H100 price per hour in 2026: why the same GPU costs $1.25 to $11.07

The H100 price per hour depends almost entirely on who you rent it from. On 19 September 2026, on-demand list prices for the same NVIDIA H100 ranged from $1.25 to $11.07 per GPU-hour across twelve providers we looked at, according to getdeploying.com. That is a gap of nearly 9× for the same chip.

This post breaks down where those prices sit, why they split into three bands, and how contract and spot pricing change the picture.

A note on dates: every price here is a snapshot. On-demand list prices are from getdeploying.com on 19 Sep 2026, and SKUs vary by provider (interconnect, memory, node size). GPU prices move by double digits in weeks, so treat these numbers as a guide to the shape of the market, not a quote.

H100 on-demand prices by provider

Provider	Tier	H100 on-demand, $ per GPU-hour
Google Cloud	Hyperscaler	$11.07
Oracle	Hyperscaler	$10.00
Azure	Hyperscaler	$6.98
AWS	Hyperscaler	$6.88
CoreWeave	Neocloud	$6.16
Crusoe	Neocloud	$3.90
Nebius	Neocloud	$3.85
Lambda	Neocloud	$3.29
RunPod	Developer cloud / marketplace	$2.89
Together	Developer cloud / marketplace	$1.99
Vast.ai	Developer cloud / marketplace	$1.79
Lium	Developer cloud / marketplace	$1.25

Source: getdeploying.com, 19 Sep 2026. On-demand list prices. SKUs vary by provider.

Three price bands for one chip

The table falls into three clear tiers.

Hyperscalers: $6.88 to $11.07

Google Cloud, Oracle, Azure and AWS charge the most. You are paying for the rest of the platform around the GPU: the storage, networking, identity, compliance and enterprise contracts your company may already run on. For teams already deep inside one of these clouds, staying put can be worth the premium. For a training run that only needs GPUs, it often is not.

Neoclouds: $3.29 to $6.16

CoreWeave, Crusoe, Nebius and Lambda are GPU-first clouds. They are built around AI workloads and price well below the hyperscalers, though the spread inside this tier is still almost 2×.

Developer clouds and marketplaces: $1.25 to $2.89

RunPod, Together, Vast.ai and Lium sit at the bottom of the range. Some are developer clouds, some are marketplaces where many hosts list capacity. The prices are the lowest, and the trade-offs vary a lot from one provider to the next: availability, network setup, and how reliable a given host is.

H100 vs other GPUs: the price bands

The H100 is not the only chip with a wide spread. Thunder Compute's September 2026 data shows the same pattern across the A100, H200 and B200:

GPU	Price range, $ per hour
A100	$1.09 to $5.07
H100	$2.89 to $11.06
H200	$3.44 to $10.60
B200	$3.50 to $27.04

Source: Thunder Compute, Sep 2026.

You may notice the H100 low end here ($2.89) is higher than the $1.25 in the provider table above. The two come from different sources, on different dates, covering different sets of providers. They don't contradict each other. They show that the answer to "what does an H100 cost?" depends on whose list you check and when.

Each GPU type also moves on its own. During one stretch, H100 rental prices rose 10% while A100 and B200 stayed flat. The cheapest chip for your job this month is not guaranteed to be the cheapest next month.

On-demand vs contract vs spot

On-demand list prices are only one of three ways to pay.

Pricing tier	What it means	What happened to H100 pricing
On-demand	Pay by the hour, no commitment	$1.25 to $11.07 depending on provider (getdeploying.com, 19 Sep 2026)
One-year contract	Commit for a year in exchange for a lower rate	Rose from $1.70 to $2.35 per GPU-hour between Oct 2025 and Mar 2026, about +40% (SemiAnalysis)
Spot	Spare capacity at a discount, can be pulled mid-job	Runs 60% to 91% below on-demand

Contracts are usually cheapest per hour, but they lock you in, and their price has been climbing. Spot is the cheapest of all, but your capacity can disappear in the middle of a training run. That is fine for jobs that checkpoint often, and painful for those that don't.

Why H100 prices differ so much

A few forces keep the spread wide:

• The market is fragmented. SemiAnalysis tracks 209 GPU clouds. Each has its own console, contract and API, so comparing them all by hand is impractical.
• Capacity is tight. SemiAnalysis reported on-demand sold out across all GPU types in 2026, and Thunder Compute reports most new NVIDIA allocation is reserved into 2027. When cheap capacity disappears, buyers pay whatever is available.
• Prices move fast. AWS cut its H100 on-demand price by 44% in a single announcement in June 2025. A few months later, contract prices were up 40%. See how GPU rental prices moved in 2025 and 2026.
• Switching is expensive. Moving to a cheaper provider means a new account, a new contract, a new API and a new bill. So most teams pick once and stay, even after their provider stops being the best deal.

How to find the cheapest H100 that will actually run your job

The cheapest list price is not always the best option. Before you choose, check:

1. Availability. A low price on capacity you can't get is worth nothing.
2. Region. Data residency and latency may rule out some providers.
3. Reliability. A cheap host that fails halfway through a run costs you the whole run.
4. Pricing tier. Decide between on-demand, contract and spot based on how long the job runs and whether it can survive interruption.
5. Total cost. Storage, egress and setup time all add to the hourly rate.

Then check again next quarter, because the answer will probably have changed. Our guide to switching GPU cloud providers covers how to do that without starting over every time.

Where Opengrid fits

Up front: Opengrid is pre-launch. The market view ships first and the routing API is in development.

Opengrid is building one view of GPU prices and availability across providers, and then a routing API on top of it. You describe the job (for example 8× H100 in the US, under $2.50 an hour, done within 12 hours) and Opengrid places it with whichever provider wins on price, availability, region and reliability. When the market moves, your code doesn't have to.

For a look at the tools that exist today, see our comparison of GPU cloud aggregators and marketplaces.