Cheapest Cloud GPU Providers: September 2026 Price Comparison
Published · Updated · NVGPU
Compare a dated GPU rental snapshot, minimum instance costs and the conditions that can change which provider is cheapest.
There is no single cheapest GPU provider for every workload. The answer changes with the GPU model, memory, region, number of GPUs and whether the job can tolerate interruption.
This comparison was reviewed on September 26, 2026. It uses NVGPU's catalog rather than the unsupported personal experiences and fixed price ranges in the earlier version.
What our catalog covers
We currently track AWS, Google Cloud, Azure, Oracle Cloud, Lambda, RunPod, DataCrunch/Verda, Nebius and CloudRift. The feed publishes each provider independently. Eight provider snapshots are dated September 26; Google Cloud's snapshot is dated September 24.
A source date is not a promise that capacity is available.
Lowest listed on-demand examples
| GPU | Provider | Region | USD per GPU-hour | Minimum instance | USD per instance-hour |
|---|---|---|---|---|---|
| RTX 4090 24GB | RunPod | FR | 0.340 | 1 GPU | 0.340 |
| L40S 48GB | RunPod | EU-NL-1 | 1.090 | 1 GPU | 1.090 |
| A100 40GB | DataCrunch/Verda | FIN-01 | 1.282 | 1 GPU | 1.282 |
| A100 80GB | RunPod | US | 1.390 | 1 GPU | 1.390 |
| H100 80GB | RunPod | US-KS-2 | 2.890 | 1 GPU | 2.890 |
| H200 141GB | Nebius | us-central1 | 4.500 | 1 GPU | 4.500 |
| B200 180GB | Lambda | asia-northeast-1 | 6.690 | 8 GPUs | 53.520 |
This is a September 26 snapshot. It excludes spot pricing and providers outside our automated coverage. It does not establish which provider has the best uptime, support or performance.
For prices after this snapshot, use the GPU price board and provider comparisons. Confirm the final quote with the provider.
Why the cheapest headline can be misleading
Minimum GPU count matters. A per-GPU rate for an eight-GPU instance is not the price of a single-GPU rental. Use the full instance cost for your budget.
Memory matters. An A100 40GB offer is not equivalent to an A100 80GB offer. A partition exposing less VRAM can be unsuitable even when it shares a familiar GPU name.
Region matters. A low rate in another continent may increase transfer time or conflict with workload requirements. Compare a usable region rather than the global minimum.
Spot and on-demand need separate comparisons. The spot discount is only valuable if the job can checkpoint, recover and still meet its deadline. Include lost work and restart time in your estimate.
Providers outside the automated price board
You can also evaluate Vast.ai, DigitalOcean, Massed Compute, Thunder Compute, Database Mart, CUDO Compute, Vultr, Hyperstack and TensorDock. This is a research list, not a verified cheapest-provider ranking. Check each provider's current pricing, billing period, available models and service terms directly.
We have not benchmarked those services for this article. Explore more options in the provider directory.
Calculate the cost of your job
Use this planning formula:
Job cost = full-instance hourly rate x billed hours + storage + transfer + other charges.
For example, 20 billed hours on the snapshot's single H100 at $2.89/hour is $57.80 in compute. Keeping the same instance active for 730 hours would be $2,109.70 before extras. Neither figure predicts how quickly your model will finish.
A short representative run is more useful than choosing from hourly prices alone. Record initialization, data download, processing and shutdown time. Check whether storage continues billing after compute stops.
A practical selection process
- Set the GPU-memory and software requirements.
- Choose an acceptable region and minimum instance size.
- Compare on-demand offers first, then assess spot separately.
- Check the provider's live quote, available capacity and extra charges.
- Run a small job and compare measured cost before scaling.
Start with H100 rental prices, A100 rental prices or L40S rental prices.