Nebius

Nebius AI Cloud rents full VMs on NVIDIA GPU platforms: B300, B200, H200, H100, RTX PRO 6000 and L40S, in regions across Europe, the US, Israel and the UK. Each platform has fixed sizes, called presets, from 1 to 8 GPUs. gpuprov boots Nebius's Ubuntu 24.04 image for NVIDIA GPUs, which has the driver, CUDA and the NVIDIA Container Toolkit preinstalled. VMs can be on-demand or preemptible (--spot), and preemptible VMs pay a live spot price.

nebius profile create                     # once; or export NEBIUS_IAM_TOKEN=...
gpuprov provision --cloud nebius --name dev --gpus H100:1
gpuprov provision --cloud nebius --name big --gpus B300:8 --region us-north1 --spot

Acceptance testing pending

This adapter is covered by unit, mock-API and vendor-stub tests. It has not yet passed the end-to-end acceptance run against a live account (allocate → SSH → k3s/GPU ready → terminate → nothing billable left), so treat it as preview.

Credentials, tenant and projects

gpuprov uses the Nebius REST API (https://api.nebius.cloud) with an IAM access token. It looks for one in this order:

1. [nebius] token in the config file
2. $NEBIUS_IAM_TOKEN
3. nebius iam get-access-token, run with [nebius] profile if you set one. The token is cached for an hour (tokens are valid for 12).

The CLI route covers user accounts and service-account profiles alike. A service account needs a group with the editor role in the tenant or project.

Nebius projects are regional: a tenant has one project per region. gpuprov lists the tenant's active projects and maps them by region, so a region (eu-north1, us-central1, …) is all you name. If a region has several projects, the one with the lowest id is used. The tenant comes from [nebius] tenant_id, then $NEBIUS_TENANT_ID; failing both, it is the only tenant your credentials can see. To use specific projects and skip discovery, pin them:

[nebius]
projects = { eu-north1 = "project-e00…", us-central1 = "project-e00…" }

Planning: live platforms, prices and capacity

For each candidate region, gpuprov lists the platforms and presets available in that region's project. Each preset that matches the ask becomes one offer:

• instance_type = platform/preset, e.g. gpu-h100-sxm/1gpu-16vcpu-200gb
• region = the project's region; there are no zones (--zone pins a region)
• ordered by hourly price, cheapest first

Without --region, every region where you have a project is a candidate.

Prices. On-demand offers are priced in one batch by Nebius's billing calculator. Preemptible offers use the public spot-price list, which quotes one GPU hour per platform and region, multiplied by the preset's GPU count. An offer that can't be priced is kept and tried last.

Capacity advisor. When the tenant is known, gpuprov reads Nebius's capacity advisor for your quotas and current stock. It drops an offer whose fresh advice says nothing can launch, and among offers at the same price it tries the more available one first. The advisor is only a hint: if it can't be read (for example, without permission), planning goes on without it.

--gpus matches the platform by whole tokens of its id or name: H100 matches gpu-h100-sxm but not gpu-h200-sxm, L40 doesn't match L40S, and RTX6000 and RTXPRO6000 both match gpu-rtx6000. The count must equal the preset's GPU count (H200:1 or H200:8). --instance-type takes platform/preset, or a bare platform for all of its presets. --spot keeps only platforms that allow preemptible VMs.

When a launch fails with Nebius's "not enough resources" error, gpuprov fails over to the next offer. A quota error moves on to the next region.

What gets created

Resource	Notes
Compute instance	named gpuprov-<name>, labeled gpuprov-vm-name=<name>; one network interface on the project's default subnet with a dynamic public IP
boot disk	VM-managed NETWORK_SSD disk of --disk GiB, named gpuprov-<name>-boot; created and deleted together with the instance

Nothing else is created. The subnet belongs to the project. Every VM is in its network's default security group, which Nebius defines as allowing all traffic and which can't be changed. --port therefore needs no firewall, and the VM is reachable on every port from the internet. Secure the services you run on it.

--image takes an image family (e.g. ubuntu24.04-cuda12) or an image id (computeimage-…). Without it, gpuprov follows Nebius's recommendation: ubuntu24.04-cuda13.0 on GPU platforms and ubuntu24.04-driverless on CPU-only ones.

SSH

The login is the spec's user (default ubuntu) on port 22, created with your key by cloud-init.

Kubernetes

Instances accept cloud-init user data, so the k3s installer runs while the VM boots and gpuprov verifies it over SSH. The GPU image already has the driver, so the installer adds the NVIDIA runtime and device plugin. Clusters use direct API access on port 6443.

Spot (preemptible) VMs

--spot creates a preemptible VM that follows the current spot price, with recovery policy FAIL, as Nebius requires. On preemption Nebius stops the VM rather than deleting it. A stopped VM's compute isn't billed, but its boot disk is, and its quota stays held. gpuprov health reports a preempted VM as UNHEALTHY. Run gpuprov down to delete it.

Health

Besides the instance state, gpuprov health reports:

• preemption: a stopped preemptible VM is critical
• a stopped on-demand VM (still holding quota and disk) is critical
• a maintenance event that affects the VM is a warning

Teardown

gpuprov down deletes the instance, which also deletes its VM-managed boot disk, then waits until Nebius no longer lists it. There is nothing else to clean up.