When the premium is worth it
Memory-bound work — long contexts, large KV caches, bigger batches per GPU — gets more from the 141 GB than from a faster chip. Compute-bound training often does not.
GPU Pricing
Sourced H200 on-demand rates across providers — list prices, H100-equivalent rates, and dated public price-page links.
H200 on-demand capacity is sourced from 5 provider price pages, each linked and dated below. The H100e column normalizes every rate to an H100-equivalent hour, so the H200 premium reads against the H100 table on one axis.
Lowest to highest sourced public list price, as of Jun 11, 2026
Public price pages with an H200 on-demand rate that cleared editorial review
The prior-generation read the H200 premium is usually priced against
Sourced rates
5 sourced provider rows, lowest list price first. Every row links the public price page it was observed on.
| Provider | Region | List $/GPU-hr | H100e $/hr | Availability | No change in | Source |
|---|---|---|---|---|---|---|
| Hyperstack | Multi-region | $3.50 | $3.04 | Available | 1 day | Open price page |
| Crusoe | Multi-region | $4.29 | $3.73 | Available | 0 days | Open price page |
| RunPod Secure | Multi-region | $4.39 | $3.82 | Available | 0 days | Open price page |
| CoreWeave | us | $6.30 | $5.48 | Available | 1 day | Open price page |
| Oracle | us-ashburn-1 | $10.00 | $8.70 | Available Limited | 26 days | Open price page |
Public on-demand list prices from provider price pages — not negotiated quotes, reserved terms, or spot rates. Each row carries its own source link and price date. How we source, label, and date evidence: methodology.
See every reviewed provider row