Skip to content
,

AWS vs Google Cloud vs Azure GPU Pricing (2026): H100, H200, B200 and A100 Compared

How much H100, H200, B200 and A100 GPUs cost per hour on AWS, Google Cloud, Azure and Oracle in 2026, why hyperscalers charge more than GPU clouds, and how reservations, Capacity Blocks and spot pricing change the picture.

4 min read

AWS vs Google Cloud vs Azure GPU Pricing — cover image from TensorCore.si

Short answer

On demand, the big clouds charge roughly $6.88 to $11.07 per H100 GPU-hour in November 2026: AWS about $6.88, Azure about $6.98 (H100 NVL), Oracle about $10.00 and Google Cloud about $11.07. That is around two to three times the median price at specialist GPU clouds. Hyperscalers close much of the gap with reservations, savings plans, AWS Capacity Blocks and spot capacity. Compare every provider on our live GPU price tracker.

On-demand GPU prices on the big clouds

Prices per GPU-hour (multi-GPU instances divided by GPU count) from our tracker in mid-November 2026:

GPUAWSAzureGoogle CloudOracle Cloud
H100$6.88 (p5.4xlarge)$6.98 (NCads H100 v5, NVL)$11.07 (a3-highgpu-8g)$10.00 (BM.GPU.H100.8)
H200$7.91 (p5en.48xlarge)$10.60 (ND96isr H200 v5)$10.85 (a3-ultragpu-8g, reservation)$10.00 (BM.GPU.H200.8)
B200$14.24 (p6-b200.48xlarge)not listed on demandnot listed on demand$14.00 (8x B200)
A100$2.74 (p4d.24xlarge, 40GB)$3.40 (ND96asr A100 v4, 40GB)$3.72 (a2-megagpu-16g, 40GB)$3.05 (BM.GPU4.8, 40GB)

For comparison, the median H100 price across all 22 providers we track is about $3.92, and the cheapest listing is $1.80. Hyperscaler GPU instances are also mostly sold as full 8-GPU machines: AWS's p5.48xlarge with eight H100s costs about $55.04 per hour on demand.

Why hyperscalers cost more

  • Enterprise features: security, compliance certifications, private networking, identity management and support.
  • Integration: GPUs sit next to your existing storage, databases and data pipelines, which can save transfer costs and engineering time.
  • Networking at scale: high-speed interconnects (such as EFA on AWS) for large training clusters.
  • Business model: hyperscalers expect large customers to commit to long-term spending in exchange for big discounts.

How to pay less on AWS, Azure and Google Cloud

Reservations and savings plans

Committing to one or three years lowers prices substantially compared with on-demand rates. Industry estimates commonly put one-year commitments at around 30 to 40% off and three-year commitments at around 50 to 60% off for GPU instances, but exact discounts vary by instance and negotiation.

AWS EC2 Capacity Blocks

Capacity Blocks let you reserve GPU instances for a set period (days or weeks) at a fixed price. They became more expensive during 2026: AWS raised Capacity Block prices several times, including increases in January, April, July and October. By October 2026, published rates were about $5.97 per H100 GPU-hour (P5), $7.90 for H200 (P5en) and $14.21 for B200 (P6-B200).

Spot and preemptible capacity

Spot instances can cost far less than on-demand prices, but they can be reclaimed at short notice and GPU spot capacity is often scarce. They suit training jobs that save checkpoints frequently and batch inference that can pause.

AWS vs Azure vs Google Cloud: which is cheapest?

  • H100: AWS and Azure are the cheapest hyperscalers on demand, at about $6.88 and $6.98 per GPU-hour; Google Cloud's 8-GPU A3 instance works out to about $11.07.
  • H200: AWS is the cheapest of the big clouds at about $7.91; Oracle, Azure and Google Cloud are around $10 to $11.
  • B200: AWS and Oracle list on-demand B200 at about $14 per GPU-hour.
  • A100: AWS is the cheapest for A100 40GB at about $2.74.

In practice, the best deal depends on your existing cloud, the discounts you can negotiate and whether capacity is available in your region.

Hyperscaler or GPU cloud?

Choose a hyperscaler if you need tight integration with existing infrastructure, strict compliance or enterprise support. Choose a specialist GPU cloud if price per GPU-hour matters most; see cheapest cloud GPU providers. Many teams use both: GPU clouds for training and experiments, hyperscalers for production.

Frequently asked questions

How much does an H100 cost on AWS?

About $6.88 per GPU-hour on demand (p5.4xlarge with one H100), or about $55.04 per hour for the 8-GPU p5.48xlarge in US East, as of November 2026.

Is Google Cloud more expensive than AWS for GPUs?

For on-demand H100s, yes in our data: about $11.07 per GPU-hour on Google Cloud versus about $6.88 on AWS. Committed-use discounts change this comparison.

Do hyperscalers offer B200 GPUs?

Yes. AWS offers P6-B200 instances and Oracle offers B200 bare-metal servers on demand; other clouds offer Blackwell mainly through reservations.

Sources and further reading

Think you understand it now? Prove it.

Solve real Tensor Core math in the TensorCore.si Challenge and earn a certificate with your name.

Take the challenge

Leave a Reply

Your email address will not be published. Required fields are marked *