Apple deploys Google-trained models in iOS 27 Siri via Private Cloud Compute on Nvidia GPUs

Apple deploys Google-trained models in iOS 27 Siri via Private Cloud Compute on Nvidia GPUs

Apple's senior vice president of software engineering Craig Federighi disclosed the technical architecture behind iOS 27's Siri AI in a post-WWDC tech talk on June 8, 2026, revealing that Apple's third-generation foundation models were trained using outputs from Google's Gemini frontier models.

Federighi clarified that Apple uses "none" of Google's client code, customer-facing models, deployment infrastructure, or knowledge bases like Google Search. Instead, Apple built a family of proprietary models refined using Gemini training data.

The AFM model family

According to Amar Subramanya, Apple's vice president of AI, the third-generation Apple Foundation Models (AFM) include:

• AFM Core: On-device dense architecture model, next generation of current shipping models
• AFM Core Advanced: On-device sparse architecture model with native multimodal capabilities, enabling features like invitation detection and expressive voices
• AFM Cloud: Server-side workhorse optimized for latency and serving cost
• AFM Cloud Image: Image generation and editing model supporting spatial reframing
• AFM Cloud Pro: Most capable model for agentic tool use and complex reasoning, with "quality similar to Gemini frontier models"

All models except AFM Cloud Pro are custom-built for Apple Silicon, trained with proprietary data, and refined using outputs from Gemini frontier models. Pricing was not disclosed.

Nvidia GPU deployment

For AFM Cloud Pro specifically, Apple collaborated with Google and Nvidia to extend Private Cloud Compute infrastructure to Nvidia GPUs in Google's cloud. Federighi emphasized that Apple's privacy architecture remains intact: requests are never stored, never accessible to Apple or third parties, and third-party researchers can continuously verify the privacy guarantees.

The System Orchestrator determines whether to process requests on-device or route them to Private Cloud Compute models based on complexity. Apple's World Knowledge Service provides grounding for current events and world knowledge queries, separate from Google Search.

Architecture details

The iOS 27 Siri system includes:

• System Orchestrator: Coordinates privacy-preserving request routing
• App Toolbox: Provides access to app actions
• Spotlight Semantic Index: Accesses personal content
• On-screen context understanding: Processes visual environment
• Private Cloud Compute: Extends iPhone privacy guarantees to cloud processing

Subramanya stated the goal is "to match every user request to the model which provides the best response at the lowest latency" across the model family.

What this means

Apple's approach represents a hybrid strategy: leveraging Google's frontier model capabilities for training while maintaining architectural independence and privacy controls. The Nvidia GPU deployment for the most demanding tasks indicates Apple's willingness to use external infrastructure when its own silicon cannot match performance requirements, provided privacy guarantees hold. The sparse architecture in AFM Core Advanced suggests Apple is adopting mixture-of-experts techniques for on-device efficiency, a departure from previous dense model designs.