The smarter Siri that Apple has been promising since 2024 finally has a release window, and the catch nobody wanted to talk about is now in the open: when Siri starts handling real questions later this year, most of the heavy thinking will happen on someone else's servers, not your iPhone.
Ars Technica, citing a new report from The Information, says Apple has spent the months since signing its Gemini deal trying to shrink Google's multi-trillion-parameter Gemini models down to something an iPhone can run by itself. The technique is called distillation: a smaller, faster model is trained to mimic the answers of a much larger one. The goal is to keep as much of Siri on the device as possible, in line with the privacy pitch Apple has built its brand on.
The reality is messier. Apple's Gemini-infused Siri will reportedly run partly on the phone and partly in the cloud. The local model handles simple requests. The bigger stuff, the kind of multi-step conversational work Siri keeps failing at today, will be shipped to a remote server running a full-size Gemini.
Editor
Kai T chevron_right
Table of Contents
The local AI problem nobody fixed
Phone makers spent years telling buyers their chips were ready for AI. That is half true. The Neural Engine inside an iPhone is built for fast, low-power contextual tasks, not for holding a trillion-parameter conversational model in memory. Google does not even pretend to run a chat-class Gemini on Android handsets; those queries go to the cloud by default.
Google does ship a smaller variant called Gemini Nano, but it powers contextual features like notification summaries, not a full conversational assistant. To match what people expect Siri to do, including the kind of follow-up questions that ChatGPT and Gemini already handle on the web, Apple needs the bigger model. The bigger model needs a data centre.
What it means for Malaysian iPhone buyers
This matters for Malaysians for two reasons. First, the iPhone is the dominant premium smartphone here. Apple Intelligence has technically been available in Malaysia since iOS 18.1 launched in October 2024, but the catch was that you had to switch your Siri request language to US English to turn the AI features on, per Apple's own newsroom guidance covered by soyacincau at launch. The next wave, the Gemini-powered Siri, gets its first reveal at WWDC on 8 June 2026, with a broader rollout expected in September alongside iOS 27.
Second, the cloud component is where the small print lives. Ars Technica reports that Apple's own Private Cloud Compute infrastructure, which runs on M-series Mac chips, has struggled to host an undistilled Gemini at all. The workaround is a deal with Nvidia. Heavier Siri queries will reportedly run inside Nvidia's Confidential Computing platform, which keeps user data encrypted while the GPU processes it. Apple may keep its Private Cloud Compute branding on top of that, but the silicon underneath is not Apple's.
The privacy story gets harder to tell
For years, Apple's pitch to Malaysian buyers worried about cross-border data flows was that the iPhone did its AI quietly, locally, on a chip you paid for. That story still holds for short, simple Siri tasks. It does not hold once your question lands in a Google or Nvidia-operated data centre to be answered.
This is not a scandal on its own. Encrypted cloud inference is a legitimate engineering choice, and every other consumer AI assistant already does it. The open question is whether Apple keeps charging the same ringgit premium for a privacy posture that now depends partly on partners.
What to watch at WWDC
The language Tim Cook uses on 8 June will tell you more than any demo. If Apple insists on labelling the entire stack as Private Cloud Compute, expect a careful asterisk about Nvidia and Google. If it gives the Gemini-powered side a separate name, that is Apple quietly admitting the old privacy story has a seam in it. Either way, for Malaysian owners of an iPhone 15 Pro or newer, Siri is about to get a lot more useful, the US English language toggle is unlikely to disappear at launch, and the answers you ask out loud will travel further than they used to.