AWS EC2 vs Azure: GPU cloud pricing compared
AWS and Azure are the two largest hyperscaler GPU clouds and the default shortlist for any organisation with a multi-region cloud estate. Both publish list prices for H100 (P5 on AWS, ND H100 v5 on Azure). AWS cut its P5 on-demand H100 list rate by roughly 44 percent in mid-2025, so its headline rate ($6.88/GPU-hr) now sits well below Azure's ($12.29/GPU-hr); both still price above most specialist clouds before negotiated discount. For H200, Azure publishes an ND H200 v5 on-demand rate of $10.60/GPU-hr equivalent on the Azure Retail Prices API, while AWS offers H200 through P5en on-demand (now available in more regions after its June 2025 update) and P5e via Capacity Blocks for ML but does not surface an H200 on-demand list rate as cleanly, so budget from the EC2 pricing page for your region. Integration depth (IAM, networking, MLOps surfaces) is the real reason teams pick one over a specialist.
Side-by-side
| Dimension | AWS EC2 | Azure |
|---|---|---|
| Headline H100 list rate (per GPU equiv) | $6.88/GPU-hr (p5.48xlarge on-demand, $55.04 node) | $12.29/GPU-hr (ND96isr H100 v5 on-demand) |
| H200 list rate (per GPU equiv) | P5en on-demand + P5e Capacity Blocks; on-demand list rate not cleanly published (check EC2 pricing page) | $10.60/GPU-hr (ND96isr H200 v5 on-demand, $84.80 node, Azure Retail Prices API) |
| Cheapest published H100 path | $1.475/GPU-hr (A100 Capacity Blocks; H100 Capacity Blocks priced separately) | $9.00/GPU-hr (ND H100 v5 Spot per-GPU equiv) |
| Reservation path | Savings Plans, Reserved Instances, Capacity Blocks for ML | Reserved Instances (1, 3 year), Savings Plans |
| Networking | EFA on supported instances | InfiniBand on ND H100 v5 |
| Integration depth | Deepest in the category (IAM, SageMaker, S3, EFA) | Tightest Microsoft estate integration in the category |
Your data, identity, and tooling are already on AWS. You need SageMaker, S3, EFA networking, and a Savings Plan that covers more than GPU compute. You can negotiate a Spot or Capacity Blocks strategy.
Your estate is Microsoft-anchored (Entra ID, Azure AI Foundry, Synapse). You want ND H100 v5 with InfiniBand on the same control plane as the rest of your Azure infrastructure and you can negotiate Reserved Instances or a Savings Plan.
Worked example
Acme MLOps Co. (illustrative example, not a real company) needs an 8-GPU H100 cluster for 30 days at 18 hours per day (4,320 GPU-hours). At AWS EC2's published H100 rate ($6.880/GPU-hr, H100 80GB SXM (p5.48xlarge) On-demand, 8x per node, per-GPU equiv) that is roughly $29,722; at Azure's published H100 rate ($12.290/GPU-hr, H100 80GB (ND96isr H100 v5) On-demand, per-GPU equiv), roughly $53,093 for raw GPU compute, before storage, egress, and MLOps overhead.
Last verified July 2026.