Siri AI: Gemini as a teacher, not as an engine - what the WWDC didn't say

The Essential

The press has taken note of Apple's admission of dependence on Google, but the agreement is more nuanced: Google would provide Gemini as a « teacher » through distillation during training, while inference would take place in the cloud. This technical division nuances the media narrative and illustrates the complexity of the partnership between Apple and Google.

Create a watch on this topic Receive the next developments if any, with an AI watch configured from this article.

Create my free watch

The dominant reading of the June 8 WWDC has taken hold in a few hours: by aligning « Siri AI » with Google's Gemini models, Apple would have signed the admission of its lag in artificial intelligence. This framing, adopted by most French and American press, is not entirely false. But it misses the real mechanics of the agreement - and this mechanics deserves to be unpacked, as it determines what « dependence » means here.

What Apple announced

« Siri AI » is a complete overhaul of the assistant: multi-turn conversations, taking into account the context displayed on the screen, execution of actions across applications. It relies on the new generation of Apple Intelligence, and Apple confirmed during the keynote that it also uses Google's Gemini models in addition to its own Apple Foundation Models (AFM). Google subsequently published a blog post for Apple ecosystem developers, "Bringing the latest Gemini models to Apple developers," officially announcing the platform aspect of the agreement.

Professor is not the engine

This is where the reading "Apple runs under Gemini" deserves correction. According to technical details, notably from AppleInsider, the third-generation AFMs do not include any Gemini code at runtime. Gemini operates at two distinct stages: as a "teacher" model during the training of AFMs - a distillation process, where the large model is used to generate data and signals for the smaller one - and as a cloud model, separately called upon for queries exceeding the device's on-device capabilities.

The distinction is not a minor detail. A dependency during training is occasional and reversible: switching teachers for the next generation is a purchasing decision. A dependency during inference, however, involves privacy, latency, and costs for every user query. Apple has managed the first on the device, and accepted the second in the cloud - two fundamentally different choices, which the word "dependency" obscures.

The infrastructure, the other stage of the agreement

According to The Information, part of the cloud inference would run on Nvidia Blackwell B200 chips hosted by Google - information that Apple has not confirmed. If it comes to fruition, 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 queries on Nvidia hardware in Google data centers, even under contract, moves the boundary of that control.

Europe will wait

Last stage, the regulatory one: Siri AI will not be available in the European Union at launch on iPhone and iPad (iOS 27 and iPadOS 27), whereas it will arrive on macOS 27 and visionOS 27. Apple invokes the DMA, arguing that regulators' interpretation would require opening up nearly unlimited access to third-party AI systems on the device, without sufficient safeguards. "We are deeply disappointed that our European users will not have Siri AI on iPhone or iPad when we release our new software versions," said Craig Federighi, without providing any timeline for the EU.

Overall, the Apple-Google agreement outlines less a surrender than a hierarchy of concessions: Apple maintains control over the on-device model, outsources the cutting-edge training, and gives ground on cloud inference. It is on this last point — not on distillation — that the future will be decided.