Nebius

What it is for

Nebius is an AI-centric GPU cloud: you launch an on-demand or spot VM on a published NVIDIA SKU and pay per GPU-second. Use it when the job is a listed H100 / H200 / B200 price, not when you want the cheapest marketplace bid.

It is not open source. On-demand H100 NVLink is $4.50 / GPU-hour on the public table (from 1 Oct 2026). Preemptible H100 starts from $0.79.

What makes it unique: the cover is Nebius’s official GPU price table — HGX H100 $4.50, H200 $5.40, B200 $8.50, B300 $9.50, plus spot “from” columns.

Key features

• GPU virtual machines — create and run VMs that attach published NVIDIA GPUs for training or inference (Compute; Prices).
• Token Factory inference — serve open models through an OpenAI-shaped API (serverless or dedicated endpoints, plus post-training) (Token Factory).
• Managed Kubernetes — a hosted Kubernetes control plane with GPU node pools for containerized ML jobs (Managed Kubernetes).
• Soperator / Slurm training — managed Slurm-on-Kubernetes clusters for multi-node training jobs (Soperator).
• Object Storage — S3-compatible buckets for datasets, checkpoints, and model artifacts (Object Storage; Storage).
• Shared filesystems — POSIX volumes several VMs can mount at once for streaming data or writing checkpoints (Compute disks).

Who it's for

Best fit: training and inference teams who want a published European / US SKU and will read the Oct 2026 price change.

Poor fit: people hunting the lowest interruptible bid on a marketplace (Vast.ai).

Pricing

Verified 2026-10-22 (Asia/Shanghai) from nebius.com/prices (cover) and Compute pricing. USD except Israel (ILS). Taxes extra. On-demand H100 moved to $4.50 on 1 Oct 2026; spot became dynamic on 8 Oct 2026. Re-check the row.

On-demand / spot examples (USD / GPU-hour, cover)

H100 NVLink is documented as eu-north1 only. L40S preemptible is a flat rate, not the dynamic spot graph. There is no public $0 VM.

Limitations / costs

• $4.50 H100 is the post-Oct-2026 list. Older $3.85 docs are stale.
• Spot can move every 15 minutes. Checkpoint.
• Stopped VMs still pay disks.
• Not a Pod marketplace. For bids, see Vast.ai.

Alternatives

• Lambda — 8× H100 SXM $3.99/GPU/hr if you want one US operator.
• Runpod — per-second Pods; H100 PCIe $2.89/hr on the Oct 8, 2026 card.
• Vast.ai — live H100 SXM from $1.73/hr on the public feed we last copied.

Official tutorial

We searched official Nebius YouTube / docs video for a current getting-started console walkthrough we could oEmbed. None resolved as a first-party product-UI tutorial, so nothing is embedded. Use the written Compute pricing plus Prices.

Website

https://nebius.com

Source

Catalog row curated for AI Tools Nav. Cover is Nebius’s official GPU table (H100 $4.50/hr on-demand). No official getting-started embed found.

Last verified

2026-10-22 (Asia/Shanghai) editorial check.