What a GPU-hour costs
2,476 prices for 28 chips from 16 clouds and neoclouds across 92 regions, per GPU per hour. H100 on-demand from $1.99.
15 providers for H100, cheapest first
| Provider | On-demand | |||||
|---|---|---|---|---|---|---|
spot $0.55 · H100 SXM (marketplace low) | $1.99 | $0.55 | – | – | 1 | H100 SXM (marketplace low) |
H100 PCIe (Community Cloud) | $1.99 | – | – | – | 1 | H100 PCIe (Community Cloud) |
vultr vbm-112c-2048gb-8-h100-gpu (NVIDIA H100) | $2.99 | – | – | – | 1 | vbm-112c-2048gb-8-h100-gpu (NVIDIA H100) |
scaleway 2 regions | $3.26 | – | – | – | 2 | H100-1-80G, fr-par-2 |
NVIDIA H100 PCIe (80 GB) | $3.29 | – | – | $5.54contract | 1 | NVIDIA H100 PCIe (80 GB) |
datacrunch spot $1.73 · 1H100.80S.30V (1x H100 SXM5 80GB) | $3.45 | $1.73 | – | – | 1 | 1H100.80S.30V (1x H100 SXM5 80GB) |
spot $0.79 · NVIDIA HGX H100 | $3.85 | $0.79 | – | – | 1 | NVIDIA HGX H100 |
NVIDIA H100 80GB HGX | $3.90 | – | – | – | 1 | NVIDIA H100 80GB HGX |
spot $1.99 · NVIDIA HGX H100 | $3.99 | $1.99 | – | $3.19contract | 1 | NVIDIA HGX H100 |
NVIDIA HGX H100 | $4.41 | – | $3.26 | – | 1 | NVIDIA HGX H100 |
spot $2.46 · NVIDIA HGX H100 (8 GPUs) | $6.16 | $2.46 | – | – | 1 | NVIDIA HGX H100 (8 GPUs) |
 AWSspot $3.91 · 15 regions | $6.88 | $3.91 | – | – | 15 | p5.48xlarge, us-east-1 |
spot $1.29 · 35 regions | $6.98 | $1.29 | $5.24 | $3.84 | 35 | Standard_NC40ads_H100_v5, eastus |
Compute - GPU - H100 (B98415) | $10.00 | – | – | – | 1 | Compute - GPU - H100 (B98415) |
spot $6.62 · Iowa (us-central1) | $11.06 | $6.62 | $7.67 | $4.86 | 1 | a3-highgpu-8g, us-central1 |
Every H100 instance and region
Questions
How is the per-GPU price worked out?
Instance prices are divided by the number of GPUs in the instance (8 for an H100 or B200 server, 4 for a GB200 VM on Azure). TPU and Trainium prices are per chip-hour. Neoclouds that publish per-GPU prices are shown as published.
What do 1-year and 3-year mean?
Reservations (Azure), committed-use discounts (Google Cloud) and 12-month plans (DigitalOcean), all turned into an hourly rate. Contract means a published cluster price for a fixed term (Lambda, Together).
Why is spot so much cheaper?
Spot and preemptible capacity can be taken back at short notice. AWS and Azure spot prices come from their feeds and change often; neocloud spot prices are "from" prices.
How often is it updated?
List prices every Monday; AWS spot and the Vast.ai marketplace every day. Every change is kept: the Changes page and /api/changes show old and new prices with the dates.
Which regions?
AWS and Azure prices come per region; Google Cloud GPU prices are for Iowa (us-central1) and TPU prices for every listed region; Oracle list prices are the same everywhere; neoclouds publish one price.