Hyperscaler
Amazon Web Services
The biggest cloud. The deepest catalog, the highest list prices, the most enterprise glue.
GPU capacity spans P (training) and G (inference/graphics) instance families, plus Trainium/Inferentia ASICs. On-demand list prices run far above neoclouds, but Savings Plans, spot, and Capacity Blocks close much of the gap. The reason to be here is everything around the GPU: VPC, IAM, S3, EKS, compliance.
GPU pricing on AWS
On-demand list per GPU-hour, last reviewed 2026-09. Methodology.
| GPU | $/GPU-hr | Notes |
|---|---|---|
| T4 | $0.53 | g4dn.xlarge |
| L4 | $0.98 | g6.xlarge |
| A10 | $1.01 | g5.xlarge (A10G) |
| L40S | $1.96 | g6e.xlarge |
| V100 | $3.06 | p3.2xlarge, legacy pricing |
| A100 80GB | $4.10 | p4de.24xlarge ÷ 8 |
| H100 SXM | $4.90 | p5.48xlarge ÷ 8, post-2025 price cut |
| H200 | $6.20 | p5e.48xlarge ÷ 8 |
| B200 | $8.65 | p6-b200.48xlarge ÷ 8 |
Strengths
- Enterprise ecosystem and compliance
- Capacity Blocks for guaranteed short-term clusters
- Spot fleet depth
Watch for
- On-demand list prices 2-4× neoclouds
- GPU quota requests can take days
- Egress fees