Provider record · updated 2026-09-10
Nebius — GPU pricing & deliverability
6 tracked offerings · grade mix A:3 · B:3 · C:0 · D:0 · U:0. Nebius prices per GPU per hour with both on-demand and preemptible tiers published. Cluster docs (fetched) publish the clearest fabric spec of the five providers: each GPU has a 400 Gbps NIC, 8-GPU VMs give 3.2 Tbps per node, GPUDirect RDMA, InfiniBand traffic isolated per cluster; 13 named physical fabrics are listed but max cluster size is not published. Cluster-capable platforms: gpu-h100-sxm, gpu-h200-sxm, gpu-b200-sxm, gpu-b300-sxm. HGX B300 also priced ($7.85 on-demand / $4.30 preemptible). GB200 NVL72 and GB300 NVL72 are listed but 'Contact us' — no public price. Reserved/committed pricing exists only as a generic 'up to 35% less' claim — no rate card published. Fabric generation name (Quantum-2) appears on marketing pages not fetched; recorded fabric details are from docs only.
Tracked offerings
| GPU | SKU | $/GPU-hr | Terms | Grade | Conf | Fabric (as published) | Src |
|---|---|---|---|---|---|---|---|
| H100-SXM | NVIDIA HGX H100 — preemptible eu-north1 (Finland) | $2.15 | spot | B | publ | InfiniBand, 400 Gbps per GPU / 3.2 Tbps per node (per docs.nebius.com) | ↗vendor |
| H200-SXM | NVIDIA HGX H200 — preemptible eu-north1 (Finland), eu-west1 (Paris, France), us-central1 (Kansas City, MO) +1 more | $2.45 | spot | B | publ | InfiniBand, 400 Gbps per GPU / 3.2 Tbps per node (per docs.nebius.com) | ↗vendor |
| H100-SXM | NVIDIA HGX H100 (gpu-h100-sxm platform) eu-north1 (Finland) | $3.85 | on-demand | A | publ | InfiniBand, 400 Gbps per GPU / 3.2 Tbps per 8-GPU node, GPUDirect RDMA (per docs.nebius.com) | ↗vendor |
| B200 | NVIDIA HGX B200 — preemptible us-central1 (Kansas City, MO), me-west1 (Israel) | $3.95 | spot | B | publ | InfiniBand, 400 Gbps per GPU / 3.2 Tbps per node (per docs.nebius.com) | ↗vendor |
| H200-SXM | NVIDIA HGX H200 (gpu-h200-sxm platform) eu-north1 (Finland), eu-west1 (Paris, France), us-central1 (Kansas City, MO) +1 more | $4.50 | on-demand | A | publ | InfiniBand, 400 Gbps per GPU / 3.2 Tbps per 8-GPU node, GPUDirect RDMA (per docs.nebius.com) | ↗vendor |
| B200 | NVIDIA HGX B200 (gpu-b200-sxm platform) us-central1 (Kansas City, MO), me-west1 (Israel) | $7.15 | on-demand | A | publ | InfiniBand, 400 Gbps per GPU / 3.2 Tbps per 8-GPU node, GPUDirect RDMA (per docs.nebius.com) | ↗vendor |
Grades measure what sellers publish — deliverability of the written claim, not measured performance. No fleet holds a measured (Verified) badge yet; that tier arrives with the verification harness. Source classes: API = provider's live pricing API · vendor = provider's published page · mkt API = marketplace API · 3rd-party = aggregator (provider publishes no price on its own pages) · quote = quote-gated or "starting at", not a walk-up price. Quote-class prices are shown in tables but excluded from headline ranges, spreads, and fair-price bands.
Terms & trust (as published, 2026-09-14)
| Uptime SLA | 99.50% Service Uptime for Compute Cloud (single VM). Other published levels per service: Managed Kubernetes, PostgreSQL, MLflow, Object Storage, VPC each have their own service-level pages; Standalone Applications is 95.00%. No GPU-cluster-specific SLA published — compute SLA has no GPU-specific service level terms. — measured on: Single Virtual Machine — unavailability defined as 'Loss of external connectiveness or a boot disk of a Virtual Machine'; separate tier for non-replicated disk placement groups (two or more disks in error state within an 8-hour interval). Measured per Reporting Period; service levels can vary by region. DoS/DDoS on the customer's VM is excluded. |
| SLA remedy | credits — compensation as a discount against service fees in the reporting period: 10.0% of the service fee if uptime is 99.50%-99.00%, 15.0% if 99.00%-95.00%, 30.0% if below 95.00%. 'Compensation is the sole and exclusive remedy'; credits cannot exceed total cost of service usage in the period, apply only as discounts to fees, no cash refunds. |
| Capacity guarantee | Not published as a guarantee. Reserved/committed capacity ('Pay up to 35% less than on-demand rates by reserving large-scale clusters for multiple months') is sales-negotiated via addendum; docs state 'Extending the discounts is subject to the available capacity at the time of the request' and cancellation only 'if allowed by the terms of your addendum'. No standard published capacity-guarantee terms. |
| Egress fees | No general compute/VM egress charge published: 'The Virtual Networks service in Nebius AI Cloud is provided free of charge' (VPC pricing page lists no traffic pricing). Object Storage egress IS charged: $0.0150 per GiB (standard); Intelligent tier egress 'Free*' limited to 3x the customer's monthly average stored capacity; Enhanced storage egress free. |
| Billing granularity | per-second — docs state 'Billing unit: 1 second' with 'Pricing unit: 1 hour (3600 seconds)' for GPUs, vCPUs and RAM; charges below one hour prorated; 'Computing resources of stopped VMs are not charged (this does not apply to storage volumes)'. Serverless AI also bills per-second. |
| Self-serve signup | yes |
| Security certifications | Trust Center lists: SOC 2 Type II 'With HIPAA', SOC 3, ISO 27001, ISO 27018, ISO 27032, ISO 27701, ISO 27799, ISO 22301, ISO 42001 (AI Management Systems), CSA STAR Level 1, alignment with NIS 2 and DORA, and 'GDPR and CCPA-compliant policies and practices'. |
Notes: Self-serve: individuals (bank card only) and companies (bank card except Israel, or bank transfer); activation 'Automatic after adding a card and paying the initial $25'; only Israel-based companies require manual activation. On-demand GPUs (e.g. H100 $3.85/GPU-hr) available without sales; committed discounts and GB200/GB300 NVL72 require sales. Umbrella SLA at docs.nebius.com/legal/sla defers all percentages to per-service pages; SLA applies only to fee-paying usage and excludes customer actions, third-party software, preventive maintenance, emergency works, government requests, AUP violations, quota breaches, force majeure. Compensation must be claimed and is capped at the period's spend. Compute SLA covers single VMs only; no published cluster-wide or GPU-health SLA.
Sources: ↗ ↗ ↗ ↗ ↗ ↗ ↗ ↗ ↗ · Fields marked "not published" are exactly that — the seller does not state them anywhere we could find. Absence is a finding.
Reliability record (public status feed, last 90 days)
14 incidents on Nebius's own status page in the trailing 90 days, collected 2026-09-15. Self-reported by the provider — a floor on true incident count, not a ceiling; it still beats the war-story thread you were going to read instead. High counts often reflect granular per-rack reporting, not worse infrastructure: read the incidents, not just the number.
| Date | Incident | Impact | Duration |
|---|---|---|---|
| 2026-09-09 | Unstable compute VM create ↗ | minor | 0.4 h |
| 2026-09-01 | Connectivity issues for newly created worker nodes ↗ | major | 0.3 h |
| 2026-08-19 | Region us-central1 partial degradation ↗ | major | 21.6 h |
| 2026-08-17 | Issues logging to the Management Console ↗ | minor | 2 h |
| 2026-08-15 | Partial InfiniBand unavailability ↗ | major | 3.3 h |
| 2026-08-12 | External VPC connectivity issues ↗ | major | 0.4 h |
| 2026-08-05 | Creation of new VMs is broken ↗ | none | 1.3 h |
| 2026-07-27 | Network issues ↗ | major | 1.6 h |
| 2026-07-22 | [eu-north1] Network timeouts affecting multiple services ↗ | major | 12 h |
| 2026-07-14 | InfiniBand switch failure affecting GPU workloads in the eu-north1-c fabric in the eu-north1 region ↗ | major | 0.4 h |
| 2026-07-03 | Managed SkyPilot OIDC Login is unavailable ↗ | minor | 0 h |
| 2026-07-01 | Monitoring, Object Storage and Token Factory services are partially unavailable ↗ | major | 1.7 h |
What Nebius does not publish
- max_contiguous_gpus / max cluster size
- InfiniBand fabric generation on fetched docs page
- VRAM per GPU (not stated on Nebius pages; standard SKU values used)
- reserved/committed rate card (only 'up to 35% less' claim)
- GB200 NVL72 and GB300 NVL72 prices (contact sales)
- region-to-SKU mapping
Unpublished fields are recorded as unknown and grade conservatively. Sellers improve their grades by disclosing — that is the point.
Weekly context on every provider
The Delivered Compute Report tracks how these prices and grades move.
Subscribe free