NVIDIA GPU generations & rental pricing
Every generation from Ampere to the announced Rubin / Feynman roadmap, with on-demand per-GPU-hour rates across 10 clouds and marketplaces. Marketplace figures are published floor prices; reserved contracts run 30–55% below list. Verified 2026-08-13.
Not (or no longer) on the rental market
CoreWeave publishes per-node prices; per-GPU shown = node price ÷ GPU count. Fireworks on-demand deployment prices step up on Sep 1 2026 (H100/H200 +$1, B200/B300 +$3, GB300 +$2 per hour). Vast.ai 'from' prices are the lowest live marketplace offer (often interruptible); stable on-demand typically runs 1.5–3× higher. DataCrunch rebranded to Verda in 2026; spot there is a flat 50% of on-demand.