Google Launches 'Private AI Compute' — Secure AI Processing with On-Device-Level Privacy

Google on Tuesday unveiled Private AI Compute, a privacy‑preserving technology to process AI queries in a secure cloud environment, described as a secure, fortified space akin to on‑device processing but with cloud capabilities. According to Google, Private AI Compute aims to unlock the full speed of Gemini cloud models while keeping personal data private and inaccessible to anyone else.

The system is powered by Trillium TPUs and Titanium Intelligence Enclaves, and relies on an AMD‑based hardware TEE to encrypt and isolate memory, with attested workloads and no administrative access to the workloads. It uses a multi‑layer attestation flow, including Noise encryption, Oak attested sessions, and ALTS within a scalable inference pipeline, and is designed to be ephemeral so inputs and inferences are discarded when a user session ends.

NCC Group, which reviewed Private AI Compute between April and September 2025, found a timing‑based side channel in the IP‑blinding relay component under certain conditions, though Google called the risk low due to noise from multi‑user use; the firm also identified three attestation‑mechanism issues that could allow a DoS or protocol attacks, for which Google is implementing mitigations.