Short answer
The cheapest cloud GPUs in November 2026 come from specialist GPU clouds and marketplaces, not hyperscalers. The lowest on-demand prices we track are about $1.80 per hour for an H100, $2.09 for an H200, $4.11 for a B200 and $0.89 for an A100. AWS, Azure, Oracle and Google Cloud charge roughly two to three times more on demand. The cheapest listings are often limited in stock or features, so compare reliability as well as price. Our live GPU price tracker updates every day.
Cheapest GPUs by model
Lowest and median on-demand prices per GPU-hour from our tracker in mid-November 2026:
| GPU | Cheapest listing | Median price | Providers tracked |
|---|---|---|---|
| A100 | $0.89 (Jarvislabs, 40GB) | about $2.00 | 15 |
| H100 | $1.80 (HyperAI) | about $3.92 | 22 |
| H200 | $2.09 (Beam) | about $4.92 | 14 |
| B200 | $4.11 (Beam) | about $7.34 | 14 |
For smaller jobs, consumer and workstation GPUs are much cheaper: typical rates are about $0.47 per hour for an RTX 4090 and $0.72 for an RTX 5090.
The cheapest H100 providers
| Provider | H100 price per GPU-hour |
|---|---|
| HyperAI | $1.80 |
| Hyperstack | $2.50 (PCIe) |
| UpCloud | $2.50 (SXM) |
| Civo | $2.99 |
| OVHcloud | $2.99 |
| Hyperbolic | $3.19 |
| Scaleway | $3.19 |
| Thunder Compute | $3.20 |
| Lambda | $3.29 |
Full lists for every GPU, filters by provider type and a cost calculator are on the GPU cloud pricing tracker.
Why the cheapest GPU is not always the cheapest option
- Availability: the lowest prices are often out of stock or limited to one GPU.
- PCIe vs SXM: cheap H100 listings are often PCIe cards, which have lower bandwidth and no NVLink between GPUs. That matters for multi-GPU training; see NVLink and NVSwitch explained.
- Hidden costs: storage, data transfer and idle time can add a lot.
- Reliability and support: marketplaces and spot capacity can be interrupted.
- Performance per dollar: a pricier GPU can be cheaper per job if it finishes much faster.
Types of GPU cloud providers
| Type | Examples | Typical H100 price | Best for |
|---|---|---|---|
| Marketplaces | Vast.ai | Lowest, varies by host | Experiments, flexible jobs |
| Serverless GPU | Modal, Beam | Pay per second | Bursty inference |
| Specialist GPU clouds | Lambda, RunPod, Hyperstack, Nebius | $2.50 to $4.50 | Training and serving |
| Hyperscalers | AWS, Google Cloud, Azure, Oracle | $6.88 to $11.07 | Enterprise integration and compliance |
We compare the big three in AWS vs Google Cloud vs Azure GPU pricing.
How to cut your GPU cloud bill
- Right-size the GPU: use our guide on how much VRAM an LLM needs so you do not pay for memory you do not use.
- Quantize models: 4-bit or FP8 models need fewer and cheaper GPUs; see LLM quantization explained.
- Use spot or interruptible capacity for jobs that can restart from checkpoints.
- Commit when usage is steady: reserved and contract prices are much lower than on demand.
- Shut down idle machines and keep data close to the GPUs to avoid transfer fees.
Frequently asked questions
What is the cheapest cloud GPU for AI?
For data center GPUs, A100s start under $1 per hour and H100s from about $1.80 per hour on demand in November 2026. For small models, rented RTX 4090 and RTX 5090 cards are cheaper still.
Which cloud is cheapest for H100?
Specialist GPU clouds such as HyperAI, Hyperstack, UpCloud, Civo and OVHcloud listed the lowest H100 prices we tracked, at $1.80 to $2.99 per hour. Prices change daily, so check the live tracker.
Is AWS more expensive than other GPU clouds?
On demand, yes. AWS lists an H100 at about $6.88 per GPU-hour, around 1.8 times the market median, though reservations and spot capacity reduce the gap.
Sources and further reading
- TensorCore.si, Live GPU cloud pricing tracker (updated daily)
- GetDeploying, Cheapest cloud GPUs (November 2026)
- GPU Per Hour, Cloud GPU price index




Leave a Reply