AWS EC2 GPU pricing 2026
AWS exposes GPUs through P (training-grade SXM) and G (inference / graphics) families plus the Capacity Blocks for ML reservation model. AWS cut P5 (H100) and P4 (A100) on-demand list rates by roughly 44 percent in mid-2025, moving them from the top of the category to mid-pack (H100 now $6.88/GPU-hr on-demand vs Azure $12.29 and GCP $11.06); the value is integration depth alongside the reservation-based Capacity Blocks model, not a rock-bottom headline rate.
Published rate card
Per-GPU per-hour rates pulled from the vendor pricing page linked above, Last verified July 2026. Rates exclude storage, egress, and any managed-service uplift. Reserved-capacity contracts typically improve on these rates.
| GPU | Configuration | Per GPU per hour |
|---|---|---|
| H100 80GB SXM (p5.48xlarge) | On-demand, 8x per node, per-GPU equiv | $6.880 |
| A100 80GB SXM (p4de.24xlarge) | On-demand, 8x per node, per-GPU equiv | $3.430 |
| A100 40GB SXM (p4d.24xlarge) | On-demand, per-GPU equiv | $2.740 |
| A100 (Capacity Blocks) | Pre-reserved, per-GPU | $1.475 |
| L4 (g6.xlarge) | On-demand | $0.805 |
| T4 (g4dn.xlarge) | On-demand | $0.526 |
Hidden costs to watch
- EBS storage, S3 egress, NAT Gateway throughput, and inter-AZ data transfer add 15-30% to a typical training run.
- EFA networking is included on supported instances but driver and topology tuning is on you.
- Savings Plans and Reserved Instances require 1-3 year commitments; Spot is volatile for multi-day training jobs.
- Capacity Blocks for ML lock in dated H100 or A100 reservations in advance at premium hourly rates.
What AWS EC2 is best for
Existing AWS-anchored workloads that need GPU access alongside VPC, IAM, S3, and SageMaker integration.
See the GPU cloud buying guideWorked example
Acme Vision Co. (illustrative example, not a real company) needs to train a 10 billion-parameter vision-language model on a fixed 8-GPU cluster for 30 days at 18 hours per day duty cycle.
At AWS EC2's cheapest published rate of $1.48 per GPU-hour, the run costs $6,372 for raw GPU compute, before storage, egress, and MLOps overhead. Add the typical 25 percent year-one uplift and the modelled spend is $7,965. Use the calculator on the homepage to model your own GPU class, cluster size, and duty cycle.
Last verified July 2026. AWS EC2 rates change frequently. Always obtain a vendor quote before purchase.
Visit AWS EC2 pricing page