Nebius switches preemptible GPUs to dynamic spot pricing

Nebius stopped charging fixed prices for its preemptible GPU virtual machines on October 8, 2026. Since then, the price of a preemptible B300, B200, H200, H100 or RTX PRO 6000 moves with spare capacity and demand for that GPU type in that region. Nebius announced the change in a blog post on September 22, 2026.

How the price moves

• The spot price can change as often as every 15 minutes.
• It stays between a floor and a ceiling. The ceiling is one cent below the current on-demand rate for the same GPU.
• The console shows the current price and up to 30 days of price history before you launch.
• The same spot price applies to preemptible capacity in Kubernetes clusters and Serverless AI.

Old fixed prices and the new range

Per GPU-hour, in US dollars, excluding tax. The spot figures are the minimums that Nebius’s documentation lists; the ceiling follows from the one-cent rule and today’s on-demand prices.

GPU	Fixed preemptible, before Oct 8	Spot, from Oct 8	Spot ceiling
B300	$4.30	from $0.99	$9.49
B200	$3.95	from $0.99	$8.49
H200	$2.45	from $0.79	$5.39
H100	$2.15	from $0.79	$4.49
RTX PRO 6000	$0.95	from $0.79	$1.79

The floor is far below the old fixed prices, but the ceiling is close to the on-demand rate. What you pay depends on demand at the time. The L40S is not part of spot pricing and keeps a flat preemptible rate.

Two ways to run

• Set a maximum price per GPU-hour for a GPU type and region. Your VM runs while the spot price is at or below it, and you pay the spot price. If the price goes above your maximum, the VM stops and its data is kept. A standalone VM has to be restarted by you or your automation once the price falls; Kubernetes restarts nodes on its own.
• Set no maximum. You pay whatever the spot price is, and you lose the VM only during maintenance or when Nebius reclaims the capacity.

What to do

Nebius asked customers already using preemptible VMs to choose a price setting before October 8. If you run interruptible training or batch jobs there, set a maximum price. Without one, a busy hour can cost almost as much as an on-demand VM.