Nebius AI Cloud review

Nebius AI Cloud is a hosted cloud infrastructure service for AI training and inference, not a model library or software package to install in your application. It provides GPU-based compute, custom hardware with non-virtualized GPUs and InfiniBand, plus self-service clusters and built-in MLOps tooling. Managed and serverless inference are also part of the offering.

In a stack, Nebius sits below model code and training frameworks: teams provision compute and storage for experiments, training runs or inference services. The homepage describes cluster access as self-service, with clusters available in minutes and repeatability built in. Its capacity options are intended to cover small experiments through large-scale environments, with flexible consumption and live GPU spot prices for interruptible workloads.

That architecture makes it relevant when GPU availability and infrastructure are the main constraints. It is not presented as a replacement for your application framework, model-serving code or the rest of your cloud stack.

The described developer workflow centers on self-service cluster provisioning, GPU capacity selection and managed inference. The homepage also points to live GPU spot prices, where buyers can compare models and regions, set a maximum price or follow the market. These are useful infrastructure controls, but they do not establish a particular programming interface.

The homepage does not name an API style, SDK, programming language, CLI, webhook or plugin. Treat cluster provisioning and managed services as the visible integration surface, then confirm API access, automation options, supported SDKs and languages in Nebius's documentation before designing a deployment workflow. Ask whether the controls you need are available through an API or only through the service interface.

Nebius advertises built-in MLOps tooling, but the homepage does not detail its components or integrations. Confirm how that tooling fits your training framework, job scheduler, observability stack and model-serving setup. The customer material mentions Slurm clusters, but does not establish a general integration guarantee.

The pricing page lists usage-based hourly compute and monthly-per-GiB storage rates; it lists no free plan, and free-trial terms are not published. The lowest listed compute rate is $0.06 per CPU-hour for Intel Ice Lake CPU-only instances. GPU preemptible rates start at $0.74 per GPU-hour for L40S with Intel CPUs; listed rates include $0.79 for H100 and H200, and $0.99 for B200 and B300. On-demand rates include $4.50 per H100 GPU-hour, $5.40 for H200 and $8.50 for B200. GB200 and GB300 on-demand pricing requires contacting Nebius.

Preemptible capacity is interruptible; the L40S configurations use flat preemptible rates rather than spot pricing. Storage adds to compute: Standard object storage is $0.0147/GiB/month, with $0.0150/GiB egress. Shared Filesystem is $0.0800/GiB/month, and block volumes range from $0.0530 to $0.1180/GiB/month depending on the option. Object storage tiers have separate rates, while Managed Kubernetes, networking egress/ingress and public IP addresses are listed as free. Model the full bill around GPU hours, storage and data transfer.

Nebius presents AI Cloud as hosted infrastructure. Its homepage does not establish that the platform's software is open source, name a license, or describe a self-hosted deployment option. Buyers who require source access or want to run the control plane in their own environment should ask Nebius directly before adopting it.

The homepage and pricing details also do not describe data export formats or a migration process. Your models and training code may remain portable if you keep them in your own repositories, but infrastructure configuration, managed inference setup and stored data can still create migration work. Confirm how to export datasets and artifacts, whether storage uses standard interfaces, and what happens to data when a service ends.

Treat portability as an architecture requirement: keep source, model artifacts and reproducible job definitions under your control, and validate an exit path for data and inference workloads.

Nebius advertises flexible scaling, self-service clusters, managed inference, 24/7 support and enterprise-grade compliance. Its customer stories describe production use, including Revolut's use of H100 GPUs and Nebius Token Factory. Those claims are not a substitute for confirming service-level commitments, security controls or workload-specific capacity with the vendor.

The pricing material does not publish quotas, rate limits, availability targets or a detailed interruption policy. Preemptible capacity is for interruptible workloads, so use it only where jobs can tolerate restarts or checkpointing. The live spot market also means you should confirm capacity and price behavior for the regions and GPU models you need. Ask about GPU reservations, queueing, regional availability, support escalation and the terms behind any quoted on-demand capacity.

Before production, validate storage and egress costs against expected data movement, and review compliance requirements with Nebius. The vendor advertises enterprise-grade compliance, but teams should confirm the specific certifications and controls their deployment requires.