GPU EconomyCloud GPU prices, read every hour
updated
H100 index$3.5426 providersWidestV10015.9×Mover, 7dB300−1.4%Coverage709 quotes63 providersReads, 24 h17,26926 runs

Analysis · · figures frozen on that day

H100 SXM median closed at $3.70, week of 21 September 2026

Published . Every figure below is a rate this site read on a provider's own page, listed with its source under the text and frozen as of that day - prices move, and a dated piece says what was true on its date. The live tables are linked beside each figure.

The week of 21 September 2026 closed with our H100 SXM median at $3.70 per GPU-hour, taken from 26 publicly listed on-demand rates across 24 clouds. The week of 14 September closed at $3.90 across 21 quotes and 19 clouds; the week of 7 September at $3.92 across 20 and 18. Week on week that is a fall of 5.1%. Of the quotes read in both of the last two weeks, one listed rate changed.

Nineteen of twenty rates read the same

Twenty (provider, region) pairs were quoted in both weeks. Nineteen of them were read at the same figure in both. The one that changed is Verda, the cloud our table still files under the id DataCrunch: $3.3475 per GPU-hour in the week of 14 September and $3.555 in the week of 21 September, a rise of 6.2%. Its price page printed 3.555 when this piece was written on 2026-09-28; our tables round it to $3.56, and the week before they rounded it to $3.35.

That gap between a falling median and a nearly unchanged set of rates is what the second column of the Price Index exists to measure. The base-100 index is chain-linked: each week’s link is the geometric mean of the week-over-week ratios of the pairs quoted in both weeks, so a cloud joining or leaving contributes no ratio at all. Nineteen ratios of exactly 1 and one of 1.062 give a link of 1.003, and the index moved from 100.1 to 100.4 against the week of 7 September as 100 — a rise of 0.3% in the same week the dollar median fell 5.1%.

The panel changed, the listed rates did not

Six pairs entered the basket for the first time: Atlas Cloud at $2.95, Civo and HexGrid’s Secure Cloud at $2.99, Hyperbolic at $3.19, Canopy Wave at $3.25 and Neysa at $4.39. Hyperbolic and Neysa print theirs as from rates, and Canopy Wave prices per GPU on a two-GPU machine. Five of the six sit below the previous week’s median of $3.90.

One pair left. GMI Cloud’s $2.00 was the cheapest quote in the week of 14 September; the figure itself did not change, but from the read of 19 September, an hour after our 16:38 UTC read, we file it under H100 80GB (form factor not stated) rather than H100 SXM, so it is no longer in this card’s basket.

The two ends of the basket are the same rows as the week before: Cudo Compute at $1.79 in no-kristiansand-1 and Microsoft Azure at $11.06 in eastus. What changed is the middle. The middle half of the week’s quotes was $2.89 to $5.95 in the week of 14 September and $2.99 to $4.41 in the week of 21 September.

The two halves of the panel

Cut by operator class, the four hyperscaler quotes carry a median of $9.90 per GPU-hour — the same figure, from the same four clouds, in both weeks. The specialists went from $3.49 across 17 quotes and 15 clouds to $3.37 across 22 quotes and 20 clouds. All six of the arrivals are specialists.

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. The week has a receipt of its own, 2026-W39.csv: one row per quote read for that week, with its source link, its read time and whether it entered the median. It carries 31 rows — the 26 that entered, and 5 read and kept out under a named rule: Gcore, Scaleway and Seeweb because they list in euros, and two Vast.ai rows because a marketplace floor is not a rate card. With 26 quotes the median is the midpoint of the thirteenth and fourteenth of them in order, which are Verda’s and Nebius’s $3.85. The week is marked final, which means it is not recomputed when later reads arrive. Nothing here is forecast and nothing here is advice: the piece states what 26 price pages were read at over one week, and which of them moved.

Every figure in this piece, and where it was read

As published on September 28, 2026. Each price links to the page it was read from; the last column links the page that keeps it current.
FigureAs publishedRead from, whenLive page
Weekly median, H100 SXM on-demand, week of 2026-09-21 (final)$3.70 per GPU-hour from 26 quotes across 24 providers, against $3.90 the week before (-5.1%)source
/price-index
Middle half of the week's quotes (p25-p75), week of 2026-09-21$2.99 to $4.41 per GPU-hoursource
/price-index
Weekly median, H100 SXM on-demand, week of 2026-09-14 (final)$3.90 per GPU-hour from 21 quotes across 19 providerssource
/price-index
Middle half of the week's quotes (p25-p75), week of 2026-09-14$2.89 to $5.95 per GPU-hoursource
/price-index
Weekly median, H100 SXM on-demand, week of 2026-09-07 (final)$3.92 per GPU-hour from 20 quotes across 18 providerssource
/price-index
Base-100 index (chain-linked, week of 2026-09-07 = 100)100.4 in the week of 2026-09-21, against 100.1 the week before (+0.3%)source
/price-index
Quotes read in both the week of 2026-09-14 and the week of 2026-09-2120 of the 26 (provider, region) pairs; 19 of the 20 read at the same figure in bothsource
/price-index
Verda (listed by us as DataCrunch), H100 SXM on-demand$3.555 per GPU-hour in the week of 2026-09-21, against $3.3475 in the week of 2026-09-14 (+6.2%); our tables round them to $3.56 and $3.35source
/providers/datacrunch
Atlas Cloud, H100 SXM on-demand, first week in the basket$2.95 per GPU-hoursource
/providers/atlascloud
Civo, H100 SXM on-demand, first week in the basket$2.99 per GPU-hoursource
/providers/civo
HexGrid (Secure Cloud), H100 SXM on-demand, first week in the basket$2.99 per GPU-hoursource
/providers/hexgrid
Hyperbolic, H100 SXM on-demand, first week in the basket$3.19 per GPU-hour, an advertised from ratesource
/providers/hyperbolic
Canopy Wave, H100 SXM on-demand, first week in the basket$3.25 per GPU-hour, priced per GPU on a two-GPU machinesource
/providers/canopywave
Neysa, H100 SXM on-demand, first week in the basket$4.39 per GPU-hour, an advertised from ratesource
/providers/neysa
GMI Cloud, the quote that left the basket$2.00 per GPU-hour, the cheapest in the week of 2026-09-14; read under H100 SXM up to 2026-09-19 16:38 UTC and under H100 80GB (form factor not stated) from the read an hour latersource
/providers/gmi-cloud
Cheapest quote in the basket, Cudo Compute (no-kristiansand-1, GPU only)$1.79 per GPU-hour, the cheapest in both weekssource
/providers/cudo
Dearest quote in the basket, Microsoft Azure (eastus)$11.06 per GPU-hour, Standard_ND96is_H100_v5, the dearest in both weekssource
/providers/azure
Nebius, H100 SXM on-demand$3.85 per GPU-hour, NVIDIA HGX H100source
/providers/nebius
Specialist clouds, weekly median, week of 2026-09-21$3.37 per GPU-hour from 22 quotes across 20 providers, against $3.49 from 17 quotes across 15 providers the week beforesource
/price-index
Hyperscalers, weekly median, week of 2026-09-21$9.90 per GPU-hour from 4 quotes across 4 providers, the same figure and the same count as the week beforesource
/price-index
Quotes read for the week and kept out of it5 of the 31 rows in the week's receipt: Gcore, Scaleway and Seeweb as EUR listings, two Vast.ai marketplace floorssource
/price-index

Answers the reader question: J6 - tell whether H100 rates are falling before I commit. Written because The index week of 2026-09-21 closed on 2026-09-27 at $3.70 with 26 quotes, against $3.90 with 21 the week before. All of it is open data under CC BY 4.0; the numbers here are a snapshot, theindex and the card pages are current.

Data updated: