Nebius moves spare GPU hours to spot prices from $0.79
Nebius switched its spare GPU hours from a fixed price to live spot pricing on 8 October. An H100 hour now floats between $0.79 and $4.49. Nebius gets paid more when clusters are full. The stuck step is the idle hour, and only the console shows it.
WIRE · NEOCLOUDS · GLOBAL · Data as at 9 Oct 2026
Nebius no longer sells its spare GPU hours at a fixed price. Since 8 October 2026, its interruptible (preemptible) GPU hours float. An H100 hour now costs anywhere from $0.79 to $4.49. Until last week it was a flat $2.15. The price moves with how full Nebius's clusters are, and it can change every 15 minutes. On the neocloud model, this is the idle-hour step: Nebius now gets paid more for spare chips when they are scarce, and less when they sit.
What happened
Nebius, the GPU cloud that rents Nvidia chips by the hour, updated its compute price list for October. Two changes, a week apart:
8 October
Preemptible B300, B200, H200, H100 and RTX PRO 6000 hours move to dynamic spot pricing
Nebius Compute pricing documentation
The band
Floor of $0.79 (H100, H200) or $0.99 (B200, B300). Ceiling one cent below the on-demand rate
Price set by capacity, cluster use, customer demand and pricing preferences
1 October
On-demand rates up 17% to 21%: H100 $3.85 to $4.50, B300 $7.85 to $9.50
This raised the spot ceiling too, since the ceiling tracks on-demand
Who can see the price
Current and past spot prices sit inside the Nebius console, not on a public page
Nebius price page, read 9 Oct 2026
Why cash cares
A GPU earns money only in the hours someone rents it. An hour nobody buys is gone. Preemptible hours are how a cloud sells chips that would otherwise sit idle. The renter gets a cheap price, and Nebius can take the chip back when a full-price customer wants it.
Under the old fixed price, Nebius got the same $2.15 for a spare H100 hour on a quiet night or a packed afternoon. Now it gets what the hour is worth at that moment.
Our arithmetic, not a Nebius figure: one eight-GPU H100 server run for a 730-hour month on spare time cost about $12,556 at the old fixed price. At the new floor it is about $4,614. At the new ceiling it is about $26,222. Same chips, a spread of almost six times, set by how busy the cluster is.
Figure: Spare-hour bind. The dark tick is the fixed preemptible price until 8 October. The amber band is the new spot range. For every chip, the ceiling is about twice the old fixed price, and the floor is well under half of it.
The second-order read
Who gets paid. Nebius, when its clusters are full. A spare B300 hour that earned a flat $4.30 can now earn up to $9.49. The spot price also points spare chips at whoever values them most.
Who pays. Renters running batch jobs, tests and research on spare time. They used to know their price. Now they carry the price risk, on top of the risk of being kicked off.
Who waits. Buyers who want certainty. Nebius still sells multi-month cluster reservations at up to 35% off on-demand. The new spread makes that commitment look better to a renter.
Stuck step. The idle hour. The spot price now moves with how busy the clusters are, so it is a live gauge of GPU utilization. But Nebius shows it only inside its console. Outside investors can't yet see whether spare hours clear near $0.79 or near $4.49.
Two cautions
The band is the published range, not what renters actually pay. Nebius has not said where spot prices have traded since 8 October.
Spare hours are a small slice of a GPU cloud's revenue. Most of the money comes from long contracts. This change tells you how Nebius prices its leftovers, not how much it earns.
What would change the view
- Nebius publishes spot history, or reports spot revenue, in its next results.
- Other GPU clouds move preemptible hours to floating prices too.
- Spot prices sit near the floor. That would mean spare capacity is piling up after the October rate rise.
- Spot prices sit near the ceiling. That would mean the clusters are close to full.
What to watch
- Nebius third-quarter results: any word on utilization, spot or preemptible revenue, and reservation demand.
- Rival price lists: whether others match the 1 October on-demand rise or the 8 October spot change.
- Reservation discounts: if the "up to 35% off" shrinks, demand for committed capacity is strong.
Read: Neoclouds, What Is GPU Utilization?, and What Does Lambda Do?
Get the Wire. A free brief every day, here and in your inbox. Subscribe free →
Sources: Nebius, Compute pricing in Nebius AI Cloud, read 9 Oct 2026: from 1 Oct 2026 on-demand H100 $4.50 (was $3.85), H200 $5.40 (was $4.50), B200 $8.50 (was $7.15), B300 $9.50 (was $7.85); from 8 Oct 2026 preemptible pricing for B300, B200, H200, H100 and RTX PRO 6000 switches to dynamic spot pricing, preemptible from $0.79 (H100, H200, RTX PRO 6000) and from $0.99 (B200, B300), previously fixed at $2.15 (H100), $2.45 (H200), $3.95 (B200), $4.30 (B300) and $0.95 (RTX PRO 6000); spot prices can change as often as every 15 minutes and the highest is one cent below the on-demand rate; Nebius price page, read 9 Oct 2026: commitment discounts of up to 35% off on-demand for multi-month cluster reservations, current spot prices shown in the Nebius console; GPU Cloud Advisors, 28 Sep 2026: 1 Oct increase of 17% to 21%. The monthly server cost examples, the ceilings as a multiple of the old price, and the spread are our arithmetic. Not investment advice. Do your own research.
