Private Cloud Compute: A new frontier for AI privacy in the cloud (security.apple.com)

Apple announced Private Cloud Compute (PCC), a new cloud architecture for Apple Intelligence that lets iPhones, iPads and Macs call larger foundation models while preserving device-grade privacy. PCC is designed so personal data sent to the cloud is processed only to fulfill the user’s request, deleted after inference, and — Apple says — never accessible to anyone else, including Apple staff. The move addresses a core tension in cloud AI: large models need unencrypted inputs to reason over complex data, but traditional cloud operations (logging, privileged admin access, opaque stacks) make strong, verifiable privacy guarantees difficult. Technically, PCC roots trust in custom Apple silicon servers with Secure Enclave and Secure Boot and a hardened subset OS derived from iOS/macOS. Apple removed typical datacenter admin surfaces (no remote shells, limited observability), built purpose-specific operational metrics, and implemented a Swift-on-Server ML stack for deterministic, stateless inference. Key design goals: enforceable guarantees (no persistent logs), no privileged runtime access, non-targetability (isolated per-request processing to prevent single-node user targeting), and verifiable transparency so researchers can confirm the production software matches inspected builds. Apple will publish a deeper technical follow-up and run a beta; if realized, PCC could be a generational shift for private cloud AI by bringing device-level cryptographic and runtime assurances into large-scale model inference.