Google Cloud Adds NVIDIA Blackwell GPUs to Confidential Computing Lineup

Google is betting that hardware-level encryption will set its cloud apart in the AI race. On Monday, the company announced a major upgrade to its Confidential Computing lineup on Google Cloud — new virtual machines powered by NVIDIA Blackwell GPUs, an open-source SDK for encrypting AI prompts, and beefed-up attestation features for multi-party data collaboration.

The centerpiece is the new Confidential G4 VMs, currently in preview. They pair NVIDIA’s RTX PRO 6000 Blackwell Server Edition GPUs with fifth-generation AMD EPYC Turin processors. AMD’s Secure Encrypted Virtualization (SEV) technology handles the hardware-level isolation, keeping data inside a trusted execution environment (TEE) during AI workloads.

For enterprises running sensitive models like healthcare diagnostics, financial fraud detection, or proprietary LLM training, that means the GPU memory itself is encrypted, not just the data in transit.

Alongside the hardware push, Google open-sourced a Prompt Encryption SDK. The tool lets developers encrypt both the prompts they send to AI models and the generated output end-to-end, so everything stays encrypted from the client all the way to the inference server inside the confidential computing environment. It’s a practical answer to a problem that’s been nagging enterprise AI adopters: how to use third-party cloud GPUs without exposing proprietary data or user privacy.

Google also upgraded Confidential Space, its platform for secure multi-party data collaboration. The new version integrates Intel Trust Authority, an attestation service that verifies the runtime environment hasn’t been tampered with and confirms it meets the customer’s security policies. Trust Authority isn’t new — Intel has been pushing it for confidential computing since 2023. But this is the first time it’s been baked directly into Confidential Space as a verifiable trust anchor.

Separately, Google added NVIDIA Hopper-architecture GPU support to Confidential Space, opening up federated learning scenarios where different organizations can jointly train models without exposing their private datasets to each other. Combined with the Blackwell-powered VMs, it gives Google Cloud a fairly complete story for what the industry calls “confidential AI” — data encrypted at rest, in transit, and during computation.