Apple announced that its new Siri AI system uses Google's Gemini language models and Nvidia hardware installed in Google servers. The company emphasized that it continues to uphold its privacy standards, similar to when its AI models were running locally on devices or on Apple-controlled servers. Apple has long positioned user privacy as a core benefit of its platforms, using encryption to prevent unauthorized access to data by third parties, including its own employees.
Apple’s Craig Federighi explained the architecture of Apple Intelligence, highlighting that the system still relies on an on-device model for simpler queries. Newer devices with sufficient hardware use AFM 3 Core Advanced, which combines device storage and processing power to enhance features like dictation and Siri’s voice. For more complex tasks, devices will connect to cloud-based models co-developed by Apple and Google, including AFM 3 Cloud, ADM 3 Cloud, and AFM 3 Cloud Pro, which runs on Google-owned Nvidia hardware. Apple claims that these models still maintain privacy protections through its new iteration of Private Cloud Compute, designed to operate on third-party hardware.
Apple’s privacy measures include using Nvidia’s Confidential Computing, Intel’s Trust Domain Extensions, and Google’s Titan security chip to mirror the protections it provides for its own servers. The company also maintains a cryptographically verifiable ledger of all Google Cloud hardware used in the Private Cloud Compute fleet, ensuring only Apple-signed software runs on these servers. Apple said it will gradually increase protections on Google Cloud servers to match its own standards over the summer.
Source: arstechnica
