Particle.news
Download on the App Store

Google Launches Private AI Compute to Bring Gemini Power to the Cloud With On‑Device‑Level Privacy

The system moves sensitive tasks into hardware‑secured cloud enclaves to offer on‑device‑level privacy, with early audits citing fixable weaknesses.

Overview

  • Google says data processed in Private AI Compute is accessible only to the user and not to Google, using a sealed environment linked via encryption and remote attestation.
  • Initial rollout powers Pixel 10 features, making Magic Cue suggestions more timely and enabling Recorder to summarize transcriptions in more languages as part of the November Feature Drop.
  • The platform runs on Google’s custom TPUs with Titanium Intelligence Enclaves, isolating workloads from administrators and broader infrastructure for verifiable processing boundaries.
  • Google describes the architecture as ephemeral by design, discarding inputs and model computations after each session to limit retention and exposure.
  • An external NCC Group assessment reported a timing-based side channel in an IP‑blinding relay and attestation implementation issues that could enable denial‑of‑service, with Google working on mitigations; coverage notes parallels to Apple’s Private Cloud Compute.