On 2026-09-24 Nebius lists the NVIDIA HGX H100 at $3.85 per GPU-hour on demand, and preemptible H100 capacity from $0.79 per GPU-hour. Until the 06:39 UTC read of that morning our pages carried $4.50 as its on-demand rate and $3.85 as its spot rate. Both were real numbers on the page; neither stood in the column we thought we were reading.
The rate card carries two on-demand columns, and a preemptible table of its own
The GPU table at nebius.com/prices has two price columns: one headed “On-demand, GPU-hour”, one headed “GPU-hour (Effective October 1, 2026)”. Preemptible capacity is not in that table at all. It has one of its own, “Preemptible NVIDIA GPU Instances”, under a column headed “Spot price, GPU-hour*”, whose cells are “from” prices rather than firm ones.
Our scraper chose its columns by position instead of by heading. When the dated column appeared, the figure we published as the on-demand rate became the one the page dates to October, and the figure we published as the spot rate became the on-demand rate. The preemptible table we were not reading at all.
What the page prints on 2026-09-24, read at 06:39 UTC: the H100 SXM at $3.85 on demand and from $0.79 preemptible, the H200 at $4.50 and from $0.79, the B200 at $7.15 and the B300 at $7.85, both from $0.99, the RTX PRO 6000 at $1.80 and from $0.79, and the L40S from $1.55 and from $0.74. The dated column is an announcement rather than a rate anyone can pay on 2026-09-24, so we publish none of it.
Ten quotes corrected, and the two figures they moved
That read changed ten Nebius quotes at once, and our changelog carries them with both sides: the H100 SXM on-demand rate from $4.50 to $3.85, a change of -14.4%, and its spot rate from $3.85 to $0.79, a change of -79.5%; five more cards moved with them. Read as price moves they look like a deep cut by the provider on one morning. They are not: they are the distance between the column we read before and the column the heading names.
Two computed figures moved too. The Cloud GPU Price Index for the provisional week of September 21, 2026 reads $3.92 per GPU-hour on 2026-09-24; we read the same provisional week at $3.97 on 2026-09-23. And the spot discount band now runs from 81.5% below a provider’s own on-demand rate at its deepest to 20% at its shallowest, with a typical 50%, over 37 rate pairs at 6 providers. Read on 2026-09-23, the shallowest was 0%, over 42 pairs.
That shallow end was the subject of our piece of September 23 on spot discounts by provider. Its Nebius figures came out of the columns described above and do not stand: the pairs it read as 0% were an on-demand rate set against itself. A dated piece here is a statement about its day and is never edited afterwards, so that one keeps the readings it carried, and this one is the correction.
Where the card stands on 2026-09-24
On our H100 SXM page the cheapest rate of any kind is now that Nebius floor, from $0.79 per GPU-hour, tagged high risk because preemptible capacity can be reclaimed mid-run. The cheapest on-demand rate is $1.79 per GPU-hour at Cudo Compute, 54% below the median rate card. Nebius’s $3.85 is a firm cell and its $0.79 is a floor, and the discount table never sets a floor against a firm rate, so the one Nebius pair it keeps is the L40S: from $0.74 against from $1.55, 52.3% apart.
How to check any number here
Every figure above is listed under this piece with the page it was read from and the time it was read. The six cards sit in the two tables named here on the Nebius rate card; our figures for them are on our Nebius provider page with a read time on every row, and the ten corrections are on our changelog. Nothing here is modelled, forecast or estimated: these are listed rates read on the hour, and a “from” price is a floor rather than a rate any one machine is offered at.