The dominant reading of the June 8 WWDC settled in within hours: by tying “Siri AI” to Google’s Gemini models, Apple would have effectively admitted that it was lagging in artificial intelligence. That framing, repeated by much of the French and American press alike, is not wrong. But it misses the real mechanics of the deal — and those mechanics deserve to be unpacked, because they determine what “dependency” actually means here.
What Apple announced
“Siri AI” is a complete overhaul of the assistant: multi-turn conversation, awareness of the context displayed on screen, and the ability to carry out actions across apps. It is built on the new generation of Apple Intelligence and, as Apple confirmed during the keynote, on Google’s Gemini models in addition to its own Apple Foundation Models (AFM). Google then published a post for developers in Apple’s ecosystem, “Bringing the latest Gemini models to Apple developers,” formalizing the platform side of the agreement.
Teacher is not engine
This is where the reading that “Apple runs on Gemini” needs correcting. According to the technical details highlighted notably by AppleInsider, third-generation AFM includes no Gemini code at runtime. Gemini comes into play at two distinct stages: as a “teacher” model during AFM training — through distillation, where the larger model is used to generate training data and signals for the smaller one — and as a cloud model, called separately for requests that exceed on-device capabilities.
The distinction is not a minor point. A training dependency is temporary and reversible: switching teachers for the next generation is a procurement decision. An inference dependency, by contrast, affects privacy, latency, and costs on every request from every user. Apple preserved the first on-device, and accepted the second in the cloud — two choices of a very different nature, which the word “dependency” tends to flatten.
The infrastructure, the other layer of the deal
According to The Information, part of the cloud inference would run on Nvidia Blackwell B200 chips hosted by Google — information Apple has not confirmed. If verified, it would mark a significant shift: Apple built its Private Cloud Compute on in-house silicon precisely to keep the inference chain under control. Running Siri requests on Nvidia hardware in Google data centers, even under contract, moves the boundary of that control.
Europe will have to wait
The last layer is regulatory: Siri AI will not be available in the European Union at launch on iPhone and iPad (iOS 27 and iPadOS 27), although it will arrive on macOS 27 and visionOS 27. Apple cites the DMA, arguing that regulators’ interpretation would require third-party AI systems to be given near-unlimited access to the device, without sufficient safeguards. “We are deeply disappointed that our European users will not have Siri AI on iPhone or iPad when our new software releases ship,” said Craig Federighi, without offering any timeline for the EU.
Overall, the Apple-Google deal looks less like a surrender than a hierarchy of concessions: Apple keeps control of the on-device model, outsources frontier training, and gives ground on cloud inference. It is on that last point — not on distillation — that what comes next will be decided.