Skip to main content
This article lists what has been added, updated, fixed or removed in Nebius AI Cloud services.
​
  • Added: You can now set project limits (capacity allowance) on a capacity block group to cap how many GPUs each project in your tenant can consume. See Managing project limits for how to view, set, remove and disable limits from the web console or CLI.
  • Added: Managed Service for Kubernetes® load balancer services now support the nebius.com/load-balancer-pool-id annotation. Use it to allocate the load balancer’s public IP address from a specific VPC IP pool instead of the default subnet pool.
  • Updated: MysteryBox is now called SecretStash. The service, resources and functionality are unchanged; only the display name and Terms of Use were updated.
  • Updated: Creating a Compute virtual machine in the web console is now a step-by-step wizard with General, Compute, Storage and other sections. If you have capacity block groups, the General step lets you choose between a Reserved VM (capacity assured by your CBGs) or a Pay-as-you-go VM, and the Compute step exposes the full reservation configuration, including Switch to PAYG behavior when a reservation is exhausted.
​
​
  • Added: You can now create disk snapshots in Compute to capture a point-in-time copy of a disk. Use snapshots to back up disks before risky operations, clone disk state or restore data from a snapshot.
  • Added: Managed Service for Kubernetes® node groups now integrate with Virtual Networks security groups. You can assign security groups to a node group and control traffic to and from the nodes.
  • Updated: You can now resize a disk of a Managed Service for PostgreSQL® cluster without any need to create the cluster again. Disk size can only be increased, not decreased.
​
  • Added: Managed Service for Soperator now supports ephemeral nodes. You can release specific worker nodes from a Soperator cluster and reuse the same capacity block group for inference workloads in a separate Managed Service for Kubernetes® cluster, without taking down the entire Soperator cluster.
  • Added: Serverless AI endpoints now expose each HTTP port through a managed https:// URL, and each TCP port through a managed tls://<host>:443 URL — both encrypted. A public IP address is no longer required to reach an endpoint from the internet.
  • Added: When you create a Serverless AI endpoint, you can now use the --inject-file option to mount local files into the endpoint container at launch. Mounted files are read-only and limited to 64 KiB.
  • Added: Object Storage lifecycle rules can now filter objects by tags. Use tag-based filters to apply expiration and storage class transitions to a subset of objects in a bucket.
  • Added: New tutorial: run OpenMM molecular dynamics simulations as a Serverless AI job. Package OpenMM in a Docker image, push it to Container Registry and run CUDA®-accelerated protein simulations that read and write results to an Object Storage bucket.
  • Updated: Managed Service for Kubernetes now uses a new default deployment strategy for node group updates: nodes are replaced one at a time (max_unavailable=1, max_surge=0), so rolling updates no longer require additional Compute quota by default. To keep the previous behavior of creating a replacement node before removing the existing one, set a non-zero surge value on the node group.
​
  • Added: Nebius Tunnels is now available in preview. You can expose local HTTP and TCP services to the internet through an outbound agent connection, without opening inbound firewall ports.
  • Added: You can now use local SSD disks for high-performance, low-latency ephemeral storage on supported Compute virtual machines, Managed Service for Kubernetes® node groups and Managed Service for Soperator clusters.
  • Added: You can now use budgets in Billing to track spending against a target amount and receive alerts when your costs cross configured thresholds.
  • Added: Object Storage now supports the Intelligent storage class. It automatically switches rarely accessed objects to a lower-cost tier while keeping objects in the same location and preserving storage performance.
  • Added: Documentation now includes instructions about how to share an object with a pre-signed URL in Object Storage. Pre-signed URLs grant temporary download access without requiring the recipient to have Nebius AI Cloud credentials.
  • Added: You can now attach a stopped virtual machine’s boot disk to another VM to inspect or repair it when the original VM does not start or you cannot connect to it.
  • Updated: The uk-south1 region is now public. Services available in this region are listed in Regions.
  • Updated: You can now create one-time usage data exports with the Nebius AI Cloud CLI. Use the exported FOCUS 1.2 CSV data for cost analysis and reporting in external tools.
​
​
  • Added: A free trial is now available for individual users in the eu-north1 region. The trial includes 30 days of access and $5 in credit to explore the platform.
  • Added: You can now use the Nebius AI Cloud SDK for Go to interact with Nebius AI Cloud services from Go applications. Manage resources, handle authentication and call Nebius AI Cloud services without building API requests manually.
  • Added: You can now send traces directly to Nebius AI Cloud from applications that use OpenTelemetry. Configure an OpenTelemetry exporter to write traces to a regional Tracing writer endpoint.
​
  • Added: Documentation now includes instructions for checking Object Storage performance by using Warp.
  • Updated: You can now run nebius ai create to create a Serverless AI endpoint or job through an interactive, step-by-step workflow. The command guides you through the required configuration options and pre-fills common values, making endpoint and job setup easier.
​
​
  • Added: You can now import your own image into Compute from an Object Storage bucket. Upload a .vmdk, .vhd, .raw or .qcow2 file to a bucket and create a custom image to launch boot disks and virtual machines from it.
  • Added: Audit Logs supports two roles, auditlogs.audit-event-viewer and auditlogs.audit-event-exporter. They grant users granular access to view audit events and manage audit log exports.
  • Added: Topology-aware scheduling is now available for GPU workloads in Managed Service for Kubernetes®. Nebius AI Cloud exposes InfiniBand™ topology information as node labels, so schedulers can place workloads on nodes that are physically closer in the network hierarchy. This improves communication performance.
  • Added: Managed Kubernetes documentation now includes the default kubeReserved values applied on worker nodes. For example, you can find out how the memory reservation is calculated for GPU nodes based on RAM size.
  • Added: Object Storage now supports object integrity checks. When you upload an object to a bucket, send a checksum with your upload. Object Storage verifies the checksum and integrity of the object on receipt.
  • Added: IAM documentation now includes instructions for managing federations. You can create, activate and deactivate federations, change their settings, and upload or delete certificates.
  • Added: CLI documentation now includes filtering output with JSONPath. Use the --format jsonpath='<expression>' parameter to extract values such as a resource ID, name or IP address without post-processing the full output.
  • Added: Object Storage now supports data transfers. Use them to replicate and back up your data from one bucket to another, or to migrate from a particular region or a third-party storage service.
  • Updated: Tax handling for companies has changed. Reverse charge may now apply to companies that provide a valid VAT ID (EU, UAE, UK) or GSTIN (India, Singapore) in their billing setup. Individuals in Israel are now subject to Israeli VAT, like companies.
  • Removed: NVIDIA NIM™ microservices for computational drug discovery are no longer available as standalone applications in Nebius AI Cloud.
​
  • Added: You can now access instance metadata from within a Compute virtual machine (VM) through the HTTP instance metadata service (IMDS). Retrieve instance details, labels, network and user data, service account tokens and live maintenance events without calling the Nebius AI Cloud API from outside the VM.
  • Updated: The legacy /mnt/cloud-metadata mount on older VM images is deprecated and will be removed on September 30, 2026. Use the HTTP instance metadata service instead.
  • Updated: The ubuntu22.04-cuda12 and ubuntu22.04-driverless boot disk image families will be deprecated on June 1, 2026. Existing disks remain available, but you can no longer create new disks from these families after that date. Migrate to Ubuntu 24.04 images before the deadline.
  • Updated: When setting up GPUs in Managed Kubernetes, the cuda12.4 driver preset (Ubuntu 22.04) is no longer available. Use cuda12.8 or cuda13.0 on Ubuntu 24.04.
​
  • Added: You can now create a Compute virtual machine and its disks together in a single CLI command. These managed disks share the VM’s lifecycle and are deleted automatically with it, which makes them a good fit for ephemeral workloads such as training jobs or inference clusters.
  • Added: You can now tune TCP window sizes on Linux virtual machines to improve per-flow throughput on high-bandwidth, high-latency connections.
  • Updated: You can now set the block size for a disk or shared filesystem when creating it in the web console, in addition to the CLI and Terraform provider.
  • Deprecated: The nebius storage v1alpha1 transfer CLI commands and the nebius_storage_v1alpha1_transfer Terraform resource are deprecated and supported until 2026-06-01. Use the v1 data transfer API instead; existing transfers remain accessible through the v1 API. See the CLI release notes and Terraform provider release notes.
​
  • Added: The NCCL Inspector by NVIDIA® is now available on demand for Soperator clusters in Nebius AI Cloud. When enabled for a job, it collects per-communicator, per-collective performance metrics from NCCL operations and sends them to the Grafana® installation included with your cluster.
  • Added: Soperator clusters now include built-in health checks and an auto-healing system for worker nodes. Critical GPU, RAM and network checks run on a schedule, and broken nodes are isolated and replaced automatically.
  • Added: You can now configure single sign-on with Okta.
  • Added: Compute virtual machines now expose maintenance reason codes that explain why a maintenance event was triggered.
  • Added: You can now retrieve the public IPv4 CIDR blocks available for allocations in a project. The available ranges depend on the project region.
  • Added: You can now collect diagnostic logs from Compute virtual machines and from Managed Kubernetes clusters, including GPU, InfiniBand™ and node-level logs for troubleshooting.
  • Added: A new all-reduce NCCL performance test is now included in Soperator. Use the pre-built single-node and multi-node versions to validate collective communication between GPUs over NVLink and InfiniBand and assess network performance.
  • Updated: SkyPilot as a standalone application is now available in the uk-south1 region.
  • Updated: The Observability Overview page in the web console is now the new entrypoint for monitoring, providing information on firing alerts, key resource metrics and recent logs in a single view.
​
  • Added: Compute disks and shared filesystems now support deletion protection. Enable it to prevent accidental deletion in automated environments such as CI/CD pipelines or Terraform configurations.
  • Added: The Nebius AI Cloud CLI now supports an interactive mode. To enable it, add -i or --interactive to a command. The CLI prompts for parameters step by step, validates each value and shows the resulting configuration before you apply it.
  • Added: The log export to Object Storage now supports the Nebius AI Cloud CLI.
  • Added: You can now get a list of public boot disk images available in a given region. The list includes image families, recommended platforms and unsupported platforms for each image.
  • Added: When Prometheus takes too long to retrieve metrics of large clusters, shard and split their scraping jobs.
  • Updated: The AWS CLI is now documented as a supported interface for configuring lifecycle rules in Object Storage, alongside the Nebius AI Cloud interfaces.
​
  • Added: The capacity advisor is now available in the web console. It shows GPU capacity availability across regions, platforms and presets so you can check where you can launch virtual machines before creating them.
  • Added: You can now monitor data plane events for Audit Logs in the web console. Graphs show processed events, quota usage and dropped events so you can detect when Object Storage activity approaches the quota.
  • Updated: You can use the web console to create, configure and delete Serverless AI endpoints.
  • Updated: You can manage authorized keys for service accounts from the web console and the Terraform provider, in addition to the CLI.
  • Updated: The SkyPilot integration tutorial now includes instructions to use the Managed SkyPilot API Server, a standalone application available in the Nebius AI Cloud console.
  • Updated: The Nebius Service Health Status page now displays all timestamps in UTC (GMT+00:00).
​
​
  • Added: The NVIDIA® RTX PRO™ 6000 virtual machine platform is now available in the us-central1 region. It features 96 GB of GDDR7 memory per GPU and comes in single-GPU and eight-GPU presets.
  • Added: Serverless AI CLI commands now support MysteryBox secrets. You can pass secrets as environment variables, registry credentials or tokens when creating endpoints and jobs, without exposing sensitive values in plain text.
  • Updated: You can now manage capacity reservations for Compute virtual machines and Managed Kubernetes node groups in the web console. Previously, this required the CLI or Terraform provider.
  • Updated: You can now export logs directly from individual resource pages in the web console, in addition to the central Observability Logs page.
  • Updated: The default retention period for logs in Observability services is now 14 days.
​
  • Added: Object versioning is available in Object Storage. When enabled, each update to an object creates a new version, which protects data from loss and allows you to restore previous versions.
  • Added: You can now create custom boot disk images to tailor images to your needs and create multiple virtual machines with custom disk architecture.
  • Added: You can also create custom boot disk images by using the Nebius Packer plug-in. Use it to pre-install packages to a disk, prepare versioned images and standardize environments across teams or workloads.
  • Added: You can now use the Nebius Terraform provider with Pulumi to declare your Nebius AI Cloud resources in the preferred programming language instead of using Terraform configuration.
  • Added: You can now export logs from Observability to an Object Storage bucket in JSON or Parquet format for long-term retention or analysis in external tools.
  • Updated: You can now attach and detach disks on running Compute virtual machines without stopping them first.
​
​
​
  • Added: Virtual machines now support secondary private IP addresses that you can use as a backup option in case of incidents.
  • Added: Nebius VPN Gateway, an open-source, VM-based IPsec gateway provided by Nebius Professional Services, is now available. You can deploy it to establish site-to-site VPN connectivity with external networks in other cloud providers or on-premises data centers.
  • Added: Nebius AI Cloud now calculates taxes for the residents of Rhode Island in accordance with the United States tax law.
  • Updated: For viewing ingested traces from Managed Kubernetes clusters in Grafana®, the connection URL is now https://read.tracing.api.nebius.cloud/projects/<project_ID>/tempo instead of https://read.tracing.api.nebius.cloud/tempo/<project_ID>.
​
​
​
  • Added: In Serverless AI, you can now debug failed jobs by keeping the job’s VM alive to connect via SSH.
  • Added: Single sign-on with JumpCloud is now available. Users can sign in to Nebius AI Cloud with their JumpCloud credentials instead of a separate password.
  • Added: You can now export audit events to an Object Storage bucket for a specific time range. Use exported data for long-term retention, data portability or analysis in external tools.
  • Added: You can now collect traces with Nebius Observability Agent for Kubernetes and view traces in Grafana. Send traces from cluster applications via OTLP and explore them in Grafana with Tempo.
  • Updated: We now recommend GPU boot disk images with Ubuntu 24.04 and CUDA 12 or CUDA 13.0. The ubuntu22.04-cuda12 image is not recommended and will be deprecated.
  • Updated: When setting up GPUs in Managed Kubernetes, you can now choose between driver presets cuda12.8, cuda13.0 and cuda12.4. You can also change the driver preset by updating the node group when you need a different driver or CUDA version.
​
  • Added: Serverless AI jobs and endpoints are now available. You can use them to run AI workloads without any need to create and maintain infrastructure for them.
  • Added: Taxes are now included in invoices for individuals from India, Singapore, the United Arab Emirates and the United Kingdom.
​
​
  • Added: Object Storage now supports logs of two types: control plane logs and data plane logs. By using them, you can get detailed information about actions on buckets and objects.
​
​
  • Added: You can now set up billing data export and automatically get usage data for cost analysis.
​
  • Updated: The me-west1 region is now public. Services available in this region are listed in Regions.
​
  • Updated: The default Compute quotas on the numbers of NVIDIA B200 and NVIDIA H200 GPUs in the us-central1 region per tenant are now zero, down from 32 and 16, respectively. This means that in tenants created after December 1, you need to request a quota raise before creating VMs with any number of these GPUs in us-central1 projects. The change does not affect existing tenants.
​
  • Added: Two new virtual machine platforms are now available for Compute virtual machines, Managed Kubernetes nodes and other resources:
    • NVIDIA® B300 NVLink with Intel Granite Rapids (ID gpu-b300-sxm) is now available in the private region uk-south1. It introduces NVIDIA B300 GPUs to Nebius AI Cloud.
    • NVIDIA® B200 NVLink with Intel Emerald Rapids (ID gpu-b200-sxm-a) is now available in the private region me-west1. This is the second platform with NVIDIA B200 GPUs; a platform with the same name (ID gpu-b200-sxm) is available in the public region us-central1.
  • Added: When creating a support ticket in the support center, you can now choose the topic of the ticket (technical, billing or digital rights). Available priority levels for technical tickets have also changed: the standard level is renamed medium, and new high and low levels are added to the existing levels. This helps us process your requests faster and more efficiently.
​
​
​
  • Updated: The default Compute quota on the number of NVIDIA H200 GPUs in the us-central1 region per tenant is now 16, down from 32. The change does not affect existing tenants.
  • Updated: When you manually install the NVIDIA GPU Operator on your InfiniBand-interconnected Managed Kubernetes nodes, GPUDirect RDMA is now enabled by default.
​
  • Added: Monitoring dashboards for Managed PostgreSQL clusters now include two new connection pooler metrics, Pooler clients active and Pooler clients waiting.
​
  • Added: Managed Kubernetes now supports Kubernetes 1.32. This version is selected for new clusters and node groups by default. See the overview of Kubernetes versions.
  • Added: The list of roles has been extended. In addition to general roles (admin, editor, viewer, auditor), you can now assign new service roles to custom groups, allowing for more granular access to your resources. The new roles control access to Object Storage and MysteryBox resources, as well as Data Subject Requests to support.
  • Updated: Starting October 22, usage of virtual machines and applications that use NVIDIA B200, H200 and H100 GPUs is recorded under new billing items that cover GPUs, vCPUs and RAM together rather than separately. Usage from October 1–21 is recalculated retroactively. This makes monitoring your usage simpler and does not affect your costs. Learn more.
​
​
  • Updated: Automatic security updates are now disabled by default in boot disk images for non-GPU virtual machines in Compute, in addition to GPU VMs, to prevent errors in running workloads. You can enable them back.
  • Updated: Computing resources of all Compute virtual machines are now eligible for committed usage discounts, as opposed to only VMs with NVIDIA H100 GPUs.
​
  • Added: The MysteryBox service is launched, allowing you to store sensitive data in an encrypted form to reuse them in your scripts, configuration files or applications.
  • Added: You can now create routing tables and routes to customize the routing of egress traffic and implement advanced networking scenarios, such as custom NAT gateways or segmented architectures.
  • Added: Your committed usage discounts are now grouped in orders that correspond to the agreement addendums that you sign.
  • Updated: Automatic security updates are now disabled by default in boot disk images for GPU virtual machines in Compute to prevent errors in running workloads. You can enable them back.
  • Updated: Default Compute quotas for GPU and non-GPU virtual machines are now separate, allowing new customers to create more VMs overall and control the remaining capacity in their projects more precisely.
​
  • Added: You can now create Compute virtual machines with container images deployed on them: containers over VMs. You can use images provided by Nebius AI Cloud or custom Docker images from public registries.
  • Added: Two new extensions for PostgreSQL are now available on all Managed Service for PostgreSQL clusters: RUM offers enhanced full-text search, and pg_repack allows you to remove bloat from tables and indexes.
  • Updated: If you pay for Nebius AI Cloud services by bank card, we can now suspend your account after one unsuccessful attempt to charge your card rather than after multiple attempts. Check regularly that your payment method is up to date.
​
  • Added: Managed Kubernetes now performs health checks on nodes, and automatically recovers nodes that are not ready or have issues with the boot disk or GPUs. You can enable and disable these health checks.
  • Added: Audit Logs now support filtering events by more fields, so you can find events that occurred in a specific region, are associated with operations that have a specific status, etc.
  • Updated: The navigation in the web console has undergone a major overhaul to group and present services and resources more clearly.
  • Updated: The upload and download speed for the Standard storage class in Object Storage is now limited to 20 GBps per tenant in all regions.
Postgres, PostgreSQL and the Slonik Logo are trademarks or registered trademarks of the PostgreSQL Community Association of Canada, and used with their permission. InfiniBand and InfiniBand Trade Association are registered trademarks of the InfiniBand Trade Association.