Apple's WWDC 2026 keynote made official what months of reporting had sketched: Siri is rebuilt around Google's Gemini, under a multi-year deal reported at roughly $1 billion a year. The new Siri answers world-knowledge questions conversationally, holds context across apps, gets deeper visual intelligence through the camera, and arrives both as a system layer and a standalone app. TechCrunch's wrap leads with the obvious headline, Siri powered by Gemini, and on the same day AppleInsider ran sources insisting there is 'not a drop of Gemini' in what ships. Both are accurate, and the gap between those two sentences is the actual story of the keynote.
The reconciliation is the architecture. Apple is not renting Gemini as a service, it is renting Gemini as a teacher: Google loans frontier models, Apple distills them into Apple Foundation Models, rebuilt and optimized for Apple silicon, trained into weights that are pure Apple code by the time they ship on the device. Requests too heavy for the phone route to AFM Cloud, Apple's models running on Private Cloud Compute, the stateless compute layer Apple introduced two years ago, where the company says data is used only to execute your request, never stored, never used for training, with system images published for external verification. The privacy story survives the Google deal intact, because the thing users touch is never Google's model, it is Apple's distillation of it.
Around the Siri core, the sweep was wide: systemwide AI dictation replacing the old dictation stack, Shortcuts that build themselves from natural language, Siri context that follows you across apps and into calls, Photos gaining Reframe, Extend, and Cleanup edits, AI tab management in Safari, new parental controls, and menopause tracking in Health. iOS 27 runs on iPhone 11 and newer, a six-and-a-half-year-old phone kept on the upgrade train, with photo loading 70 percent faster and AirDrop 80 percent faster. Developers also found foldState and angleDegrees APIs in the betas, the clearest signal yet that September's hardware event has a hinge in it.
The strategic read: Apple just demonstrated the third way to live with frontier labs. One pattern rents the capability and wears the dependency openly. Apple's pattern rents the teacher and owns the deployment, paying about a billion a year for distillation rights while keeping the weights, the silicon, and the privacy narrative. It concedes the frontier, Apple is no longer pretending to train its own competitor to Gemini, but it converts that concession into a supply contract instead of a platform dependency. Whether a distilled student stays close enough to the teacher as Gemini moves is now one of the most commercially consequential open questions in AI, and the first data arrives when iOS 27 ships this fall.