GPU rental prices · checked October 2026
Rent a GPU by the Hour: Compare Cloud GPU Rental Prices
You can rent a GPU in minutes and pay only for the hours you use — from about $0.35/hr for an RTX 4090, $0.89/hr for an A100, and $1.80/hr for an H100 on specialized clouds (spot prices go lower). Pick a GPU with enough VRAM for your model, choose a provider, launch an instance, and stop it when you’re done.
GPU rent prices per hour (October 2026)
Lowest on-demand rates we found on specialized GPU clouds, with typical market medians. Hyperscalers (AWS, Google Cloud, Azure) usually charge 2–5x more for the same GPU.
| GPU | VRAM | Rent from | Typical median | Best for |
|---|---|---|---|---|
| RTX 4090 | 24 GB | $0.35/hr | ~$0.50/hr | Stable Diffusion, 7B–13B LLMs |
| RTX 5090 | 32 GB | $0.35/hr | ~$0.70/hr | Image/video gen, fine-tuning small models |
| A100 80GB | 80 GB | $0.89/hr | ~$1.80/hr | Budget training, fine-tuning |
| L40S | 48 GB | $0.96/hr | ~$1.60/hr | Inference, rendering |
| H100 | 80 GB | $1.80/hr | ~$3.29/hr | Training and serving large models |
| H200 | 141 GB | $3.14/hr | ~$4.80/hr | Long-context, memory-bound inference |
| B200 | 180 GB | $3.49/hr | ~$8.19/hr | Frontier-scale training |
Spot (interruptible) H100s start near $0.94/hr. Prices move weekly — see the full cost to rent a GPU breakdown or compare H100 vs A100 before you commit.
How to rent a GPU in 4 steps
-
Size the GPU to your workload
VRAM decides what fits. A 7B model runs on 24 GB; a 70B model needs ~40 GB quantized or 140 GB at full precision. Check the VRAM guide first — it’s the step that saves the most money.
-
Pick a GPU rental provider
Marketplaces are cheapest, specialized clouds balance price and reliability, hyperscalers cost most. Compare them in our best cloud GPU providers guide.
-
Choose on-demand or spot
On-demand runs until you stop it. Spot is 50–80% cheaper but can be interrupted — use it for training that saves checkpoints.
-
Launch, connect, and stop when done
Start from a PyTorch or vLLM template, connect via SSH or Jupyter, and shut the instance down after the job. Idle GPUs still bill. Full walkthrough: how to rent a GPU in the cloud.
Where to rent a GPU: best GPU rental providers
The right provider depends on whether you care most about price, reliability, or scale.
| Provider | Type | Billing | Best for |
|---|---|---|---|
| RunPod | Specialized cloud | Per second | Best all-round balance, serverless |
| Vast.ai | Marketplace | Per second | Lowest prices, interruptible jobs |
| Lambda | Specialized cloud | Per hour | ML support, A100/H100 clusters |
| CoreWeave | Enterprise neocloud | On-demand / reserved | Large-scale training |
| AWS / Google Cloud / Azure | Hyperscaler | Per second / hour | Compliance and existing cloud stack |
Head-to-head: RunPod vs Vast.ai.
What GPU rent really costs (worked example)
Renting one H100 for 100 hours costs about $180 at the lowest on-demand rate ($1.80/hr), around $330 at the market median, and $600 or more on a hyperscaler. The same job on spot can drop under $100.
Budget for the extras the hourly rate leaves out:
- Storage — persistent volumes often bill even while the GPU is stopped.
- Egress — moving data out can cost $0.08–$0.12/GB on hyperscalers.
- Idle time — a forgotten instance bills around the clock.
- CPU, RAM, and networking — some providers price these separately.
Rent a GPU or buy one?
Renting wins if you use a GPU less than about 4 hours a day or need data-center cards like the H100. Buying makes sense only if you run a GPU most of the day for well over a year. An RTX 4090 workstation costs roughly the same as several years of part-time GPU rent — before electricity.
| Your usage | Better choice | Why |
|---|---|---|
| Under 20 hrs/week | Rent | No upfront cost, no idle hardware |
| Bursty training jobs | Rent (spot) | Big GPUs for days, not years |
| Daily local AI + gaming | Buy | One card covers both |
| 24/7 for 18+ months | Buy | Hardware pays for itself |
New to this? Start with what a cloud GPU is.
Can you rent a GPU for gaming?
Yes, but cloud GPU rental is built for compute, not games. For gaming, a cloud gaming service (such as GeForce NOW) streams games from rented RTX hardware with low latency and a monthly fee. Renting a raw GPU server and streaming a Windows desktop works too, but it’s usually slower to set up and costs more per hour of play. GPU rental shines for AI, rendering, and video work.
Explore GPU rental guides
- Cost to rent a GPU — hourly and monthly prices by GPU
- Best GPU rental providers — compared on price and reliability
- Cloud GPU models — RTX, A100, H100, H200, B200 explained
- How to rent a GPU — step-by-step setup
- VRAM for LLMs — how much GPU memory you need
- H100 rental prices — every provider compared
GPU rental FAQs
How much does it cost to rent a GPU?
Renting a GPU costs from about $0.35/hr for an RTX 4090, $0.89/hr for an A100, and $1.80/hr for an H100 on specialized clouds as of October 2026. Spot instances are 50–80% cheaper, while AWS, Google Cloud, and Azure typically charge 2–5x more for the same GPU.
Can I rent a GPU by the hour?
Yes. Most GPU rental providers bill per hour, and RunPod and Vast.ai bill per second. You start an instance, use it, and stop it — you’re charged only while it runs, plus any storage you keep.
Where is the cheapest place to rent a GPU?
GPU marketplaces such as Vast.ai and spot instances on specialized clouds are usually the cheapest places to rent a GPU. They trade some reliability for price, so save checkpoints if your job runs for hours.
Is it cheaper to rent or buy a GPU?
Renting is cheaper if you use a GPU less than about 4 hours a day or need data-center GPUs like the H100. Buying pays off only when you run a GPU most of the day for well over a year.
Can you rent a GPU for gaming?
You can, but cloud gaming services like GeForce NOW are usually the better way to play on rented hardware. Raw cloud GPU rental is designed for AI, rendering, and other compute work.
Do I pay for a rented GPU when it’s idle?
Yes, if the instance is still running. Stop or terminate it when your job finishes. Persistent storage may keep billing even after the GPU is stopped.