On 2026-09-26 our table holds 581 rate quotes from 51 clouds across 40 GPU models. Ninety-six of those quotes are one day old: thirteen providers were read for the first time on the evening of 2026-09-25 — Akamai Cloud (Linode), EmpirioLabs, Civo, Daytona, Hugging Face Inference Endpoints, HexGrid, Lyceum, Canopy Wave, Atlas Cloud, Neysa, Nova Cloud, Enverge and EcoHash. Between them they list an on-demand rate card on 29 of the 40 cards. On 26 of those 29, the cheapest rate card belongs to a seller that was already in the table.
The cheapest H100 SXM is still $1.79
Twenty-five clouds list an on-demand H100 SXM on 2026-09-26. The median rate card — each provider counted once, at its own cheapest on-demand rate, community capacity and marketplace floors left out — is $3.61 per GPU-hour, over a 6.2x spread from $1.79 at Cudo Compute in no-kristiansand-1 to $11.06 at Microsoft Azure in eastus.
Five of the thirteen list the card, and all five landed in the middle of it: Atlas Cloud at $2.95, Civo and HexGrid at $2.99, Canopy Wave at $3.25 (priced per GPU on a two-GPU machine) and Neysa at $4.39, which its page prints as a from rate. The cheapest of the twenty-five, Cudo’s $1.79, has read the same in every one of our hourly reads since 2026-09-07, the day we first read it.
Across all 29 cards the thirteen quote, 35 of their 78 rate cards sit below their card’s median, 36 above it and 7 exactly on it.
Three cards where the cheapest did change
The GeForce RTX 3090 has two rate cards in the table: EmpirioLabs at $0.40 and RunPod’s Secure Cloud at $0.50, a median of $0.45. On the RTX 5090, HexGrid’s $0.55 is the cheapest of seven rate cards, ahead of CloudRift at $0.60, against a median of $0.72 and a dearest of $1.40. The RTX PRO 6000 Blackwell Max-Q has one rate card at all, EmpirioLabs at $1.80; the card’s other readings are Vast.ai marketplace floors, which our tables keep apart from rate cards.
Two of those three cards carry one or two sellers in the table; the H100 SXM carries twenty-five.
The two near misses
Daytona came second on the B300 at $6.25 for a GPU-only rate, behind Sesterce’s $4.95 from rate, against a median of $7.99 across fifteen providers. Hugging Face came second on the T4 at $0.50 for a single card on AWS, behind Google Cloud’s $0.35 in us-east5, against a median of $0.56 across eight providers.
How to check any number here
Every figure above is listed under this piece with the page it was read from and the time of the read; all of them come from one run, at 06:43 UTC on 2026-09-26. The medians, the counts and the spread are arithmetic over those same readings, computed the way /methodology states, and rate cards leave out marketplace floors and community capacity here exactly as they do on every other page of this site. Nothing here is forecast and nothing here is advice: the piece counts what 51 price pages said on one morning, thirteen of them for the first time.