TL;DR
Apple rebuilt Siri on a custom 1.2T-parameter Gemini model running on Nvidia Blackwell GPUs in Google Cloud. Federighi says requests are never stored. The company unveiled five new AI models and a three-tier privacy architecture.
Apple’s most important AI announcement at WWDC 2026 was not a feature. It was an architecture.
The rebuilt Siri runs on a custom 1.2-trillion-parameter model built on Google’s Gemini technology, hosted on Google Cloud servers powered by Nvidia Blackwell B200 GPUs. For the company that made privacy its premium product, outsourcing AI inference to its largest competitor’s cloud requires an extraordinary amount of trust engineering.
The three-tier system
Apple now routes Siri queries through three layers. Simple tasks stay on-device using Apple’s own models. Moderately complex requests go to Apple’s Private Cloud Compute servers.
The 💜 of EU tech
The latest rumblings from the EU tech scene, a story from our wise ol' founder Boris, and some questionable AI art. It's free, every week, in your inbox. Sign up now!
The heaviest reasoning tasks route to Google Cloud. At each tier, Apple says queries are anonymised and tokenised so neither Apple staff nor Google can link requests to individual users.
What Federighi said
“We use none of the models that Google deploys to its customers,” software chief Craig Federighi said at a WWDC media event. “Your requests are completely private to you. They’re never stored. They’re never accessible to anyone.”
The contract with Google reportedly bars the company from training future models on Apple user data. Nvidia’s confidential computing feature encrypts data while it is being processed on the Blackwell GPUs, adding a hardware-level safeguard on top of the contractual one. No independent audit of the Google Cloud tier has been published, and contractual bans on training can be renegotiated in future deals.
The five new models
Apple unveiled the third generation of its Apple Foundation Models (AFM), a family of five models distilled from Gemini: AFM Core, Core Advanced, Cloud, Cloud Pro, and Cloud Image. The most powerful, AFM Cloud Pro, offers quality that is “similar” to Google’s frontier Gemini models, according to AI VP Amar Subramanya, though no independent benchmark has confirmed the comparison.
All five are custom-built for Apple Silicon, trained with proprietary data and reinforcement learning. The on-device models handle basic tasks without any data leaving the phone.
Why this is awkward
A year ago, Federighi and marketing chief Greg Joswiak dismissed the idea of a “bolted-on chatbot” at WWDC 2025. Now Siri is a conversational chatbot. When asked what changed, Federighi said: “We see Siri not as a separate chatbot, but rather as an integral but conversational tool that you use in the moment.”
Apple also settled a $250 million class action last month over marketing AI features in 2024 that were not ready when the iPhone 16 launched. The company acknowledged through Siri engineering lead Mike Rockwell that previous attempts to revamp the assistant “didn’t meet Apple’s standards.”
The Google dependency question
The deal with Google is reportedly worth roughly $1 billion per year. It gives Apple access to frontier-class AI without building it from scratch, but it also creates a dependency on a company that is simultaneously Apple’s biggest rival in mobile operating systems and its largest source of search revenue.
For users, the question is whether Apple’s privacy architecture is strong enough to survive the combination of Google models, Nvidia hardware, and cloud inference. For investors, the question is whether Apple’s late entry into AI can recapture the ground it lost while trying to build everything in-house. WWDC 2026 is Apple’s answer to both. September, when the features ship, is when users get to decide if they believe it.
Get the TNW newsletter
Get the most important tech news in your inbox each week.