Compute Engine release notes

Generally available: Compute flexible committed use discounts (CUDs) are available for G2 and G4 GPU accelerator-optimized machine series. The supported resources include vCPUs, memory, Local SSD disks, and GPUs.

Compute flexible CUDs are spend-based CUDs that apply to eligible Google Cloud spend across Compute Engine, GKE, and Cloud Run. For G2 and G4 machine series, compute flexible commitments provide the flexibility to switch between eligible machine series and regions depending on your workload needs. For GPUs that belong to these machine series, compute flexible commitments don't require attached reservations.

For more information, see Compute flexible CUDs.

A vulnerability (CVE-2026-6726) in the Trusted Computing Group's TPM 2.0 reference implementation code was discovered and is being addressed. For more information, see the GCP-2026-054 security bulletin.

Changed: The minimum provisioned throughput for a Hyperdisk ML volume attached to more than 20 instances is 20 MiB/s per instance, reduced from 100 MiB/s.

For more information, see Share a Hyperdisk ML volume between instances.

Public preview: you can create a machine image that includes or excludes specific non-boot disks attached to a source Compute Engine instance. For more information, see Create a machine image from specific disks. You can also create a new disk by restoring a specific individual disk from a machine image without creating a new Compute Engine instance. For more information, see Create a disk from a machine image.

Hyperdisk Balanced volumes on C4D instances have increased maximum throughput limits for these machine types:

• c4d-*-96: 3,125 MiB/s (up from 2,800 MiB/s).
• c4d-*-192: 6,250 MiB/s (up from 4,800 MiB/s).
• c4d-*-384: 12,500 MiB/s (up from 10,000 MiB/s).

For detailed performance limits, see Hyperdisk Balanced performance limits when attached to an instance.

The maximum IOPS per GiB for Hyperdisk Balanced Storage Pools have increased from 4 IOPS per GiB. The new limits depend on the provisioning type:

• Standard performance: 30 IOPS per GiB
• Advanced performance: 6 IOPS per GiB

For more information, see Limits for Hyperdisk Storage Pools.

You can observe real-time virtual machine (VM) distribution across zones, machine types, and instance states in your managed instance groups (MIGs) by using the GCE MIG Instance Distribution Monitoring dashboard in Cloud Monitoring. When your group uses location flexibility across zones, instance flexibility across machine types, or both, this visibility helps you monitor capacity allocation and diagnose runtime fallback behavior. For more information, see Monitor instance distribution in MIGs.

Generally available: You can configure a regional managed instance group (MIG) to allow a VM repair in an alternate zone when the MIG can't repair the VM in its original zone. Repairing a VM in an alternate zone can help to improve your application's resiliency and resource obtainability. For more information, see Repair a VM in an alternate zone.

Encrypting disks, snapshots, images, and machine images with customer-supplied encryption keys (CSEKs) is deprecated and will be disabled on July 20, 2027.

For more information and alternatives to CSEKs for your Compute Engine resources, see Deprecation of customer-supplied encryption keys (CSEK) in Compute Engine.

Hyperdisk Balanced High Availability volumes on C4 instances have increased performance limits for several machine types. For example, an instance that uses the c4-standard-16 machine type can reach up to 1,600 MiB/s of throughput, up from 600 MiB/s.

For detailed performance limits, see Hyperdisk Balanced High Availability performance limits when attached to an instance.

Changed: The following operations on the boot disk of a Compute Engine instance that has a service account attached no longer require the iam.serviceAccounts.actAs permission. In the following list, the boot disk of such an instance is referred to as the source disk.

• Creating a standard or archive snapshot of the source disk.
• Cloning the source disk.
• Creating a machine image of the instance.
• Creating a custom image of the source disk.
• Starting asynchronous replication of the source disk to another region.
• Creating a new disk when you create an instance, if the new disk is created from an instant snapshot of the source disk.

Preview: The network-optimized C4N machine series offers machine types with 375 GiB to 12,000 GiB of attached Titanium SSD.

To use C4N machine types with attached Local SSD disks, you can Request preview access.

For more information, see C4N machine series.

Preview: Advanced Compute Images provide high-performance images to support your artificial intelligence (AI), machine learning (ML), and high-performance computing (HPC) workloads on Google Cloud.

Advanced Compute Images provide a single source of trusted, performance-tuned OS images that remove the need for manual image building for specialized workloads. Each image version is pre-installed with the necessary drivers, network fabrics, and Slurm agents to help you run your workloads.

Generally available: The network-optimized C4N machine series is generally available for Compute Engine and Google Kubernetes Engine (GKE) customers. Powered by 5th generation Intel Xeon Scalable processors (Emerald Rapids), C4N instances are purpose-built for network- and block storage-intensive workloads such as:

• Network and security appliances
• High-performance databases
• High-scale data analytics
• Distributed filesystems

The C4N machine series delivers the highest I/O performance available in Compute Engine, supporting up to 400 Gbps of network bandwidth and up to 95 million packets per second (Mpps) of sustained packet processing performance. C4N also offers leading block storage performance with Hyperdisk Extreme that scales up to 25 GiB/s of bandwidth and 1M IOPS. C4N instances are available in predefined machine shapes with three different vCPU to memory ratios, ranging in size from 2 to 192 vCPUs and up to 1,488 GB of DDR5 memory.

For C4N machine types with attached Local SSD disks, you can Request preview access.

For more information, see C4N machine series.

Generally available: You can cancel a future reservation request in calendar mode to prevent Compute Engine from provisioning your requested resources and incurring unnecessary charges. For more information, see Delete a future reservation request in calendar mode.

Generally available: In a managed instance group (MIG), you can use a health check to monitor your application health without triggering repairs for an unhealthy VM, if the application fails the health check. You can prevent the MIG from repairing an unhealthy VM by turning off autohealing. For more information, see Turn off repairs in a MIG.

Preview: You can use Gemini in the Google Cloud console as an AI-powered interface to evaluate hardware options, estimate deployment costs, and view recommended configurations for your Compute Engine instances. Prompting Gemini helps you reach an optimal configuration for your workload before you create or modify a compute instance. For more information, see Design your compute infrastructure with Gemini.

Generally available: You can create instances all at once in a regional managed instance group (MIG) by using resize requests. For more information, see About resize requests in a MIG.

Generally available: Resource-based committed use discount (CUD) recommendations are available for GPUs, Local SSD disks, and premium operating system (OS) licenses.

CUD recommendations provide insight into any additional commitments that you can purchase to optimize the costs of the resources that you run. You can use these recommendations and purchase commitments for resource usage that isn't covered by commitments and is being charged at list prices. Google Cloud analyzes your compute instance spending trends with and without a commitment and generates CUD recommendations on a monthly basis.

For more information about how CUD recommendations are generated, what resource types are supported, and how to use recommendations to purchase commitments, see Get recommendations for committed use discounts (CUDs).

For resource-based committed use discounts (CUDs), the default value of CUD scope for most Cloud Billing accounts has changed from Project to Billing account. If the CUD scope is set to Billing account, then resource-based CUDs from a commitment are shared across all projects in that account. If the CUD scope is set to Project, then resource-based CUDs from a commitment are available to only the project in which you purchased that commitment.

Depending on the Cloud Billing account's creation date and the active commitments in that account, this change applies in the following way:

• Cloud Billing accounts created on or after June 16, 2026: The CUD scope is Billing account (CUD sharing enabled) by default.
• Cloud Billing accounts created before June 16, 2026:
  • If the account has no active resource-based commitments on June 16, 2026, then the CUD scope has changed to Billing account (CUD sharing enabled).
  • If the account has any active resource-based commitments on June 16, 2026, then the CUD scope remains unchanged and Google Cloud continues to use your existing configuration.

For more information, see Share resource-based CUDs across projects.

Preview: Before you create Spot VMs, you can view the following information for a specific machine type and location:

• You can view the real-time obtainability and estimated uptime. This information helps you maximize your chances of successfully creating Spot VMs, as well as help ensure that your workload starts and runs efficiently.

• You can view historical and current preemption rate and pricing. This information helps you compare and choose the configuration that best fits your workload needs and budget.

For more information, see View the availability of Spot VMs and View the preemption rate and pricing for Spot VMs.

In an autoscaled managed instance group (MIG), you can configure the stabilization period to manage how quickly the autoscaler deletes instances after a decrease in the load. This configuration can help optimize costs or maintain extra capacity based on your workload requirements. For more information, see Configure stabilization period.

A vulnerability (CVE-2025-10263) about bypass of translation stages or GPT protections in some Arm core families was discovered and has been addressed. For more information, see the GCP-2026-036 security bulletin.

Generally available: The C4D machine series supports Hyperdisk Balanced High Availability disks.

For more information, see About Hyperdisk Balanced High Availability and Performance limits for machine series.

Generally available: You can gradually create Flex-start VMs in a managed instance group (MIG) as capacity becomes available. Unlike resize requests for MIGs that wait for full capacity before creating VMs, this method might create only a portion of your requested Flex-start VMs if capacity is unavailable. The MIG creates the remaining VMs later as capacity permits. Flex-start VMs run for up to seven days and help you obtain high-demand resources, such as GPUs, at a discounted price.

For more information, see Create a MIG that uses Flex-start VMs.

Generally available: Compute Engine supports the Google's custom-developed accelerator Tensor Processing Unit (TPU), providing a converged experience across AI accelerators on Google Cloud. You can use the Compute Engine instance API and managed instance group (MIG) API to create and manage TPU VMs. You can perform standard VM configurations such as using a custom OS or configure boot disk size. Compute Engine APIs support the creation and management of TPU slices across all consumption options, enabling small-scale experimentation and large-scale training and inference workloads.

For more information, see TPU resources in Compute Engine.

Generally available: In a managed instance group (MIG), obtain the requested number of virtual machine (VM) instances all at once by using bulk mode of the target size policy. Using bulk mode helps you avoid partial VM provisioning in a MIG. Bulk mode is particularly beneficial for batch workloads, such as high performance computing (HPC) or distributed training, that require full capacity before they can start. For more information, see About bulk mode.

Generally available: Two C4A bare metal machine types are generally available:

• c4a-standard-96-metal with 96 vCPUs and 384 GB of DDR5 memory
• c4a-highmem-96-metal with 96 vCPUs and 768 GB DDR5 memory

These two machine types support Hyperdisk Balanced, Hyperdisk Extreme, Hyperdisk Throughput, and Hyperdisk ML volume storage and up to 100 Gbps of network bandwidth.

To learn more about the C4A machine family, read General-purpose machines. To see where you can create C4A bare metal instances, read Bare metal instances.

A vulnerability in AMD firmware (CVE-2025-61971, CVE-2025-61972, CVE-2024-36315) that could compromise SEV-SNP guests has been addressed. For more information, see the GCP-2026-031 security bulletin.

A vulnerability (CVE-2025-54518) about potential corruption within the micro-operation (OP) cache in Zen 2 microarchitecture processors was discovered and has been addressed. For more information, see the GCP-2026-032 security bulletin.

Preview: In an autoscaled managed instance group (MIG), you can monitor individual autoscaling events and view details to understand the reasons behind each autoscaling decision. For more information, see Monitor autoscaling events.

Generally available: Compute Engine now offers support for AI zones. To learn more, see AI zones.

Generally available: The G4 accelerator-optimized machine series now supports the creation of virtual machine (VM) instances with less than one GPU attached (fractional GPUs). When you create VM instances with fractional GPUs, you can select 1/2, 1/4, or 1/8 of a G4 GPU. Fractional GPUs let you optimize costs for workloads that don't require the resources of a full GPU.

For more information, see the G4 machine series overview.

Generally available: You can use the Compute Engine remote Model Context Protocol (MCP) server to let AI agents and AI applications manage Compute Engine resources, such as Compute Engine instances, managed instance groups, disks, and snapshots. For more information, see Use the Compute Engine remote MCP server.

Generally available: You can create a Hyperdisk Balanced High Availability disk by cloning a zonal Hyperdisk Balanced or Hyperdisk Extreme disk. This lets you make your zonal workloads highly available by adding a replica of the data in another zone within the same region.

For more information, see Create a regional disk clone from a zonal disk.

Generally available: To ensure data consistency when backing up multiple disks, you can use consistency groups of instant snapshots to back up a group of disks at the same point in time.

For more information, see About instant snapshots.

Preview: You can specify a 120-second preemption notice duration while creating Spot VMs. Use this feature for workloads on Spot VMs where you want up to an additional 120 seconds for handling preemption. If you want to migrate existing Spot VMs workloads, make sure you update your workload to handle preemption outside of a shutdown script and test preemption. For more information, see Spot VMs and Create and use Spot VMs.

Generally available: You can rotate the customer-managed encryption key (CMEK) used to encrypt a disk, standard snapshot, or archive snapshot to a new key version without downtime.

Generally available: You can change the CMEK used to encrypt a disk, standard snapshot, or archive snapshot to a different key without downtime.

For more information, see Rotate the CMEK for a disk or standard snapshot and Change the CMEK for a disk or standard snapshot.

You can view the physical location of your Compute Engine instances in a zone to understand your cluster topology. This information helps you reduce network latency between your compute instances. For more information, see View Compute Engine instance topology.

Generally available: You can control the physical location of the Compute Engine instances in a MIG by using workload policies. Workload policies help you to, for example, place your compute instances close together to minimize network latency when running AI or ML workloads. For more information, see About workload policies in MIGs.

A vulnerability (CVE-2025-54510) about AMD SEV-SNP guest memory integrity has been addressed. For more information, see the GCP-2026-019 security bulletin.

A vulnerability affecting AMD SEV-SNP Confidential VM instances was discovered and has been addressed. For more information, see the GCP-2026-021 security bulletin.

Generally available: Hyperdisk ML disks are supported by the following machine series:

• C3 bare metal
• C4, including bare metal instances.
• C4A bare metal instances
• N4A

For more information, see Hyperdisk ML overview.

Generally available: Hyperdisk ML disks are supported by the following machine series:

• A3 Ultra
• C4D
• N4
• N4D

Hyperdisk ML offers the highest throughput of all Google Cloud Hyperdisk types, up to 2 TiB/s (2,097,152 MiB/s). For more information, see Hyperdisk ML overview.

Preview: To control the use of the deprecated container startup agent, an option for deploying containers on Compute Engine instances, you can enforce the constraints/compute.managed.disableVmsWithContainerStartupAgent organization policy constraint. This constraint prevents the creation of Compute Engine instances that use the container startup agent and the gce-container-declaration metadata.

You can also enforce this organization policy in dry-run mode to identify projects that use the deprecated metadata, without blocking resource creation.

For more information, see Prevent the creation of VMs that use the container metadata and Migrate containers deployed on VMs during VM creation.

Generally available: The maximum throughput for a Hyperdisk ML disk is increased to 2,097,152 MiB/s from 1,200,000 MiB/s. Hyperdisk ML provides the highest throughput per disk for machine learning and for workloads that require high read throughput on immutable datasets.

For more information, see About Hyperdisk ML.

A vulnerability (CVE-2026-23268) about CrackArmor was discovered and has been addressed. For more information, see the GCP-2026-015 security bulletin.

Generally available: The maximum throughput for a Hyperdisk Balanced High Availability disk is increased to 2,400 MiB/s from 1,200 MiB/s. Hyperdisk Balanced High Availability provides high availability block storage for mission-critical workloads by synchronously replicating data between two zones within a region.

For more information, see Hyperdisk Balanced High Availability overview.

Preview: The instance flexibility policy of a managed instance group (MIG) lets you override the minimum CPU platform and disk definition that is specified in the MIG's instance template. With these overrides, you can select machine types that run on different CPU platforms and that have different architectures.

For more information, see About instance flexibility in MIGs.

Changed: The following operations on the boot disk of a Compute Engine instance that has a service account attached require the iam.serviceAccounts.actAs permission on the service account. In the following list, the boot disk of such an instance is referred to as the source disk.

• Creating a standard or archive snapshot of the source disk, including application consistent snapshots
• Cloning the source disk
• Creating a machine image of the instance
• Creating a custom image of the source disk
• Starting asynchronous replication of the source disk to another region
• Creating a new disk when you create an instance, if the new disk is created from an instant snapshot of the source disk

If you already have the Compute Instance Admin (v1) (roles/compute.instanceAdmin.v1) role and the Service Account User (v1) (roles/iam.serviceAccountUser) role on the project, no action is required.

Otherwise, ask your administrator to grant you the iam.serviceAccounts.actAs permission on the service account. For instructions, see Manage access to other resources.

To address high-severity kernel vulnerabilities (including CVE-2025-21756 and CVE-2025-38052) in Rocky Linux 8 and 9, updates are available for the Compute Engine images maintained by CIQ. If your VM instances use images dated before September 2025 (version v20250912), you must take action to ensure you continue to receive security patches.

How to determine if your Compute Engine VMs are affected

You are affected if your VM instance uses a Rocky Linux image from an -optimized-gcp or -optimized-gcp-nvidia family with a version date older than v20250912 (for example, rocky-linux-9-optimized-gcp-v20250807). To check your VM's source image, see View VM instance image details. You can view details for these image families in Rocky Linux OS details.

Action required

• If your image version is v20250912 or later: Your VM is already configured to use the newer SIG/Cloud Next (SCN) repositories and is receiving security updates. No action is required.

• If your image version is older than v20250912: Your VM is configured to use legacy SIG/Cloud repositories that no longer receive regular kernel updates and won't receive future security patches. While running sudo dnf update applies a one-time patch for the vulnerabilities listed, you must manually migrate the VM to the SCN repositories to receive ongoing updates by following the CIQ migration guide.

When you autoscale a managed instance group (MIG), you can monitor the configured group size and the size recommended by the autoscaler on a chart. For more information, see Monitor group size.

For Red Hat Enterprise Linux (RHEL) operating system, VM Manager provides vulnerability scanning results based on the latest minor version for each major version released. If your VM runs an earlier minor version of RHEL, you might get inaccurate results in the vulnerability reports. For more information about supported operating systems for vulnerability reports, see supported operating systems.

Generally available: You can use managed constraints with Organization Policy Service for centralized, programmatic control of your Compute Engine resources. Managed constraints replace legacy compute.* constraints and are identifiable by the compute.managed.* prefix. They also include built-in support for safe rollout tools like Policy Simulator and dry run mode.

For more information, see Organization policies for Compute Engine and Managed constraints.

Generally available: You can apply compact placement policies to standalone Flex-start VMs. These policies let you colocate your standalone Flex-start VMs as close to each other as possible. Applying a compact placement policy minimizes network hops and improves the performance of latency-sensitive workloads. For more information, see About compact placement policies and About Flex-start VMs.

Generally available: H4D VMs, designed for high performance computing (HPC) workloads, are now generally available. Based on 5th generation AMD EPYC Turin with Cloud RDMA 200 Gbps networking, H4D VMs offer 192 cores (SMT disabled), up to 1,488 GB of memory, and 3,750 GiB of Local SSD. H4D is optimized for tightly-coupled applications that scale across multiple nodes.

For more information, see H4D machine series.

Generally available: You can use Hyperdisk Exapools for large-scale workloads, such as AI and machine learning, that require between 500 TiB and 5 EiB of block storage and more than 100 GiB/s of concurrent performance in a single zone. With Hyperdisk Exapools, you purchase storage and performance in bulk and share those resources across as many as 500,000 disks in a single project.

To use Hyperdisk Exapools with your projects, contact your account team to get access.

To learn more about Hyperdisk Exapools, see Hyperdisk Exapools overview.