Google Cloud announced general availability of Confidential Computing on its G4 GPU instances on October 21, 2026. The company says the service covers 36 zones; its supported-zone list names 15 in the United States.

The announcement covers Compute Engine VMs and Google Kubernetes Engine (GKE) nodes powered by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs.

The 15 U.S. zones in Google’s list

Google’s supported-configuration page lists these U.S. zones for G4 Confidential VM:

us-central1-d, us-central1-f, us-east1-b, us-east1-d, us-east4-b, us-east4-c, us-east5-a, us-east5-b, us-east5-c, us-south1-a, us-south1-b, us-west1-b, us-west1-c, us-west3-a, and us-west4-a.

What the confidential configuration protects

Confidential computing here is Google’s approach to protecting data and model parameters while they are in use. Google describes hardware memory encryption, non-exportable keys and remote attestation as parts of that protection. Its supported configuration pairs the g4-standard-48 machine type with AMD SEV and NVIDIA Confidential Computing.

Deployment requires a new VM and GPU setup

G4 supports standard on-demand, Spot and Flex-start provisioning; reservations are also supported. Before creating a VM, teams need a supported zone and operating-system image, along with enough GPU quota. Google supports Container-Optimized OS cos-125-lts and newer, and Ubuntu ubuntu-2204-lts and newer.

An existing VM cannot be converted to Confidential VM, so deployment requires creating a new instance. After creation, confidential-computing mode must be enabled on the attached GPU. Google’s GPU VM guide describes driver and Linux configuration steps and recommends version 580 or later of the NVIDIA open kernel module drivers. G4 use also automatically incurs a license fee for NVIDIA Confidential Computing software and services.

Google’s documentation says GPU-accelerated Confidential VM instances cannot create clusters for multi-node workloads. It also lists a possible GPU-attestation failure involving a mismatched measurement record at index 9. The documented workaround is to fully stop and restart the VM; rebooting from the guest operating system does not resolve it.

Google and Vertebrae report different performance measures

Google reports less than 3.5% throughput degradation in tests using g4-standard-48 with Llama 3, Llama 3.3, Resnet-50 and SDXL. The benchmark measured tokens or images per second, and Google says results vary with system setup, hardware and software.

Vertebrae separately reports a 5.2% increase in end-to-end latency on its workloads when comparing Confidential G4 with standard G4 running identical software. Throughput and latency are different measures, and the two results come from different workloads.

Google described a later Confidential Space rollout

In its October 21, 2026 announcement, Google said Confidential Space support for secure multi-party data sharing was expected to roll out to general availability in the coming weeks.