Apple unveiled Siri AI at WWDC 2026 on June 8, per its newsroom, the long-delayed conversational overhaul of its voice assistant. The disclosure underneath the consumer reveal is that the cloud-tier model behind it, AFM Cloud Pro, was built in collaboration with Google and runs on Nvidia GPUs inside Google Cloud, per CNBC. A new LanguageModel protocol lets iOS apps swap between Apple Foundation Models, Gemini, and Anthropic's Claude with no session-code changes.
The technical detail to read carefully is the architecture, not the Siri demo. Apple split its model stack into two tiers: an on-device AFM for privacy-bound tasks, and AFM Cloud Pro for frontier-quality workloads. The cloud-tier model is Google's Gemini Frontier, repackaged inside Apple's Foundation Models API, running on Nvidia accelerators in Google's data centers, per TechCrunch's WWDC coverage. Apple did not phrase it that way onstage; the documentation said it for them. The practical effect is that Apple stopped trying to win the frontier-model race itself and shipped the best Siri it could ship by buying the model layer from Google.
The developer-facing piece reinforces the same posture. The LanguageModel protocol means an app developer writes once, and the call routes to whichever model the user has selected as default (Apple, Gemini, Claude, with more partners coming), per TechTimes. Apple is positioning the device as the router and the user as the choice, with Apple capturing a platform position on every routed inference call. That is a different business than selling frontier models, and a more defensible one given Apple's relative position on training compute and AI research talent.
Bottom Line
WWDC was Apple admitting the frontier race is decided and reaching for the platform layer above it. For app developers shipping AI features, the LanguageModel protocol is the API to build against now. For Google, the cloud-AI revenue line just gained the iPhone install base as a routed customer.