Apple Details Private Cloud Compute Architecture for Private AI in the Cloud

Summary

Apple describes Private Cloud Compute (PCC), a cloud intelligence system built to handle Apple Intelligence requests that require larger foundation models than devices can run locally. PCC uses custom Apple silicon, a hardened operating system derived from iOS and macOS, Secure Enclave, Secure Boot, code signing, sandboxing, and a purpose-built machine-learning stack. Apple says requests are encrypted from the user device directly to validated PCC nodes, while load balancers and other supporting services remain outside the decryption trust boundary. The system is designed to use personal data only to fulfill an inference request, delete it after returning the response, and prevent access even by Apple staff with administrative responsibilities. PCC excludes remote shells, interactive debugging, general-purpose logging, and mechanisms that could expand privileged access at runtime; only predefined and audited metrics and logs can leave a node. To reduce targeted attacks, Apple describes hardware inspection and revalidation, third-party-observed supply-chain checks, blind-signature credentials, an external OHTTP relay to hide source IP addresses, and target diffusion that distributes requests across subsets of nodes without user-identifying metadata. Apple also plans to publish measurements and binary images for every production build in an append-only transparency log, allowing researchers to verify that production nodes run inspected software. The company says it will provide research tools, a virtualized PCC environment, selected security-critical source code including sepOS and iBoot, and expanded Apple Security Bounty coverage. This is an initial technical overview; Apple says further implementation details will follow when PCC becomes available in beta.