Spot vs on-demand vs reserved GPUs: which should you rent?

Spot, on-demand or reserved: it is one of the first choices you make when renting GPUs, and one of the most expensive to get wrong. The same H100 can cost very different amounts depending on the pricing tier, and each tier carries a different risk. This guide compares the three, using recent prices, and gives a practical way to choose by workload.

The three pricing tiers

Tier	What you get	Price	Main risk
On-demand	GPUs by the hour, no commitment	H100 from about $1.25 to $11.07 an hour across 12 providers	Capacity may not be there when you need it
Spot	Spare capacity at a discount	Runs 60% to 91% below on-demand	Capacity can be pulled mid-job
Reserved / contract	A fixed rate for a committed term	One-year H100 contracts rose from $1.70 to $2.35 per GPU-hour, Oct 2025 to Mar 2026 (SemiAnalysis)	You pay whether you use it or not

On-demand prices from getdeploying.com, 19 Sep 2026. SKUs vary by provider.

On-demand GPUs

On-demand is the default. You rent by the hour and stop paying when you stop. No commitment, no contract.

The spread is the first thing to know. An on-demand H100 lists from about $1.25 an hour at developer clouds and marketplaces to $11.07 an hour at the most expensive hyperscaler. Where you rent matters as much as which tier you pick. See why the same H100 costs $1.25 to $11.07 for the full breakdown by provider.

The second thing is availability. SemiAnalysis reported on-demand capacity sold out across all GPU types in 2026. On-demand is flexible on paper. In practice, the provider you want may have nothing free when your job is ready.

Spot GPUs

Spot is spare capacity sold at a discount, running 60% to 91% below on-demand. That is the cheapest way to rent a GPU by a wide margin.

The catch is that the provider can take the capacity back mid-job. For work that can stop and resume, that is a fair trade. For a long training run without frequent checkpoints, losing the GPUs partway through can cost more than the discount saved.

Reserved and contract GPUs

Reserved capacity means committing to a term, often a year, in return for a fixed rate and a guarantee the GPUs are yours.

Two things changed this tier recently. First, contract prices rose fast: one-year H100 contracts went from $1.70 to $2.35 per GPU-hour between October 2025 and March 2026, up about 40% (SemiAnalysis). Second, the supply is spoken for. Thunder Compute reports most new NVIDIA allocation is reserved into 2027.

Reserved capacity also has a hidden cost: you pay for every hour, busy or not. A VentureBeat enterprise survey (n=107) found 83% of enterprises run their GPUs at 50% utilization or less. More on that in our post on GPU utilization statistics.

Which tier fits your workload

There is no single right answer, but the workload usually decides it.

Workload	Usually fits	Why
Long training runs	On-demand, or reserved if you train continuously	Interruptions are costly. Spot works only with frequent checkpoints and a job that can resume
Short or experimental training	On-demand or spot	Low commitment. A lost spot run costs little
Production inference	Reserved for the steady base, on-demand for peaks	Needs to stay up. Steady load uses reserved capacity well
Batch jobs	Spot	Work can usually be retried, so interruptions are cheap and the discount is large
Deadline-bound jobs	On-demand	You need capacity that won't be pulled before the deadline

A few practical rules:

• Commit only to load you are sure of. Reserve for the steady part of your usage and rent the rest.
• Checkpoint before you use spot. If a job can't survive losing its GPUs, it doesn't belong on spot.
• Compare across providers, not just tiers. An on-demand price at one provider can be lower than another's contract rate. The spread between providers is often bigger than the spread between tiers.
• Recheck often. Prices move by double digits in weeks. AWS cut its H100 on-demand price by 44% in a single announcement in June 2025. See how fast GPU rental prices move.

Where Opengrid fits

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

Choosing a tier is one decision. Choosing a provider for each tier, every time prices move, is the harder one. With Opengrid, you describe the job, like GPU type, count, region, price ceiling and deadline, and Opengrid places it on whichever provider wins on price, availability, region and reliability. The market view shows you what each tier costs across providers before you commit. Read more about how GPU routing works.