Best Cloud GPU Providers (2026): 11 Places to Rent GPUs

A single high-end GPU can sit idle for most of the week in a small company. Renting one only when you need it can be the difference between trying an AI project and shelving it.

The eleven providers below are listed alphabetically and in no particular order, and every description comes from the provider's own website, read this week. If cloud services in general are new to you, my explainer on what cloud hosting is and its benefits sets the scene.

Cloud GPU providers compared

I checked the free plan column on each tool's own website, as of Nov 2026. Yes means the site names a free plan, a free account, a free tier or an open source version. Trial only means a time-limited trial. Not stated means the pages I read named neither, even where there's a button inviting you to start for free, so check the vendor's page before you sign up.

Tool	Type	Good for	Free plan
CoreWeave	AI cloud for training and serving models	Large-scale training clusters	Not stated
Hyperstack	On-demand cloud GPU service	On-demand GPUs by the hour	Not stated
Jarvislabs	GPU rental for AI and deep learning	Simple GPU rental with support	Not stated
Lambda	AI compute in the cloud	Teams training their own models	Not stated
Lium	GPU cloud marketplace	Renting whole machines by the second	No
Modal	Platform for production AI	Serverless inference and sandboxes	Not stated
Nebius	AI cloud from silicon to API	Self-service GPU clusters	Not stated
Paperspace	Cloud GPU platform from DigitalOcean	Existing customers of that cloud	Yes
Runpod	AI developer cloud	GPU pods and serverless endpoints	Not stated
TensorDock	Cloud GPU platform	A wide choice of GPU types	Not stated
Vast.ai	Platform for renting GPUs from many hosts	Renting GPUs from many hosts	Not stated

Eleven cloud GPU providers to consider

CoreWeave

Best for: AI companies that need a cloud purpose-built for training and serving models at very large scale.

Hyperstack

Best for: Startups that want fast, on-demand cloud GPUs for AI and machine learning work, with hourly rates shown on the site.

Jarvislabs

Best for: Researchers and solo developers who want simple GPU rental for AI and deep learning with friendly support.

Lambda

Best for: Research and engineering teams that want cloud compute built around AI training, from single instances to clusters.

Lium

Best for: Developers and AI agents who want to rent a whole GPU machine, from RTX cards to B300, and pay by the second.

Lium is a GPU cloud marketplace for developers and AI agents. You rent a whole machine over SSH, from RTX cards right up to H100, H200, B200 and B300, and you're billed by the second only while it runs, with no long-term commitment. Everything works from a command line tool, a Python SDK or a REST API. What I like most is the MCP server, which lets Claude Code and Codex rent and manage GPUs on their own, so an agent can start a machine, run the job and shut it down within a budget you set. Nothing sits idle on your bill: when the machine stops, so does the cost.

Modal

Best for: Developers who'd rather write code than manage servers, using serverless inference and sandboxes for AI workloads.

Nebius

Best for: Teams that want to go from zero to GPU clusters quickly with self-service access on a dedicated AI cloud.

Paperspace

Best for: Developers who already know DigitalOcean and want cloud GPUs for AI and machine learning workloads there.

Runpod

Best for: AI developers who want GPU pods and serverless endpoints across many regions, plus ready templates for open models.

TensorDock

Best for: Hobbyists and small teams who want a wide choice of GPU models for training, inference, rendering or gaming.

Vast.ai

Best for: Thrifty developers who want to rent GPUs from many independent hosts for AI, rendering or transcription work.

machine0

machine0 is a recent Y Combinator launch, from the Summer 2026 batch, and it sits outside the eleven above. It gives you persistent virtual machines of up to 60 vCPUs and 240 GB of RAM, with optional GPUs, billed per minute. Each machine has a static public IP and an HTTPS endpoint, and you control it from a CLI or through MCP, so an AI agent can run it too. It's built for agents and developers with long-lived or compute-heavy workloads. GPU sizes are offered only in the us-east, uk, eu and asia regions, so check that your location fits first.

Where to start for your project

• Writing code without managing servers: Modal.
• Pods and serverless endpoints for open models: Runpod.
• Renting from many independent hosts: Vast.ai.
• Letting an AI agent rent its own GPU, billed by the second: Lium.
• Large training clusters: CoreWeave or Nebius.
• Developers already using DigitalOcean: Paperspace.

The NCSC's first cloud security principle

The National Cyber Security Centre's cloud security principles begin with data in transit protection. It says your data should be adequately protected against tampering and eavesdropping as it moves across networks inside and outside the cloud.

That matters with GPUs because you're often sending your most valuable data and models to someone else's machines. Ask each provider how data is encrypted in transit and at rest, and what happens to your files when an instance is deleted.

On-demand instances, serverless and clusters

On-demand instances give you a whole GPU machine for as long as you keep it running. They suit training runs and experiments where you want full control.

Serverless GPUs start when a request arrives and stop when it's done, which suits AI features with uneven traffic. Clusters link many GPUs together for large training jobs that one machine can't handle.

Want your tool on this list? Featured spots and reviews here are paid partnerships. Here's how to partner with me.

Questions to ask before you rent a cloud GPU

1. Which GPU models are available right now?
2. Is billing per second, per hour or per month?
3. Can I stop an instance without losing my files?
4. Which regions are the data centres in?
5. Is there serverless inference as well as full machines?
6. What security and compliance details can they share?

Renting the hardware is only part of the job, and our guide to MLOps platforms weighs up platforms that manage experiments and models.

Signs it's time to rent GPUs

• Training runs take days on a laptop. Move them to the cloud.
• You're about to buy a GPU for one project. Rent first.
• Your AI feature has spiky traffic. Look at serverless.
• Machines run overnight unused. Set them to stop.

Do you make a cloud GPU service? Tool makers can ask to be considered for a featured place on this list.

Four more GPU clouds, matched to the job

Updated 5 November 2026. Short answer: if you want flexibility rather than one fixed contract, these four belong on your shortlist. Here's who each one suits.

• You want the newer NVIDIA cards: Hyperbolic lists H100, H200 and B200 GPUs on demand or reserved, plus managed Kubernetes and a private cloud option.
• You want a clear hourly price: GMI Cloud shows H100 from $2.60 /GPU-hour and H200 from $3.20 /GPU-hour as of Nov 2026, with serverless inference and an agent runtime on the same platform.
• You book capacity in blocks: SF Compute lets you reserve capacity and sell back idle compute through its scheduler, with VMs, bare metal and managed Slurm.
• You want many clouds on one bill: Shadeform deploys GPUs across 30+ clouds and 100+ regions from one console, with centralized billing.

My verdict: start with Shadeform if you want to shop around, and GMI Cloud if you need a published hourly rate to budget against.

My plain advice on choosing a cloud GPU provider

Match the provider to the job. A weekend experiment needs a quick, simple rental. A product feature with real users needs reliability, support and security answers. A big training run needs clusters and a provider that can guarantee capacity.

And always switch machines off when you're done. Idle GPUs are the easiest way to waste money in the cloud. If you're still learning how the models themselves work, my plain guide to transformer models is a gentle introduction.

Inference across four chip makers from one private AI cloud

Some GPU clouds now sell the software layer as well as the chips. In September 2026 one provider put several accelerator brands behind a single inference platform, which matters if your data has to stay inside your own environment.

Cirrascale Cloud Services

In September 2026, Cirrascale Cloud Services announced the production release of its Inference Platform, which its press page calls a complete software stack for enterprise inference. It describes itself as a neocloud built for Private AI, with NVIDIA, AMD, Qualcomm and Tenstorrent clouds. The vendor says each request is routed to a suitable model and accelerator without code changes, and every pipeline is run from one web console. You can also fine-tune on private data that never leaves your environment. Teams with strict data rules should look here, and check which accelerators your models are tested on before you commit.

Frequently asked questions

What is a cloud GPU?

A cloud GPU is a graphics processor you rent from a provider over the internet, used for AI training, inference, rendering and other heavy computing.

Why rent a GPU instead of buying one?

Renting lets you use powerful hardware only when you need it, without a large upfront purchase or the work of running it yourself.

Are there free cloud GPUs?

Most providers charge for GPU time, though some offer credits or trials. Check each provider's own page for current offers.

What's serverless GPU?

Serverless GPU starts compute when a request arrives and stops afterwards, so you don't keep a machine running between jobs.

Is my data safe on a rented GPU?

It depends on the provider. Ask how data is protected in transit and at rest, and what happens when an instance is deleted.

Which GPU do I need for AI?

It depends on the model size and job. Larger models and training runs need more GPU memory, while smaller inference jobs can run on modest cards.