New Apple Intelligence features rely on external cloud providers while offering developers better on-device AI capabilities through Core AI library.
Summary
- Apple announced new Siri AI features at WWDC 2026 that use vision LLMs to extract screen information, avoiding need for app-specific integrations
- Most demanding AI tasks will run on Google Cloud using NVIDIA hardware, not Apple’s own infrastructure
- New Core AI library integrates with PyTorch ecosystem to help developers leverage Apple hardware for AI models
- Features require iOS 27 Developer Beta and waitlist approval, with limited real-world testing data available
Scepticism Warranted After 2024 Promises
Apple’s track record with AI announcements suggests caution. The company’s 2024 WWDC Apple Intelligence promises left early adopters disappointed, creating reasonable doubt about the latest batch of AI features announced at WWDC 2026. The new Siri AI capabilities do appear technically feasible with current technology, particularly given Apple’s licensing arrangement for a custom Gemini-derived model.
Vision LLMs Enable Screen-Based Intelligence
The standout technical approach involves using vision LLMs to extract information directly from user screens. This sidesteps the complex integration challenges that would require every application to ship custom code for Apple Intelligence compatibility. Vision LLMs have matured significantly since June 2024, making this approach more viable than previous attempts.
Google Cloud Dependency for Complex Tasks
Despite Apple’s privacy-focused messaging, the most demanding AI tasks will run on Google Cloud infrastructure using NVIDIA GPUs rather than Apple’s own hardware. Apple’s Security Research blog explains that “for the most demanding tasks, including agentic tool-use and complex reasoning, we worked with Google and NVIDIA to extend our PCC infrastructure to Google Cloud systems using NVIDIA GPUs, while maintaining Apple’s powerful security and privacy protections.”
Security Architecture Spans Multiple Providers
Apple maintains its Private Cloud Compute security model across the Google Cloud deployment. The architecture includes dedicated processes for network data parsing, short-lived shared inference software, and isolated confidential VMs for attested keys. Apple promises all binaries will be published for public inspection, extending their transparency approach to the third-party infrastructure.
Core AI Library Opens Developer Access
The new Core AI library represents a significant step forward for developers wanting to run AI models on Apple hardware. It integrates with Meta’s PyTorch ecosystem through Core AI PyTorch Extensions, allowing developers to convert existing PyTorch models into Core AI programs. The system works by traversing PyTorch’s FX graph node-by-node and mapping operators to Core AI operations.
Limited Testing Access Constrains Evaluation
While iOS 27 Developer Beta is available today, accessing the new Siri AI features requires navigating a waitlist system. MacRumors’ Aaron Perris reports successfully gaining access, suggesting credible real-world testing reports may emerge soon. Until broader testing occurs, the actual performance and reliability of these features remains unknown.
Why it matters
This development creates new security considerations for enterprise environments where Apple devices process sensitive information through third-party AI infrastructure, while potentially offering better on-device AI capabilities for applications that don’t require cloud processing.
What to do now
- Monitor real-world testing reports as they emerge from early beta users
- Review Apple’s published binaries when available for security assessment
- Evaluate the Core AI library for potential internal AI model deployment if your organisation uses PyTorch
