Apple's WWDC 2026 Siri AI runs Gemini on Google Cloud with NVIDIA GPUs

Apple's WWDC 2026 Siri AI announcements are met with skepticism in the post, which cites how badly Apple's 2024 Apple Intelligence promises failed to materialize. The new features are described as at least looking feasible with current technology, partly because Apple is licensing a custom Gemini-derived model it can run on its own Private Cloud Compute infrastructure. Vision LLMs are being used to read information directly from the user's screen — a practical workaround that avoids requiring every existing app to ship custom integration code, and one the post notes was less viable in June 2024 when vision LLMs were a less mature category.

On the developer tooling side, Apple introduced a Core AI library designed to let developers run their own models on Apple hardware.

On the developer tooling side, Apple introduced a Core AI library designed to let developers run their own models on Apple hardware. It integrates with Meta's open-source PyTorch ecosystem through `coreai-torch`, a Python package that bridges PyTorch and Core AI by taking a model exported as a `torch.export.ExportedProgram`, traversing its FX graph node-by-node, and mapping ATen operators to Core AI operations to produce an `AIProgram` ready to run on Apple hardware.

An update in the post clarifies that the Private Cloud Compute Gemini models run in Google Cloud on NVIDIA GPUs. Apple's Security Research blog describes the architecture as applying many of the same layered security patterns used for PCC on Apple silicon — including dedicated processes with namespace isolation for network data parsing, short time-to-live recycled inference software, and attested keys held in a separate confidential VM. As with PCC on Apple silicon, all binaries will be published for public inspection. An iOS 27 Developer Beta is available, though access to the new Siri AI features requires joining a waitlist.

Key facts

1. 01Apple announced new Siri AI features at WWDC 2026, powered by a custom Gemini-derived model licensed by Apple.
2. 02Private Cloud Compute has been extended to Google Cloud using NVIDIA GPUs for 'the most demanding tasks, including agentic tool-use and complex reasoning.'
3. 03Vision LLMs are used to extract information from the user's screen, avoiding the need for individual apps to ship custom Apple Intelligence integration code.
4. 04A new Core AI library bridges PyTorch and Apple hardware via the `coreai-torch` Python package, mapping ATen operators to Core AI operations.
5. 05An iOS 27 Developer Beta is available, but access to the new Siri AI features requires getting through a waitlist.
6. 06PCC on Google Cloud applies layered security patterns including namespace isolation, short-TTL inference software recycling, and attested keys in a dedicated confidential VM.
7. 07All PCC binaries will be published for public inspection, consistent with the existing Apple silicon PCC approach.

Topics