Summary

  • From 8 October, Nebius prices pre-emptible GPU capacity dynamically by platform and region, using capacity and demand signals; the rate can change at 15-minute intervals.
  • A maximum price is a stop condition, not a capacity reservation. Following the rate avoids a price-triggered stop but accepts a changing bill and possible capacity reclamation.

Analysis

The important change is not the word “spot” but the control it gives each side. Nebius says its algorithm reads available capacity and real-time demand signals from its own users, then sets a rate for a GPU type and region. The provider’s ceiling remains one cent below the current on-demand GPU-hour price. Its launch note says the feature became effective on 8 October; the operating guide says the rate can change while a resource is running.

The public pricing page currently displays “from” rates of $0.99 per GPU-hour for HGX B300 and B200, and $0.79 for H200, H100 and RTX PRO 6000. These are starting figures, not a rate promised for every region or hour. The same page says L40S pre-emptible instances remain outside the dynamic scheme and use a flat rate. Nebius says customers can inspect up to 30 days of pricing history in its console, but the public pages do not provide a customer-independent price series.

Buyers choose between following the live rate and setting a maximum. With a policy, they pay the spot price rather than their ceiling; if the rate moves above that cap, the VM stops. The cap therefore limits the per-GPU-hour rate, not total project spend, and it does not reserve capacity. The console’s hourly estimate assumes the rate stays unchanged for that hour, while Nebius documents billing at successive 15-minute prices. GPU count and time still determine the bill.

Stopping is an operational event. Nebius says attached-volume data is retained, but dynamic public IP addresses are released and local SSD data is ephemeral; its documentation also describes a 60-second warning before forced shutdown. A standalone VM needs a manual or automated restart. Nebius says Kubernetes restarts pre-emptible VMs in a cluster, but that does not remove checkpoint, rescheduling or completion-time costs. The company reports that only a “small fraction” of PVM workloads have been interrupted by capacity reclamation, without giving a denominator or period. That is a company claim, not a measured rate readers can reproduce.

Sources