GPU rental prices in 2025 and 2026: how fast they move

GPU rental prices used to be something you checked once a year. In 2025 and 2026 they moved by double digits in weeks. If you rent GPUs, the price you locked in last quarter may already be well above or below the market.

This post walks through how GPU rental prices moved over fifteen months, using the NVIDIA H100 as the example, and what that means for anyone buying compute.

The short version

When	What happened	Move	Source
5 Jun 2025	AWS cuts its H100 on-demand price in a single announcement	−44%	AWS, via CloudZero and IntuitionLabs
Jun to Jul 2025	Other providers follow with their own cuts, in single moves	−20% to −25%	CloudZero, IntuitionLabs
Oct 2025 to Mar 2026	One-year H100 contract price climbs from $1.70 to $2.35 per GPU-hour	about +40%	SemiAnalysis
Dec 2025 to Jan 2026	Silicon Data H100 rental index rises from $2.00 to $2.20 in four weeks	+10%	Silicon Data
2026	On-demand capacity sold out across all GPU types	n/a	SemiAnalysis

In fifteen months the H100 went from a sharp price war to a shortage, with prices rising again.

Mid-2025: the price war

On 5 June 2025, AWS cut its H100 on-demand price by 44% in a single announcement. That is not a gradual drift. It is nearly half the price, overnight, from the largest cloud provider in the world.

Other providers did not wait long. Through June and July 2025, several cut their own prices by 20% to 25%, again in single moves rather than small steps.

For buyers, this was the moment to renegotiate or move. Teams that had signed earlier contracts, or picked a provider and stayed put, kept paying the old price.

Late 2025 into 2026: the reversal

Then the direction flipped.

• Contract prices rose about 40%. SemiAnalysis tracked one-year H100 contracts rising from $1.70 to $2.35 per GPU-hour between October 2025 and March 2026.
• The rental index jumped 10% in four weeks. Silicon Data's H100 rental index went from $2.00 to $2.20 between December 2025 and January 2026.
• Capacity dried up. SemiAnalysis reported on-demand sold out across all GPU types in 2026, and Thunder Compute reports most new NVIDIA allocation is reserved into 2027.

The teams that locked in contracts at the low point did well. The teams that waited paid 40% more for the same commitment a few months later.

Each GPU moves on its own

It is tempting to think of "GPU prices" as one number. They are not. During the same stretch where the H100 rental index rose 10%, A100 and B200 prices stayed flat (Silicon Data).

Every GPU type has its own price band, and the bands are wide. Thunder Compute's September 2026 data:

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.

So "is now a good time to rent?" has no single answer. It depends on the chip, the provider and the pricing tier. For a provider-by-provider breakdown, see H100 price per hour in 2026.

Price is not the only thing moving

Three other variables shift alongside the price:

• Capacity. A provider can be the cheapest on paper and have nothing available. With on-demand sold out across GPU types, finding capacity is often harder than finding a low price.
• Pricing tier. Spot runs 60% to 91% below on-demand, but the capacity can be pulled mid-job. Contracts are cheaper per hour but lock you in, and as above, contract prices can rise fast.
• The provider list itself. SemiAnalysis tracks 209 GPU clouds, up from 124. New supply keeps appearing, and it is spread across more places than anyone can check by hand.

What this means for GPU buyers

The main lesson from 2025 and 2026: the right provider this quarter is the wrong one next quarter.

That would be fine if switching were cheap. It isn't. Every switch means starting over:

• A new account and a new contract
• A new API and new tooling
• A new bill to reconcile
• Re-testing who is actually reliable

So most teams pick once and overpay for as long as they stay. When AWS cut prices 44%, teams locked in elsewhere did not benefit. When contract prices climbed 40%, teams that had not shopped around had no fallback.

A few practical habits help:

1. Re-check prices every quarter at least. Monthly, if GPU spend is a big line in your budget.
2. Keep your workloads portable. Containers, checkpoints and provider-agnostic tooling make a move cheaper.
3. Match the tier to the job. Spot for jobs that checkpoint often, contracts for steady baseline load, on-demand for everything in between.
4. Track more than price. Availability and reliability matter as much as the hourly rate.

Our guide to switching GPU cloud providers goes deeper on keeping the cost of a move low.

Where this goes next

Prices that move this much are starting to be treated like a commodity. CME lists H100 and B200 rental futures from 5 October 2026, built on Silicon Data's indices. We explain how that works in CME GPU futures explained.

Where Opengrid fits

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

Opengrid is building a live view of GPU prices and availability across providers, then a routing API that places each job wherever price, availability, region and reliability are best at that moment. The goal is simple: when the market moves, your code shouldn't have to. You describe the compute you need once, and switching providers stops being a project.