Premium

Google unveils ‘Private AI Compute’ for Gemini models with on-device-level privacy

Private AI Compute represents a major milestone in responsible AI processing, reinforcing user security and privacy in the cloud.

Private AI Compute bridges the gap between on-device performance and cloud intelligence. It powers new features such as Magic Cue on Pixel 10 devices, which delivers timely suggestions, and enhancements to the Recorder app, enabling multilingual summarisation of transcripts. (Image: Google)

Google has been advancing privacy-enhancing technologies (PETs) to optimise a wide range of AI applications. Its latest development, Private AI Compute, introduces a new cloud-based AI processing platform integrated with Google’s Gemini models. According to Google, this system delivers the same level of security and privacy typically associated with on-device processing.

This innovation underscores Google’s ongoing commitment to safety and responsibility in AI development. As AI evolves, it is becoming more capable — moving beyond simple tasks to proactively anticipating user needs, offering personalised suggestions, and performing actions at the right moments. Achieving this sophistication requires advanced reasoning and computational power that often exceed on-device limitations.

To address this, Private AI Compute was developed to unlock the full potential of Gemini’s cloud models while maintaining strict privacy standards. The platform enables faster, more intelligent interactions without exposing users’ personal data — even to Google itself. This ensures that users can enjoy seamless access to information, smarter recommendations, and more efficient task execution with enhanced privacy protection.

Story continues below this ad

Data protection and security

Private AI Compute represents a major milestone in responsible AI processing, reinforcing user security and privacy in the cloud. Built on the foundations of Google’s AI Framework, AI Principles, and Privacy Principles, it provides a secure and isolated environment for handling sensitive data — the kind traditionally processed only on personal devices.

Also read | Nano Banana 2: Google’s next AI model promises smarter image understanding

At its core, the system relies on Google’s advanced technology stack, including custom Tensor Processing Units (TPUs) and Titanium Intelligence Enclaves (TIE), which deliver robust privacy and security guarantees. This architecture enables Gemini models to process data efficiently in the cloud while upholding stringent privacy requirements.

To ensure that user data remains protected, Private AI Compute employs hardware-level safeguards such as remote attestation and end-to-end encryption. These measures create a sealed, hardware-secured environment where sensitive information is processed privately — ensuring that even Google cannot access it. Ultimately, users retain full control over their data, reinforcing transparency and trust in AI-powered experiences.

Enhancing on-device capabilities

Story continues below this ad

This development marks the beginning of a broader expansion in AI experiences that blend the strengths of on-device and cloud-based processing for sensitive applications. Google plans to share further updates, including a technical brief detailing the privacy and security innovations behind Private AI Compute.