Skip to main content
This document defines Service Level of Compute Cloud (“Service”) and is an integral part of Nebius Services Agreement (“Agreement”) and SLA available at: https://docs.nebius.com/legal/sla (the “General SLA”). Capitalized terms used herein but not defined herein shall have the meanings set forth in Agreement and Linked Documents, including the Terms of Use of Compute Cloud available at https://docs.nebius.com/legal/specific-terms/compute. In the event of a conflict between this document and the General SLA, this document shall govern solely with respect to the definition of Unavailability, the commencement and end of Service Downtime for the Service, and the Compensation amounts set out herein. This document sets out separate Service Levels for Virtual Machines and for Rack-Scale GPU System. Service Levels for Disks, Shared Filesystems and other storage services are set out in the applicable Service Levels available at https://docs.nebius.com/legal/sla-levels.

Service Level for Virtual Machines

This section applies to Virtual Machines other than Virtual Machines within a Rack-Scale GPU System.

Compensation Amount

Service Uptime Percentage is measured per single Virtual Machine.

Definition of Unavailability of a single Virtual Machine

A Virtual Machine is Unavailable during any period, in which any of the following conditions is not met: (a) all GPUs allocated to the Virtual Machine under the selected platform and preset, if any, are present and not in a Failed State; (b) where the selected platform and preset provide NVLink connectivity between the GPUs allocated to the Virtual Machine, such NVLink connectivity is available; (c) where the selected platform and preset provide connectivity to a Cluster Interconnect Fabric, the Virtual Machine has connectivity to such Cluster Interconnect Fabric; (d) the Boot Disk of the Virtual Machine is accessible to the Virtual Machine; and (e) the Virtual Machine has external network connectivity. Where Nebius notifies Customer via the Management Console that the Virtual Machine can be restarted on another Host, Unavailability ends upon such notification. The period between the notification and the completion of the stop and start of the Virtual Machine by Customer or by Nebius is not Unavailability. If, following the notification, the Virtual Machine cannot be restarted on the Host made available for reasons attributable to Nebius, Unavailability is deemed to have continued. Where a Virtual Machine or a Rack-Scale GPU System exists for part of a Reporting Period only, Maximum Service Uptime for that Virtual Machine or Rack-Scale GPU System is the number of minutes from the later of its creation and the start of the Reporting Period until the earlier of its deletion and the end of the Reporting Period. A Virtual Machine or a Rack-Scale GPU System exists from the creation until the deletion of the corresponding resource in the Nebius Compute API. For the avoidance of doubt, a Virtual Machine shall not be deemed Unavailable due to any of the following: (i) Customer workload slowdowns or failures; (ii) errors at the model, framework, library, container, or application level; (iii) distributed training inefficiency or benchmark degradation; (iv) reduced bandwidth or increased latency of NVLink or Cluster Interconnect Fabric that does not result in loss of connectivity; (v) issues with Customer’s software or configuration, including the guest operating system and any drivers or firmware modified by Customer; (vi) Unavailability of Disks other than the Boot Disk, of Shared Filesystems, or of other storage services, which is governed by the applicable Service Levels for such services; (vii) loss of external network connectivity due to circumstances beyond Nebius’ control, including DoS or DDoS attacks on Customer’s Virtual Machine; or (viii) periods during which the Virtual Machine is stopped or deleted by Customer or by software acting on Customer’s behalf.

Service Level for Rack-Scale GPU System

This section applies to Virtual Machines within a Rack-Scale GPU System. Such Virtual Machines are not subject to the Service Level for Virtual Machines and are not eligible for Compensation under it.

Rack-Scale Product Table

Compensation Amount

Service Uptime Percentage is measured per Rack-Scale GPU System. The compensation schedule below applies to each Rack-Scale GPU System of Customer.

Definition of Unavailability of a Rack-Scale GPU System

A Rack-Scale GPU System is Unavailable during any period in which fewer Nodes in its Rack are Functional and Usable than the number of SLA-Covered Nodes specified in the Rack-Scale Product Table for the applicable product. A Node is not Functional and Usable during any period in which any of the following conditions is not met for the Virtual Machine deployed on it: (a) all GPUs allocated to the Virtual Machine are present and not in a Failed State; (b) all GPUs allocated to the Virtual Machine have NVLink connectivity to the Rack’s NVLink domain via the Rack’s NVLink switches; (c) the Virtual Machine has connectivity to the Cluster Interconnect Fabric; (d) the Boot Disk of the Virtual Machine is accessible to the Virtual Machine; and (e) the Virtual Machine has external network connectivity. Where no Virtual Machine is deployed on a Node, the Node is not Functional and Usable during any period in which the Service does not report it as available for launching a Virtual Machine, as described in the Documentation. For the avoidance of doubt, impairments affecting a number of Nodes equal to or lower than the Maintenance Buffer at any point in time do not constitute Unavailability of the Rack-Scale GPU System and are not counted as Service Downtime. A failure of a Rack-Level Component constitutes Unavailability of the Rack-Scale GPU System only if, as a result, fewer Nodes than the number of SLA-Covered Nodes remain Functional and Usable. Where Nebius notifies Customer via the Management Console or other means described in the Documentation that a replacement Rack is available for relocation in accordance with clause 1.5 of the Terms of Use of Compute Cloud, Unavailability ends upon such notification. The period between the notification and the completion of the relocation is not Unavailability. If, following relocation, fewer Nodes than the number of SLA-Covered Nodes are Functional and Usable on the replacement Rack for reasons attributable to Nebius, Unavailability is deemed to have continued. The exclusions listed in items (i) through (viii) of the Service Level for Virtual Machines apply to this section accordingly.

Definitions

Boot Disk means the disk from which the Virtual Machine’s operating system is loaded, as attached to the Virtual Machine at the time of its creation or subsequently in accordance with the Documentation. Cluster Interconnect Fabric means high-bandwidth network connecting Virtual Machines within a cluster for inter-node communication, distinct from the external network, as described in the Documentation. Failed State means, in respect of a GPU, that the GPU is not exposed to the Virtual Machine or is reported by Nebius’ monitoring system as failed and pending reset or replacement. Correctable errors, performance degradation, and errors resulting from Customer’s workloads or software do not constitute a Failed State. Functional and Usable means, in respect of a Node, that none of the circumstances listed in the section “Definition of Unavailability of a Rack-Scale GPU System” applies to it. Maintenance Buffer means the Nodes of a Rack in excess of the number of SLA-Covered Nodes, as specified in the Rack-Scale Product Table. Nodes in the Maintenance Buffer may be used by Customer but are not counted toward Service Uptime calculations. SLA-Covered Nodes means the number of Nodes per Rack counted toward Service Uptime calculations under the Service Level for Rack-Scale GPU System, as specified in the Rack-Scale Product Table.
Web address: https://docs.nebius.com/legal/sla-levels/compute Publication date: October 6, 2026 Effective date: October 6, 2026 Previous version of the document: https://docs.nebius.com/legal/archive/sla-levels/compute-20240925