{"id":68530,"date":"2026-05-28T16:44:33","date_gmt":"2026-05-28T14:44:33","guid":{"rendered":"https:\/\/www.apfelpatient.de\/?p=68530"},"modified":"2026-05-28T16:45:31","modified_gmt":"2026-05-28T14:45:31","slug":"apple-google-ki-deal-details-nvidia-chips","status":"publish","type":"post","link":"https:\/\/www.apfelpatient.de\/en\/rumors\/apple-google-ki-deal-details-nvidia-chips","title":{"rendered":"Apple and Google: New details about the AI technology behind Siri"},"content":{"rendered":"

Just days before WWDC, a new report outlines how Apple plans to technically implement its upcoming AI features. At the heart of the plan is a miniaturized Gemini processor for the iPhone, combined with a cloud solution using Google servers and Nvidia chips. Apple remains committed to its privacy promise \u2013 but will have to rebuild its internal architecture to achieve it.<\/strong><\/p>\n\n\n\n

With the keynote on June 8th just over a week away, Apple's next generation of AI is drawing ever closer. While most reports so far have focused on the visible features \u2013 such as the completely redesigned Siri with its own app and system-wide search gesture<\/a> \u2013 a recent report delves into the underlying technology. It's not about what Siri will be able to do, but rather the infrastructure on which these capabilities actually run. And here it becomes clear that Apple's partnership with Google has a deeper impact on the architecture than previous headlines have suggested.<\/p>\n\n\n\n

A large model that is supposed to shrink to the size of an iPhone<\/h2>\n\n\n\n

At the heart of the strategy is a process called distillation. This involves using a large AI model to train a significantly smaller one, which then runs locally on the device. Specifically, Apple is said to be using a version of Google's large Gemini model to derive a compact variant that runs directly on Apple devices (via The Information<\/a>).<\/p>\n\n\n\n

This underscores a principle that has guided Apple since the beginning of its AI offensive: processing should ideally take place on the device itself, not in the cloud. According to the report, on-device processing remains a key selling point for the next wave of features.<\/p>\n\n\n\n

To further accelerate this shrinking of models, Apple is reportedly looking for smaller companies specializing in precisely this task. One acquisition target Apple is said to have considered is Liquid AI \u2013 a Cambridge, Massachusetts-based startup that focuses on running AI models locally.<\/p>\n\n\n\n

Why the cloud remains indispensable<\/h2>\n\n\n\n

As much as Apple relies on local processing, there seems to be no way around the cloud. Many AI queries will reportedly still require support from the data center. The reason lies in the sheer size of the complete Gemini model: According to the report, it has trillions of parameters and demands so much computing power that Apple has not yet been able to run it on its own server infrastructure called Private Cloud Compute.<\/p>\n\n\n\n

The solution is said to lie in using third-party infrastructure. Some of the requests to the new Siri version will run in the Google Cloud, specifically on a licensed version of Google's Gemini model. This would mean Apple would be reliant on its partner's servers for computationally intensive tasks \u2013 a remarkable move for a company that otherwise prefers to run its services in-house. That Google would provide the AI technology for Siri was already apparent when Google confirmed the use of Gemini for Siri by 2026<\/a>.<\/p>\n\n\n\n

Nvidia chips and the attempt to protect privacy<\/h2>\n\n\n\n

The most sensitive issue when outsourcing to external data centers is data privacy \u2013 and this is precisely where Nvidia comes into play. Apple is said to have recently approved the use of Nvidia's privacy technology for this environment, suggesting that Nvidia's AI chips are used for at least some of the computing work in the Google Cloud.<\/p>\n\n\n\n

The feature in question is Confidential Compute, a security function in Nvidia's graphics processors. It encrypts data and AI models while they are being processed. Activating the feature does slightly slow down request processing, but it could help Apple meet its privacy promises. The decision to implement this method was reportedly made only in recent weeks \u2013 an indication of how quickly Apple is adjusting the architecture.<\/p>\n\n\n\n

A familiar name for a changed technology<\/h2>\n\n\n\n

One interesting consequence of this restructuring, which many users may not initially notice, is that Apple will reportedly retain the term "Private Cloud Compute" for the next generation of Apple Intelligence features \u2013 even though these will no longer run exclusively on Apple's own servers.<\/p>\n\n\n\n

The brand name remains, while the underlying technology expands. What was once a purely in-house solution is evolving into one that also integrates third-party infrastructure and chips, but aims to maintain its data privacy promise through encryption during processing. According to the report, Apple continues to explore ways to operate AI functions in the cloud while preserving its own data protection standards.<\/p>\n\n\n\n

Apple's AI architecture is taking shape<\/h2>\n\n\n\n

The reports about the visible features and the reports about the underlying technology combine to form a coherent picture: Apple wants to develop Siri into a powerful assistant, but cannot handle the necessary computing load alone. The answer is a multi-layered model consisting of locally running mini-models, an in-house cloud, and \u2013 new this time \u2013 the Google Cloud with Nvidia hardware for the most demanding tasks.<\/p>\n\n\n\n

The extent to which Apple relinquishes control over its own AI can also be seen in the strategic importance of AI choice<\/a>. Which of these technical decisions Apple will actually reveal on stage and which will remain hidden behind the scenes will only become clear at the keynote. An overview of expectations for WWDC 2026 provides context<\/a>. (Image: Shutterstock \/ Mamun_Sheikh)<\/p>\n\n\n\n