Nebius AI Cloud
NVIDIA GPU cloud with a European sovereign inference API
Tool details
Nebius AI Cloud is the pay-as-you-go AI cloud of Nebius, a Dutch publicly listed company. All of its compute comes from NVIDIA GPUs, and its data centers are concentrated in Finland, France and the UK. It brings bare-metal GPUs, managed Kubernetes and the Token Factory inference API into a single account.
Core capabilities
- On-demand GPU instances: H100 at $4.50 per hour, H200 at $5.40, B200 at $8.50, B300 at $9.50, and preemptible H100 from $0.79
- Token Factory: more than 60 open-source models billed per token, from $0.06 per 1M tokens, OpenAI SDK compatible
- Slurm as a first-class citizen: Managed Soperator runs Slurm on Kubernetes, with a $0 control plane
- In-house racks: servers and racks designed in-house, smart air cooling, roughly 20% less energy under the same conditions
- Managed services: Kubernetes, PostgreSQL and S3 object storage, with egress, ingress and public IPs all at $0
- Cluster scale: InfiniBand at 3.2 Tbit/s, with a single cluster scaling to thousands of GPUs
Where it fits AI teams that need EU data residency, projects moving HPC workloads to multi-node training, product teams looking to cut open-source model inference costs, and companies like Revolut and Shopify that already run at scale. It is not a good fit for general-purpose web hosting or for systems that need disaster recovery across availability zones.
Unique strengths One of the few neoclouds that open both a bare-GPU path and a per-token API through a single account, with the contracting entity in the Netherlands and every data center in Europe. Its in-house ODM racks deliver a lower TCO, and the company claims 83% savings on fine-tuning and 82% on inference versus AWS.
Editorial assessment Gartner 4.8/5 and SemiAnalysis ClusterMAX Platinum, with far stronger recognition on the professional side than consumer platforms. There is no free trial: registration requires a $25 prepayment, and the Builder Program offers $25 in inference credits. Note that revenue is highly concentrated in Microsoft and Meta, that compute depends 100% on NVIDIA, and that there is no cross-availability-zone disaster recovery.
Pricing
Subscription$25 minimum prepaid balance (L40S GPU from $1.55 per GPU-hour)Adding a card prepays your balance, which is then deducted by actual usage; no subscription plans and no free tier
Entry-level inference and graphics compute; RTX PRO 6000 is $1.80 per GPU-hour
The mainstay training and inference cards; preemptible from $0.79 per GPU-hour
Blackwell generation; preemptible from $0.99 per GPU-hour
Full-rack configurations; capacity and delivery timelines must be confirmed with sales
Inference API for open-source models, with input and output priced separately
Commit to annual usage in exchange for a discount; third-party sources cite up to about 35%
Additional notes
No free trial: registration requires a card and a $25 prepayment; you can apply to the Builder Program for $25 in Token Factory credits (learning and testing only). Managed Kubernetes, Soperator, egress and ingress traffic and public IPs are all $0, and managed PostgreSQL is $0.034 per vCPU-hour. Preemptible and live spot prices vary by region.
FAQ
User reviews
User Reviews
No reviews yet
No one has reviewed this tool yet
