GPU Rental Pricing
The same GPU costs 4–5× more depending on where you rent it. Compare real on-demand rates and estimate your monthly bill. Prices checked 2026-08-02.
- Per day
- $—
- GPU-hours / month
- —
- 24/7 equivalent
- $—
- At median market rate
- $—
On-demand rates by provider
NVIDIA H100 (80GB) · 80 GB HBM3 · Training and fine-tuning large models; high-throughput inference
| Provider | Type | From ($/hr) | ≈ Monthly (8h × 22d) |
|---|---|---|---|
| Vast.ai (peer marketplace — verify host reliability) | Marketplace | $1.49 | $262 |
| Spheron (spot pricing) | Marketplace | $2.01 | $354 |
| RunPod | GPU cloud | $2.39 | $421 |
| Lambda | GPU cloud | $2.49 | $438 |
| Cudo Compute | GPU cloud | $2.99 | $526 |
| CoreWeave (enterprise clusters) | GPU cloud | $4.25 | $748 |
| Google Cloud | Hyperscaler | $6.20 | $1,091 |
| Azure | Hyperscaler | $6.98 | $1,228 |
NVIDIA A100 (80GB) · 80 GB HBM2e · Cost-effective training and inference for mid-size models
| Provider | Type | From ($/hr) | ≈ Monthly (8h × 22d) |
|---|---|---|---|
| Vast.ai | Marketplace | $0.89 | $157 |
| RunPod | GPU cloud | $1.19 | $209 |
| Lambda | GPU cloud | $1.29 | $227 |
| Cudo Compute | GPU cloud | $1.79 | $315 |
| Google Cloud | Hyperscaler | $2.90 | $510 |
| Azure | Hyperscaler | $3.40 | $598 |
NVIDIA RTX 4090 (24GB) · 24 GB GDDR6X · Budget inference, Stable Diffusion/Flux image generation, small fine-tunes
| Provider | Type | From ($/hr) | ≈ Monthly (8h × 22d) |
|---|---|---|---|
| Vast.ai | Marketplace | $0.35 | $62 |
| TensorDock | Marketplace | $0.37 | $65 |
| RunPod | GPU cloud | $0.44 | $77 |
| Cudo Compute | GPU cloud | $0.49 | $86 |
NVIDIA L40S (48GB) · 48 GB GDDR6 · Inference with larger context/models than 24GB cards allow; video workloads
| Provider | Type | From ($/hr) | ≈ Monthly (8h × 22d) |
|---|---|---|---|
| Vast.ai | Marketplace | $0.67 | $118 |
| RunPod | GPU cloud | $0.86 | $151 |
| Cudo Compute | GPU cloud | $0.99 | $174 |
Rates are public on-demand starting prices as of 2026-08-02; actual price varies by region, configuration and availability. Some links may earn us a commission at no cost to you.
How to pick a tier (and save 60–80%)
- Marketplaces (Vast.ai, TensorDock, Spheron): cheapest by far; hardware quality and uptime vary by host. Ideal for experiments, checkpointed training, and batch jobs.
- GPU clouds (RunPod, Lambda, Cudo, CoreWeave): the sweet spot for most teams — reliable, fast provisioning, 40–60% below hyperscaler rates.
- Hyperscalers (AWS, Google Cloud, Azure): pay the premium only when you need their compliance, networking, or existing-contract credits.
Two universal savings: spot/interruptible instances (30–70% off for fault-tolerant work) and right-sizing the card — an RTX 4090 at $0.35/hr handles most image generation and small-model inference that people reflexively rent H100s for. And before renting at all, run your workload through our API cost calculator — below ~40–60% utilization, the API is almost always cheaper than any rental.
Frequently asked questions
› How much does it cost to rent an H100 GPU?
As of 2026-08-02, H100 rentals range from about $1.49/hr on peer marketplaces (Vast.ai) through $2.40–$3.00/hr on GPU clouds (RunPod, Lambda) up to $6–$7/hr on hyperscalers (Azure, Google Cloud). That is a 4–5× spread for the same silicon — provider choice matters more than any discount program.
› Why are hyperscalers so much more expensive than GPU marketplaces?
You pay for the surrounding platform: enterprise SLAs, compliance certifications, integrated storage and networking, and support. A solo developer fine-tuning a model on a marketplace GPU gets none of that — and usually needs none of it. Teams with compliance requirements or production SLAs often do need it, which is why both markets exist.
› Is it cheaper to rent a GPU or use an API?
For most inference workloads, APIs win: providers batch thousands of users on the same hardware, and you pay nothing for idle time. Renting wins when you have sustained high utilization (the GPU busy most of the day), need an open-weight model the APIs do not serve, fine-tune frequently, or have data that cannot leave your infrastructure. The break-even is roughly 40–60% sustained utilization.
› What is the difference between on-demand and spot/interruptible pricing?
Spot instances are idle capacity offered 30–70% cheaper, with the catch that the provider can reclaim them with little notice. They are excellent for fault-tolerant training with checkpointing and for batch inference, and a bad idea for anything serving live traffic.
› How many GPU hours does fine-tuning actually take?
A LoRA fine-tune of a 7–8B model on a few thousand examples typically takes 2–8 H100-hours ($5–$25 at GPU-cloud rates). Full fine-tuning of the same model runs 50–200 GPU-hours. Large-model full fine-tunes scale into thousands of GPU-hours — which is why most teams stick to LoRA or use provider fine-tuning APIs.