YOU GUYS - get ready because Apple just dropped one of the most interesting privacy engineering deep-dives I've ever seen and it completely changes how we think about "trusted" cloud computing! You remember when Apple marketed itself as the only tech company that cared about your data — well, they're trying to hold onto that brand even while partnering with Google for their new Gemini-based Siri AI. The dilemma was real: powerful enough language models can't run locally on iPhones without sacrificing capability, but sending user queries off-device breaks Apple's core promise. So Craig Federighi took a side stage at WWDC and laid out the whole architecture in detail — it's brilliant stuff that you need to hear about before dismissing "privacy-washed" marketing as just another slogan this time around.
Here’s how they do it, because the engineering is genuinely clever: on-device processing handles simpler queries with AFM 3 Core (locally) and advanced reasoning with Cloud models; when a query must go off-device via Private Cloud Compute (PCC), Apple doesn't just hand Google your data — not at all. For their model running on Nvidia GPUs in Google clouds, they deploy a fresh iteration of PCC using Intel SGX Trust Domains, Nvidia Confidential Computing hardware enclaves, and even Google's own Titan security chip to create isolated environments where neither Google nor any other third party can peek inside. To top it off, Apple maintains its own cryptographically verifiable ledger of every single server in the fleet and only runs software signed by Apple on those boxes — so Google has no access to what's happening at all, even though their hardware is hosting it!
The brilliance extends further: an on-device System Orchestrator strips down every request before it leaves your device - if you ask about a shared recipe in Messages, only the relevant text goes up; not who sent it, not when, and definitely no metadata. And here's the kicker that Federighi wanted to emphasize: PCC is architecturally transient by design — data exists long enough for the model to generate an answer and then literally vaporizes immediately afterward because it's never stored in a persistent form on any server. The whole stack launches with iOS 27, iPadOS 27, macOS 27 Golden Gate this fall — developer betas are already out but wait for the stable July build if you want to try it yourself. This is Apple showing how "privacy-first" branding can actually be backed by real hardware isolation instead of just marketing copy.
Source: https://arstechnica.com/apple/2026/06/apple-says-its-ai-is-still-private-even-when-it's-running-on-googles-servers
Here’s how they do it, because the engineering is genuinely clever: on-device processing handles simpler queries with AFM 3 Core (locally) and advanced reasoning with Cloud models; when a query must go off-device via Private Cloud Compute (PCC), Apple doesn't just hand Google your data — not at all. For their model running on Nvidia GPUs in Google clouds, they deploy a fresh iteration of PCC using Intel SGX Trust Domains, Nvidia Confidential Computing hardware enclaves, and even Google's own Titan security chip to create isolated environments where neither Google nor any other third party can peek inside. To top it off, Apple maintains its own cryptographically verifiable ledger of every single server in the fleet and only runs software signed by Apple on those boxes — so Google has no access to what's happening at all, even though their hardware is hosting it!
The brilliance extends further: an on-device System Orchestrator strips down every request before it leaves your device - if you ask about a shared recipe in Messages, only the relevant text goes up; not who sent it, not when, and definitely no metadata. And here's the kicker that Federighi wanted to emphasize: PCC is architecturally transient by design — data exists long enough for the model to generate an answer and then literally vaporizes immediately afterward because it's never stored in a persistent form on any server. The whole stack launches with iOS 27, iPadOS 27, macOS 27 Golden Gate this fall — developer betas are already out but wait for the stable July build if you want to try it yourself. This is Apple showing how "privacy-first" branding can actually be backed by real hardware isolation instead of just marketing copy.
Source: https://arstechnica.com/apple/2026/06/apple-says-its-ai-is-still-private-even-when-it's-running-on-googles-servers