Nebius removes its spot-price ceiling — preemptible GPUs can now cost more than on-demand

What changed, from Nebius's own post “Introducing spot pricing for preemptible VMs” (published Sep 22, 2026; read Nov 2, 2026): its update note, dated Nov 2, 2026, reads — “this section previously stated that the spot price always stays between a floor and a ceiling one cent below the on-demand rate. We have removed the ceiling. Customers can cap what they pay by setting a max price.” The post's how-it-works section now states plainly: “There is no upper limit, so during periods of high demand the spot price can rise above the on-demand GPU-hour rate.”

How the mechanism works, in the same post: the spot price for a GPU type in a region is set from available capacity and real-time demand, recalculated up to every 15 minutes, and it still cannot fall below a set floor — only the ceiling has gone. A renter who sets a max price keeps the old protection by hand: when spot rises above the cap, the instance stops, data is retained, and it can restart when the price falls back — automatically for Kubernetes clusters, manually (or via Terraform, SDK, API or CLI) for standalone VMs. A renter who instead chose to follow the spot price has no cap at all after today's change.

What did not change: Nebius's on-demand list. Re-read on Nebius's own prices page on Nov 2, 2026, B300 is $9.50 an hour, B200 $8.50, H200 $5.40, H100 $4.50, RTX PRO 6000 $1.80 and L40S from $1.55 — every figure the same as our Nov 1 read. This is not a price rise in the list; it is the removal of the guarantee that the interruptible lane is always the cheaper one.

Why it matters in this market: spot and preemptible capacity is where budget inference, batch jobs and short fine-tunes live, and its whole pitch was a discount in exchange for interruption risk. A spot price that can pass the on-demand rate in a busy hour inverts that pitch exactly when capacity is tightest — which, in the autumn pattern this journal has been tracking (Nebius's own across-the-board rise on Oct 1, Verda's eight rises since Oct 23, SaladCloud's demand-cited rise on Nov 1), is more and more of the time. Rising prices and now an uncapped spot market are consistent with growing demand for GPUs (our reading of the pattern, labelled as interpretation).