Private Cloud Compute: A new frontier for AI privacy in the cloud

Overall Reception

• Many are impressed by Apple’s attempt to design a privacy architecture first, then ship AI features on top of it, viewing this as unusually customer‑centric.
• Others see it as Apple finally catching up to long‑known security practices, wrapped in marketing, though some argue PCC is genuinely ahead in combining ML accelerators, enclaves, and public verifiability at consumer scale.
• Several commenters say this is the first time they’ve considered moving more of their computing to Apple because of privacy.

Architecture & Security Properties

• Praised aspects: stateless nodes with ephemeral keys, no general logging, onion‑style routing via Oblivious HTTP, non‑targetability claims, and publishing binary images plus some source, including sepOS and iBoot.
• The transparency log and tooling, plus a dedicated bug bounty with higher rewards for privacy‑breaking issues, are seen as significant.

Remote Attestation & Verifiability

• Discussion explains attestation as signed measurements from a hardware root of trust that clients verify against known hashes.
• Some highlight this as “end‑to‑end attestation” for an ML enclave, stronger than many cloud offerings.
• Skeptics argue that because Apple holds the signing keys and controls both hardware and client software, it could—under pressure—sign malicious builds or fake attestations, and that this is inherently a “trust the vendor” situation.

Government Access & Threat Models

• Debate over how resistant PCC really is to state actors and secret court orders.
• Some think architecture plus transparency logs make covert, targeted backdoors very hard without broad, visible changes.
• Others say US legal powers and gag orders mean no cloud in the US can be fully private; at best, PCC stops routine corporate misuse, not intelligence services.

On‑Device vs Cloud, Opt‑Out

• Multiple commenters want clear on‑device‑only modes, a global “PCC kill switch,” and MDM‑level controls.
• It’s unclear from the discussion exactly how visible or granular opt‑out will be, especially for integrated features like Spotlight or notification summarization.

Comparisons & Broader Implications

• Compared frequently to AWS Nitro Enclaves, Google confidential computing, and TPM/TEE‑based systems; PCC is seen as pushing similar ideas to a mainstream consumer scale.
• Some want PCC‑like infrastructure or APIs opened to third‑party developers or even self‑hosted nodes.
• There is speculation about costs, business models, and whether this is primarily a privacy product or an AI/investor story.