Nebius opens its GPU spot-price feed to everyone

Nebius announced public live prices for its GPU spot capacity on 21 October, making the quotes available on its website and through services including Shadeform, Silicon Data and GPU Finder. It is also piloting a programme under which customers can release idle GPUs back into the spot pool and recover some value.[1]

The pricing mechanism itself is already running. Nebius announced dynamic spot prices on 22 September and put them into effect on 8 October. The new development opens the market information to people outside the cloud’s existing customer accounts.[4]

What the public feed tells you

Nebius’s documentation says both the website and the price API work without an account. Quotes identify a GPU platform, region, currency, price per GPU-hour and the time that price took effect. A separate timestamp identifies when the data was captured. History is available for up to 30 days.[2]

That creates a useful starting point for a team deciding where to run a flexible job. It can compare regions, examine recent variation and build a budget before provisioning a machine. The history also helps distinguish a briefly attractive quote from a price that has persisted. It remains a record of past prices, rather than a forecast.[2]

A maximum price sets a stopping condition

Prices can change as often as every 15 minutes, including during a run. A customer can set a maximum price for a GPU platform: while the market price is below that threshold, the customer pays the market price. If it rises above the threshold, the resource stops. Capacity can also be reclaimed independently of the price.[3]

For an illustrative calculation, eight GPUs running for one hour at $2 per GPU-hour incur $16 in that compute charge. With a $3 maximum, the same hour still costs $16; the threshold controls when a rising price stops the machine. These figures explain the mechanism and are not observed Nebius quotes.[3]

Nebius bills running compute by the second, although it displays hourly prices. Its published prices exclude applicable taxes. Storage volumes have separate charges and continue to incur them while a VM is stopped, so a compute quote covers only part of a project’s cost.[5]

There is an operational consequence to the spending control. Nebius’s launch documentation says standalone preemptible VMs do not restart themselves after a price stop; customers must restart them or automate that step. Kubernetes can restart preemptible machines in its clusters automatically.[4]

For a job that can save its progress and resume, public pricing makes the trade-off easier to calculate. The relevant saving is the cost of finishing the job, including any repeated work after interruptions. A lower hourly quote becomes valuable when the application can actually make use of it.[3]

The idle-capacity rebate could add another side to that market by returning unused capacity to circulation. For now, Nebius describes it as a pilot. The immediately usable change is the ability to inspect the prices and their history before deciding how much interruption a workload can tolerate.[1]

Report an error or suggest a correction