Cloud GPU Compute Pricing: H100 / H200 per Hour (July 2026)
What it costs to run one NVIDIA H100 / H200 GPU per hour — the big three clouds (AWS / Azure / GCP) vs specialist GPU clouds (RunPod / Together / Lambda / CoreWeave), on-demand. The same GPU can cost 3–4x more depending on provider.
| Provider | Type | H100 ($/GPU/hr) | H200 ($/GPU/hr) |
|---|---|---|---|
| RunPod | GPU cloud | $2.69–2.99 | $4.39 |
| Together AI | GPU cloud | $3.99 | N/A |
| Lambda | GPU cloud | $3.29–4.29 | N/A |
| CoreWeave | GPU cloud | $6.16 | $6.31 |
| AWS | Hyperscaler | $6.88 | N/A (no public on-demand) |
| GCP | Hyperscaler | $11.06 | $10.60 |
| Azure | Hyperscaler | $12.29 * | $13.78 * |
Method & sources
Basis: per GPU, per hour, on-demand, USD. GPU clouds (RunPod/Together/Lambda/CoreWeave) are from official pricing pages; RunPod/Lambda show ranges across plans (Secure/Community, PCIe/SXM). Hyperscalers are whole-instance price ÷ GPU count: GCP from the official pricing page, us-central1, a3-highgpu-8g (8×H100 $88.49/hr) and a3-ultragpu-8g (8×H200 $84.81/hr) ÷8 (includes CPU/RAM/SSD; GCP's A3 GPUs are not separately priced); AWS H100 is p5.48xlarge ÷8 (us-east-1); AWS H200 has no public standard on-demand price (Capacity Blocks only), hence N/A; Azure is third-party-aggregated (ND H100/H200 v5; official page timed out, marked *). ⚠ Some pricing aggregators show GCP at ~$1.18/GPU ($9.46/instance), which is erroneous (identical across regions, far below official) and is not used. Prices vary by region and time; verify on official pages. Not an endorsement.
Source: https://cloud.google.com/products/compute/pricing/accelerator-optimized
Retrieved: 2026-07-22