Apple’s on‑device pivot

Apple is reorganizing its artificial intelligence work around Amar Subramanya, the former Google executive now leading Apple’s AI effort, with Siri and other features expected to lean harder on models that run on the device itself. (apple.com) Apple said on December 1, 2025 that Subramanya would join as vice president of AI, reporting to Craig Federighi, as longtime AI chief John Giannandrea stepped down and moved toward retirement in spring 2026. The company has not publicly detailed a new Siri roadmap, but Apple has separately said Siri is designed to process requests on device whenever possible. (apple.com, apple.com) “On-device” means the model runs on an iPhone, iPad, or Mac instead of sending every request to a remote data center. Apple’s June 2024 Apple Intelligence launch described that setup as the default for many tasks, with a fallback system called Private Cloud Compute for requests that need larger server-based models. (apple.com, apple.com) Apple has tied that architecture to privacy for nearly two years. In June 2024, the company said on-device processing was “a cornerstone” of Apple Intelligence, and its security documentation says Private Cloud Compute is meant to extend iPhone-style protections into the cloud when local hardware is not enough. (apple.com, apple.com) The tradeoff is scale. Apple’s own machine learning researchers said the first Apple Intelligence system paired an on-device language model of about 3 billion parameters with a larger server model on Apple silicon, a smaller local footprint than the giant cloud models many rivals have emphasized. (apple.com) That choice also fits Apple’s spending profile. Apple reported $34.55 billion in research and development expense on $416.16 billion in net sales for fiscal 2025, or about 8.3% of revenue, while continuing to frame AI as part of its hardware-and-software stack rather than as a standalone cloud service. (apple.com, apple.com) Apple has already opened some of that local model stack to outsiders. At Worldwide Developers Conference in June 2025, it said developers would be able to access the on-device large language model at the core of Apple Intelligence, and that Shortcuts could tap either on-device models or Private Cloud Compute. (apple.com) The company’s public language still leaves room for cloud partners. Apple’s support documentation says Apple Intelligence first checks whether a request can be handled on device, then sends more complex work to Private Cloud Compute, and Apple has also said some requests can be routed to ChatGPT with user permission. (apple.com, apple.com) So the shift under Subramanya is less a rejection of the cloud than a decision about where Apple wants the center of gravity. For Siri, that keeps the company close to the position it has repeated since 2024: process personal requests on the device first, and reach for bigger servers only when the job demands it. (apple.com, apple.com)