Google Cloud has made its Confidential G4 GPU instances generally available, expanding its infrastructure options for organizations that need to run artificial intelligence workloads while protecting sensitive information during processing.
The service combines Google Cloud G4 instances with NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs and confidential-computing controls. Google says the design is intended to help protect model weights, prompts, generated responses and other data while they are actively being used, rather than only when stored or transmitted.
Availability expands beyond preview
The October 21 launch moves Confidential G4 from preview to general availability across 36 zones. Customers can use the capability with Compute Engine virtual machines and Google Kubernetes Engine nodes, with multiple consumption models available.
Support for Confidential Space, Google Cloud’s environment for protected multi-party data collaboration, is expected to reach general availability for these systems in the coming weeks. That addition could be useful where several organizations need to analyze or train on combined datasets without exposing the underlying private information to one another.
How the protection works
Confidential G4 uses hardware-enforced isolation, non-exportable hardware keys, memory encryption and remote attestation. In practical terms, remote attestation lets an organization verify the software, configuration and integrity of the execution environment before it releases a sensitive dataset or starts a workload.
Google says the environment is designed to restrict access from layers outside the protected workload, including host administrators and the hypervisor. This matters for companies evaluating cloud-based inference for regulated records, proprietary models or customer conversations, although each organization still needs to assess its own threat model, identity controls and compliance obligations.
The company also says teams can deploy with standard Ubuntu images, NVIDIA drivers and CUDA-X libraries without refactoring their applications specifically for the confidential environment. That could reduce migration work for existing single-GPU inference pipelines.
Performance claims need workload-level testing
In Google’s testing on the g4-standard-48 machine shape, Confidential G4 delivered inference throughput within 3.5% of standard, non-confidential G4 instances across selected workloads using Llama 3, Llama 3.3, ResNet-50 and SDXL. Those results are vendor benchmarks under specific configurations, so actual latency and throughput can vary with model, software stack, hardware selection and traffic pattern.
For infrastructure teams, the general-availability milestone makes confidential GPU computing a production option rather than an experimental preview. A sensible evaluation should compare total cost, regional capacity, attestation integration, operational monitoring and real application performance against a standard G4 deployment.
It should also test failure recovery and key-release policies, not just headline throughput. Confidential computing strengthens the boundary around data in use, but it does not replace secure application design, least-privilege access or careful handling of data before and after inference.
Source: Google Cloud’s official announcement, published October 21, 2026.
Featured image: original AI-generated editorial illustration created for BEX News; it is a conceptual depiction, not a photograph of Google Cloud hardware.
