Architecture: Hopper
Launched: 2024

These are the chip's own figures, kept in this site's registry and the same whoever rents it out. A provider's feed is trusted for its prices, its GPU counts and the memory it states per card, never for what the board is. There is no speed score on this page: bandwidth is a published number, and turning it into a benchmark would be inventing one.

Every provider renting the H200 SXM

Cheapest per GPU-hour first. The rate is for one GPU for one hour, so a node of eight sits beside a single card at a comparable number, with the whole machine's hourly price under it. Where a provider publishes an interruptible rate it is in the same cell.

RunPod
H200 · Community Cloud
1x H200 SXM · 141 GB each · 24 vCPU · 276 GB RAM
disk none, billed separately · regions not stated · billed per second · marketplace
Community Cloud: vetted third-party hosts, cheaper and less consistent · storage billed separately · billed per second
View at RunPod ↗
$3.59per GPU-hour$3.59/hr all inno spot rate
Crusoe
NVIDIA H200 HGX
1x H200 SXM · 141 GB each · host size not stated
disk none, billed separately · regions not stated · billed per hour
Priced per GPU-hour · vCPU, RAM and disks billed separately · egress free · no minimum commitment
Crusoe ↗
$4.29per GPU-hour$4.29/hr all inno spot rate
Verda (DataCrunch)
1H200.141S.44V · 1x H200 SXM5 141GB
1x H200 SXM · 141 GB each · 44 vCPU · 170 GB RAM
disk none, billed separately · regions not stated · billed per hour
Dedicated Hardware Instance · storage billed separately
Verda (DataCrunch) ↗
$4.33per GPU-hour$4.33/hr all in$2.16 spot
Verda (DataCrunch)
2H200.141S.88V · 2x H200 SXM5 141GB
2x H200 SXM · 141 GB each · 88 vCPU · 340 GB RAM
disk none, billed separately · regions not stated · billed per hour
Dedicated Hardware Instance · storage billed separately · 900 GB/s between GPUs
Verda (DataCrunch) ↗
$4.33per GPU-hour$8.65/hr all in$2.16 spot
Verda (DataCrunch)
4H200.141S.176V · 4x H200 SXM5 141GB
4x H200 SXM · 141 GB each · 176 vCPU · 680 GB RAM
disk none, billed separately · regions not stated · billed per hour
Dedicated Hardware Instance · storage billed separately · 900 GB/s between GPUs
Verda (DataCrunch) ↗
$4.33per GPU-hour$17.31/hr all in$2.16 spot
Verda (DataCrunch)
8H200.141S.176V · 8x H200 SXM5 141GB
8x H200 SXM · 141 GB each · 176 vCPU · 1360 GB RAM
disk none, billed separately · regions not stated · billed per hour
Dedicated Hardware Instance · storage billed separately · 900 GB/s between GPUs
Verda (DataCrunch) ↗
$4.33per GPU-hour$34.62/hr all in$2.16 spot
RunPod
H200 · Secure Cloud
1x H200 SXM · 141 GB each · 24 vCPU · 276 GB RAM
disk none, billed separately · regions not stated · billed per second
Secure Cloud: RunPod's own datacentres · storage billed separately · billed per second
RunPod ↗
$4.59per GPU-hour$4.59/hr all inno spot rate
Together AI
NVIDIA HGX H200
1x H200 SXM · 141 GB each · host size not stated
disk none, billed separately · regions not stated · billed per hour
GPU Clusters, on demand and per GPU-hour · shared filesystem billed separately · reserved terms of 7 days and longer cost less and are not indexed · memory is the board's published spec, which Together's table does not give
Together AI ↗
$5.99per GPU-hour$5.99/hr all in$2.99 spot
CoreWeave
NVIDIA HGX H200
8x H200 SXM · 141 GB each · 128 vCPU · 2048 GB RAM · NVLink
disk 61.4 TB · 2 regions · billed per hour
$50.44 an hour for the whole node · egress and VPC free · IPv4 and storage billed separately · spot shown is the cheaper of the two regions
CoreWeave ↗
$6.31per GPU-hour$50.44/hr all in$2.58 spot

Billed as a function

1 rate · never ranked

These rates buy a container that runs when it is called, not a machine you hold. CPU and memory are billed on top, a chosen region or a guarantee against preemption costs more, and the figure is a floor rather than a price. They are listed for what they are and ranked nowhere on this site.

Modal
Nvidia H200 SXM
1x H200 SXM · 141 GB each · host size not stated
disk none, billed separately · regions not stated · billed per second · serverless
$0.001261 a GPU-second, the preemptible region-agnostic rate · a chosen region costs 1.15x to 1.75x and non-preemptible 3x · CPU, memory and volumes billed on top · memory is the board's published spec
Modal ↗
$4.54per GPU-hour$4.54/hr all inno spot rate

Frequently asked

Who rents the H200 SXM cheapest?: RunPod, at $3.59/GPU-hr on H200 · Community Cloud, read on 17 Sept 2026. That is the on-demand rate for one GPU for one hour; the whole machine bills $3.59/hr.
What is the NVIDIA H200 SXM?: An H100 with HBM3e: same compute, nearly twice the memory, which is what decides how big a model fits. It is a Hopper board from 2024 with 141 GB of memory at 4.80 TB/s, in the SXM package. Those figures come from the vendor, not from any provider's price page.
Can I rent a H200 SXM on spot?: Yes. 6 of the 9 configurations here publish an interruptible rate, the cheapest being $2.16/GPU-hr at Verda (DataCrunch), 50% under its own on-demand rate. The machine can be taken back mid-run, so it suits work that checkpoints.
Why do H200 SXM prices differ so much?: The spread here runs from $3.59/GPU-hr to $6.31/GPU-hr for the same silicon, because what else comes with it differs: the vCPU and RAM around the GPU, whether a disk is included, how many regions it runs in, and whether the machine belongs to the provider or to a third party on a marketplace. Each row says which.
How current are these prices?: They are read every morning from each provider's own public page or API, and this page recomputes with them. Last refresh: 17 Sept 2026. Reserved and committed rates are not indexed. Prices exclude tax.

Same generationH100 SXM GH200 H100 PCIeSimilar memoryH100 SXM · 80 GB A100 80GB SXM · 80 GB B200 · 180 GB RTX PRO 6000 · 96 GB