How to choose a GPU cloud provider, and why switching costs so much

Choosing a GPU cloud provider looks easy. You compare a few prices, open an account and launch an instance. Signing up takes minutes. The hard part comes later, when the market moves and a better option appears, because switching GPU cloud providers means starting over.

This guide covers what to check before you pick a provider, why switching is so expensive, and how to stay flexible when you can't predict next quarter's prices.

Up front: we are Opengrid, and we are building a product to make this easier. Opengrid is pre-launch. The market view ships first and the routing API is in development. The checklist below works whatever you end up using.

The market you are choosing from

There are a lot of options, and they are spread far apart:

• 209 GPU clouds. SemiAnalysis tracks 209 GPU clouds, up from 124. Each has its own console, contract, API and bill.
• Prices spread across a wide band. 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. The same is true of other chips: A100s range from $1.09 to $5.07 an hour and B200s from $3.50 to $27.04.
• Reliability varies a lot. SemiAnalysis's ClusterMAX 2.0 grades GPU clouds from Platinum down to Underperforming. Only one is rated Platinum.
• Capacity is tight. SemiAnalysis reported in 2026 that on-demand capacity was sold out across all GPU types, and Thunder Compute reports most new NVIDIA allocation is reserved into 2027.

For a breakdown by provider type, see why the same H100 costs so much more at some providers.

Why switching costs so much

Every time you move to a new provider, you pay four setup costs again:

1. A new account and a new contract. Procurement, terms and often a minimum commitment.
2. A new API and new tooling. Provisioning scripts, images, networking and storage all need rework.
3. A new bill to reconcile. One more invoice and one more dashboard for finance to track.
4. Re-testing who is actually reliable. Benchmarks and burn-in runs before you trust a new cluster with real work.

So teams pick once and overpay for as long as they stay. That made sense when prices barely moved. It makes less sense now. AWS cut its H100 on-demand price by 44% in a single announcement in June 2025. Then one-year H100 contract prices rose about 40% between October 2025 and March 2026. The right provider this quarter is often the wrong one next quarter. More on that in how GPU rental prices moved in 2025 and 2026.

A checklist for choosing a GPU cloud provider

1. Write down the job, not the instance

Start with GPU type, GPU count, region, price ceiling and deadline. If you start from an instance list, you will anchor on whatever the first provider shows you.

2. Check availability before price

With on-demand capacity this tight, the cheapest listing is worthless if it can't give you the GPUs when you need them. Confirm the provider can actually deliver your GPU count in your region.

3. Benchmark the price against the market

Compare the quote with the whole band for that chip, not just two or three providers. Rough H100 on-demand bands as of September 2026:

Provider type	Examples	H100 on-demand, per GPU-hour
Hyperscalers	Google Cloud, Oracle, Azure, AWS	$6.88 to $11.07
Neoclouds	CoreWeave, Crusoe, Nebius, Lambda	$3.29 to $6.16
Developer clouds and marketplaces	RunPod, Together, Vast.ai, Lium	$1.25 to $2.89

4. Decide between spot and on-demand

Spot capacity runs 60 to 91% below on-demand, but it can be pulled in the middle of a job. Spot suits jobs that checkpoint often or can restart. Long training runs with no checkpointing belong on on-demand or reserved capacity.

5. Check reliability data

Look at independent ratings like ClusterMAX and ask other teams about their experience. A cheap cluster that fails halfway through a training run costs more than an expensive one that finishes.

6. Keep the contract short where you can

A long commitment locks in today's price. That is good if prices rise and bad if they fall, and in 2025 and 2026 they did both. If you need certainty, the new GPU rental futures are one way to hedge without a multi-year contract.

7. Keep your code provider-agnostic

The less your code depends on one provider's APIs, the cheaper it is to switch. Orchestrators like SkyPilot or dstack help when you already hold accounts at several clouds, and aggregators like Shadeform give you one API across about 20 clouds. See our comparison of GPU cloud aggregators and marketplaces.

Where Opengrid fits

Opengrid is built on the idea that you shouldn't have to pick a provider at all. You describe the compute you need, and Opengrid routes the job to whichever provider wins on price, availability, region and reliability. You get one invoice whichever provider runs it. When the market moves, your code doesn't change.

	Doing it yourself	With Opengrid (pre-launch)
Compare prices	By hand, across a few providers	Live market view across providers
Pick a provider	You decide	Routed by price, availability, region and reliability
Switch providers	New account, API, bill and testing	No code changes
Billing	One bill per provider	One invoice
Status	Available today	Market view first, routing API in development

The short version

Start from the job, check availability first, benchmark the price against the whole market, match spot or on-demand to how your job handles interruptions, and keep your setup portable. If you want that decision made for you, join the Opengrid waitlist.