Cloud GPU On-Demand Price Index (H100, A100, L40S & more)

What does an hour of GPU compute actually cost? This index tracks the on-demand list price per GPU-hour for the main data-centre NVIDIA accelerators — H100, H200, A100 80GB, L40S, L4 and the RTX 4090 — across the big-three hyperscalers (AWS, GCP, Azure), the neoclouds (Lambda, RunPod) and the Vast.ai marketplace, with every priced cell sourced to a vendor page and dated.

It exists to put a number on a claim you see everywhere in decentralised-compute marketing: “up to 60–85% cheaper than AWS.” Against an owned baseline, that claim mostly holds — an NVIDIA H100 lists at roughly $6.88–$12.29 per GPU-hour on-demand across AWS, GCP and Azure, versus under $3 on a neocloud and from ~$1.73 on a marketplace. Where you rent matters far more than most buyers assume.

Read it with care. These are on-demand list prices — not spot, and not 1- or 3-year committed rates, which run much lower. Hyperscaler prices are per-instance (an AWS p5.48xlarge bundles eight H100s) divided down to a per-GPU figure. Marketplace prices float with host supply. Prices change constantly, so treat this as a dated snapshot and verify at the linked source before you commit.

Deciding whether to rent or own? Pair this with the AI GPU spec database (what each GPU delivers for the money) and the DePIN compute networks directory (the decentralised alternatives), part of D-Central’s sovereign AI work.