Best overall
Highest score of the 10 providers in this ranking (83/100). Teams that already live in AWS or need its service catalog.
Visit Amazon Web ServicesRent GPUs by the hour for training, fine-tunes and inference. Supply moves the real price. We cite a public floor when the vendor publishes one, and we do not invent stock.
Updated Nov 2026 · 10 providers · 1 with a published monthly plan · 2 with a free tier
Three starting points from the ranking below, each for a different kind of buyer.
Best overall
Highest score of the 10 providers in this ranking (83/100). Teams that already live in AWS or need its service catalog.
Visit Amazon Web ServicesBest free tier
The highest-ranked provider with a free plan or trial, after the picks above. Teams that want GCP data services or GPUs next to the app.
Visit Google CloudSorted by score out of 100. A from-price is the cheapest plan the vendor publishes, not a quote for your workload.
Teams that already live in AWS or need its service catalog
Teams that want GCP data services or GPUs next to the app
Teams that need a lot of GPU, not one cheap hour
GPU capacity when site and energy story matter
A single GPU VM without a hyperscaler account
On-demand training boxes with a known GPU
Notebooks and small training runs on a known platform
Fine-tunes and batch jobs that can checkpoint
A short GPU job when you are willing to pick the host
The lowest GPU hour when you can tolerate a marketplace
You pay for GPU hours, by model and by how the capacity is reserved. These are the lines that decide the invoice.
The hourly rate depends mostly on the card: consumer cards, previous-generation datacenter cards and the latest accelerators sit in very different price bands.
On-demand is the list price. Spot or interruptible capacity is cheaper and can be taken back. Reserved clusters need a contract and usually a minimum term.
Datasets, checkpoints and container images sit on persistent storage billed per GB per month. Moving data out costs extra on many clouds.
A GPU bills while it is allocated, not while it computes. Notebooks left open overnight are a classic budget leak.
| Provider | Starting price | Billing model | Free tier |
|---|---|---|---|
| Amazon Web Services | Lightsail from $3.50/mo | Pay for what you use: requests, tokens, traffic or compute time. The bill moves with volume. | Yes |
| Google Cloud | Free tier, then usage | Pay for what you use: requests, tokens, traffic or compute time. The bill moves with volume. | Yes |
| CoreWeave | Public GPU price list | Billed by the hour for the resources you keep allocated. Idle capacity still counts. | No |
| Crusoe | Public GPU cloud rates | Billed by the hour for the resources you keep allocated. Idle capacity still counts. | No |
| Hyperstack | Hourly GPU VMs | Billed by the hour for the resources you keep allocated. Idle capacity still counts. | No |
| Lambda | Public on-demand GPU rates | Billed by the hour for the resources you keep allocated. Idle capacity still counts. | No |
| Paperspace | Public hourly GPU rates | Billed by the hour for the resources you keep allocated. Idle capacity still counts. | No |
| RunPod | Community GPUs from about $0.20/hr | Billed by the hour for the resources you keep allocated. Idle capacity still counts. | No |
| TensorDock | Marketplace hourly rates | Billed by the hour for the resources you keep allocated. Idle capacity still counts. | No |
| Vast.ai | Marketplace, often from about $0.10/hr | Billed by the hour for the resources you keep allocated. Idle capacity still counts. | No |
You want the lowest published hourly floor and can tolerate variable stock or interrupts. RunPod, Vast.ai and TensorDock are typical starting points.
You need on-demand capacity that stays up. Prefer hyperscaler GPU families or vendors that sell non-spot instances you can keep attached.
Multi-GPU interconnect and longer jobs matter. Shortlist CoreWeave, Crusoe, Lambda and hyperscaler training shapes; confirm the exact SKU and network before you schedule the run.
Utilization is low or ops cost dominates. Buy tokens on an [LLM API](/llm-apis) until the GPU would stay busy enough to justify renting hardware.
GPU lists go stale quickly because stock and spot prices move. RunPod, Vast.ai, Lambda, TensorDock and Hyperstack publish hourly prices a developer can read without a quote. Amazon Web Services and Google Cloud publish instance families and add the usual cloud extras—storage, networking and idle disks. A marketplace floor is not a reserved contract. A reserved contract is not a spot interrupt.
Match the GPU to the job before you compare logos. Inference on a small model does not need an H100 node. Multi-GPU training cares about interconnect and whether the cluster stays up. CoreWeave, Crusoe and Paperspace sit in different parts of that spectrum: some lean toward large training estates, others toward notebooks and shorter jobs. Confirm VRAM, GPU count and the interrupt policy on the vendor page the day you buy.
Storage and egress often dominate short jobs. Pulling a large image or dataset once can cost more than the GPU hour. Check local disk size, network volume pricing and whether checkpoints land on object storage you already pay for elsewhere. If you only need a model behind an API key, compare LLM APIs first—running your own stack on a GPU makes sense when utilization stays high enough to beat token billing.
Community and spot capacity is cheaper because the machine can go away. That is fine when you checkpoint and can resume. It is a weak base for a latency SLA. On-demand or reserved capacity costs more and is the right comparison for a steady inference service. Do not mix those columns on a spreadsheet.
This ranking scores published pricing, a free credit when one exists, and self-serve access. It is not a tokens-per-second lab and it does not claim live stock. If you still need a plain CPU box for the control plane or a bastion, keep a VPS or cloud hosting account beside the GPU vendor instead of forcing every service onto an accelerator node.
Start with one paid hour on the image you will actually use. Driver, CUDA and the clock you were quoted should match the box you got. If the template fails to boot cleanly, that vendor is not “almost right”—it is the wrong shortlist for your stack.
Step 1
Compare that SKU. A provider that is cheap on an old card can be expensive on the card you need.
Step 2
Spot and community clouds are cheaper because the machine can go away. Fine if you checkpoint. Bad if you cannot.
Step 3
Datasets and image pulls can cost more than the GPU hour on a short job. Check both.
Step 4
Boot the image you will actually use. Driver, CUDA and the clock you were quoted should match the box you got.
A cheap hourly rate on a marketplace says nothing about availability or what a month of training costs. Estimate hours from your own runs.
Distributed training depends on the network between cards. Ask what interconnect the cluster uses before you compare price per GPU.
Interruptible capacity is fine for experiments with checkpoints. It is a risk for anything with a delivery date.
Training on one cloud with data on another adds transfer cost and latency.
Each provider is scored out of 100 against the other 9providers on this page, using only what is published on the vendor's site. The sponsor slot is not an input.
Full editorial policyMarketplace clouds often show the lowest hourly floor, with variable stock and interrupt risk. Hyperscalers publish a higher, more stable on-demand price. Compare the same GPU, not the smallest number on the homepage.
No. DrDeploy does not publish lab benchmarks. We rank public commercial terms and say so.
Yes if you pick on-demand capacity and a region that stays in stock. Community floors are a weak base for a latency promise.
When you want a key and a rate card instead of drivers and utilization math. See LLM APIs. Rent GPUs when you run the weights yourself and can keep the cards busy.
Often not. Local disk, network volumes and egress are separate lines. Price a real dataset pull, not only the GPU hour.
Not on marketplace or spot inventory. Treat public floors as commercial facts for the SKU, then confirm stock when you are ready to buy.