Nebius AI Cloud review

Nebius AI Cloud is infrastructure for training and running AI models, rather than a model-building app or a fixed monthly software subscription. Its homepage describes GPU-based compute, built-in MLOps tools, and managed and serverless inference. It also pitches self-service clusters that can be provisioned in minutes, with capacity for small experiments through global-scale workloads.

The hardware is a key part of the offer: Nebius describes custom infrastructure with non-virtualized GPUs and InfiniBand. For interruptible jobs, it offers live GPU spot pricing, which lets customers compare available models and regions, set a maximum price, or follow the market.

Pricing is usage-based. The pricing page lists hourly rates for GPU and CPU compute, plus monthly-per-GiB rates for storage. That means the relevant comparison is not just the GPU-hour price: storage and other usage charges also matter when estimating the cost of moving a workload.

Nebius combines GPU infrastructure with tools and services around the model workflow. That can be useful for teams comparing a move from a general-purpose cloud: the offer includes MLOps tooling, managed inference, and self-service clusters, rather than compute alone. The homepage says clusters are repeatable and available in minutes, though it does not publish a provisioning-time guarantee.

The listed hardware spans several workload types, from NVIDIA H100 and H200 systems to B200 and B300 options, plus RTX PRO 6000 and L40S configurations. The pricing page also lists CPU-only instances, so teams do not have to treat every part of a cloud workload as GPU compute.

• Published on-demand rates include $4.50 per GPU-hour for HGX H100, $5.40 for HGX H200, and $1.80 for RTX PRO 6000.
• Preemptible rates start at $0.79 per GPU-hour for H100 and H200, and $0.99 for B200 and B300. These rates are not listed for every model.
• Storage options include shared filesystem, WEKA, object storage tiers, and block volumes with different replication options.

The live spot-price feature adds a way to compare availability and set a price ceiling for interruptible work. We would still compare the full compute and storage bill before switching.

The pricing is not a simple monthly plan. The pricing page lists hourly compute and monthly-per-GiB storage charges, but no monthly or annual subscription price. Several high-end GPU rates are also undisclosed: Nebius says to contact it for on-demand GB200 NVL72 and GB300 NVL72 pricing, and lists no preemptible price for either. A listed starting rate can therefore be a poor basis for forecasting a full deployment.

Storage adds another calculation. Listed examples range from $0.0147/GiB per month for Standard object storage to $0.1100/GiB per month for Enhanced object storage; egress is charged for some options and free for others. The pricing page also lists some services, including managed Kubernetes and networking egress/ingress, as free, but that does not make every storage or compute charge disappear.

The published pricing information does not publish a free plan, trial terms, usage caps, or a single all-in estimate. Teams evaluating a switch should confirm the price and availability for their chosen region and GPU model. Preemptible capacity is interruptible by definition, and some L40S preemptible rates are flat rates rather than spot pricing, according to the pricing details.

Nebius lists usage-based compute rates by GPU-hour, not a monthly or annual plan. On-demand pricing starts at $1.55/GPU-hour for L40S with Intel CPUs, while HGX H100 is $4.50, H200 is $5.40, B200 is $8.50 and HGX B300 is $9.50. That’s a substantial step up for newer, higher-end GPU systems; GB200 and GB300 NVL72 on-demand prices require contacting Nebius. CPU-only instances start at $0.06 per CPU-hour for Intel Ice Lake and $0.13 for AMD EPYC Genoa.

Preemptible rates start at $0.79/GPU-hour for H100 and H200, and $0.99 for B200 and B300. Availability and pricing vary by model; L40S preemptible pricing is a flat rate, not spot pricing. Storage adds to the bill: Standard object storage is $0.0147/GiB/month, while shared filesystem storage is $0.0800/GiB/month. Egress can also cost extra: Standard object storage lists $0.0150/GiB. Managed Kubernetes, networking egress/ingress and public IP addresses are listed as free. Budget for usage and storage, not a fixed subscription.

Nebius says self-service clusters can be available in minutes and includes built-in MLOps tooling, but the published material does not specify migration services, data-export formats or compatibility with another cloud’s pipelines. Treat those as questions to verify before moving workloads. Map your model artifacts, training data, containers, checkpoints and inference endpoints, then confirm how each can be transferred and what storage or egress charges apply.

The likely switching effort depends on how much of your workflow uses Nebius-specific tooling; we don’t have sourced details on portability or lock-in to quantify it. Its GPU-hour pricing and preemptible options make it worth comparing equivalent instance types and interruption tolerance before moving. We also don’t have Reddit posts or user reports in the published sources, so we can’t responsibly claim what tends to make customers leave. Ask Nebius for export and deletion terms, and test a small workload before committing production data or training runs.