Siri AI: Apple Teams Up with Google Cloud to Revolutionize 2026

At the 2026 Worldwide Developers Conference (WWDC), Apple unveiled advancements in Siri AI. Following the disappointing announcements of 2024, caution is warranted, but the new features appear to be achievable this time. Apple has integrated a model derived from Gemini, which they run on their own Private Cloud Compute infrastructure.

The new capabilities of Siri AI rely on vision LLMs to extract information directly from the user's screen. This innovative approach allows bypassing the need for each application to develop specific code to integrate with Apple's intelligence. In 2024, vision LLMs were still an emerging technology, but they have since matured.

Apple also introduced the Core AI library, a tool that represents a significant step forward for developers. This library is designed to optimize the use of Apple hardware, integrating with Meta's open-source PyTorch ecosystem. The Core AI PyTorch extensions, known as coreai-torch, serve as a bridge between PyTorch and Core AI. They enable developers to run an existing PyTorch model, exported as a torch.export.ExportedProgram, in a Core AI AIProgram ready to be executed on Apple hardware. This process involves traversing the FX graph node by node and mapping ATen operators to Core AI operations.

Developers can now install a beta version of iOS 27, which includes these new features. However, access to Siri AI requires going through a waiting list. Aaron Perris from MacRumors reported successfully getting off this list, raising hopes for concrete feedback on the functioning of Siri AI in the near future.

Additionally, Apple announced that the Private Cloud Compute Gemini models operate on Google Cloud, utilizing NVIDIA hardware. According to Apple's security research blog, for the most demanding tasks, including the use of agentic tools and complex reasoning, Apple has collaborated with Google and NVIDIA to extend its PCC infrastructure to Google Cloud systems. These systems use NVIDIA GPUs while maintaining Apple's robust security and privacy protections.

The PCC on Google Cloud leverages many architectural security models similar to those used on Apple silicon. The initial processing of network data for each request is carried out in a dedicated process, within its own namespace. The shared inference software is recycled with a short lifespan, and the attested keys are kept in a dedicated confidential VM, isolated from external inputs. As with PCC on Apple silicon, all binaries will be released for public inspection, ensuring increased transparency and security.