33 accelerator models · 199 instance types
Cloud GPU pricing
Compare the price of every cloud GPU instance by GPU model: on-demand and spot price per GPU-hour, GPU memory and the cheapest region on AWS, Azure, Google Cloud, Akamai and Vultr.
GPUs
lowest Linux on-demand price per GPU-hour across clouds and regions
AI accelerators
AWS Trainium and Inferentia, Intel Gaudi, Qualcomm Cloud AI
| Model | Memory | Architecture | Clouds | Instances | From / GPU-hr | Cheapest |
|---|---|---|---|---|---|---|
| AWS Trainium2 | 96 GB | – | 1 | $6.447 | AWS trn2.48xlarge | |
| AWS Inferentia2 | 32 GB | – | 4 | $0.7582 | AWS inf2.xlarge | |
| Qualcomm Cloud AI 100 | 16 GB | – | 1 | $1.115 | AWS dl2q.24xlarge | |
| AWS Trainium | 32 GB | – | 3 | $1.344 | AWS trn1.2xlarge | |
| Intel Habana Gaudi | 32 GB | – | 1 | $1.639 | AWS dl1.24xlarge | |
| AWS Inferentia | 8 GB | – | 4 | $0.228 | AWS inf1.xlarge |
FAQ
What is the cheapest cloud H100?
AWS p5.4xlarge at $6.88 per H100-hour on demand in US East (N. Virginia). See the H100 page for every option, including spot.
How is price per GPU-hour calculated?
The instance's hourly Linux on-demand price in its cheapest commercial region divided by the number of GPUs it has. The instance also includes CPUs, memory and local storage, so compare total price too when your workload needs those.
Why are some GPUs fractional?
Some providers sell slices of a GPU (for example Azure NVads A10 v5 with 1/6 of an A10, or AWS g6f with a fraction of an L4) for virtual desktops and light inference. The GPU count and memory are shown as sold.