John Ternus takes over as CEO of Apple on 1 September 2026. Among the files awaiting him, AI is the most talked about: the company has built a reputation on privacy and on-device processing, and now finds itself functionally behind competitors who made the opposite choice.
We have been covering Apple from a particular angle for months: that of a company which chose a constraint that others did not.
The initial bet, and what it costs
Apple's position is consistent and demanding: process as much as possible on the device, do not send personal data to servers, and use the cloud only within a controlled framework.
This constraint carries a high technical cost. A model that has to fit in a phone's memory is far smaller than a model served from a data centre. As we explained in our article on compression, you can reduce a lot, but not indefinitely.
The result is a functional gap that is visible to the user, in an area where comparison is immediate and brutal: you ask the same thing of two assistants, and one answers better.
What could change the game
Two developments make this bet far more tenable than it was a year ago, and that is what makes the moment interesting.
Small models are improving quickly. We have documented 27-billion-parameter models fitting on a single graphics card and a 30-billion-parameter agent. And we explained why in our article on distillation: every advance at the top filters down, a few months later, to what runs on modest hardware.
Privacy is becoming a commercial argument. This is the most interesting reversal. We reported this morning that half of the Fortune 500 would favour internally hosted models, precisely to avoid letting their data out. The reasoning Apple held alone two years ago has become mainstream.
This leadership change comes as a dispute between Apple and OpenAI heads to a hearing scheduled for 1 October. Beyond the conflict itself, this case will raise fundamental questions about what a platform can demand of the AI apps that install on it. It is an extension of what we described regarding the obligation imposed on Google to open Android to rival assistants: the question of who controls access to the user.
The real issue, which is not technical
The question facing this leadership is not about catching up on a benchmark score. It is about deciding whether the privacy constraint remains a principle or becomes negotiable.
These are two incompatible strategies. You can build the best possible assistant by sending everything to servers, or you can protect data and accept doing worse on certain tasks. Attempts at compromise generally produce systems that disappoint on both sides.
What makes this choice interesting to watch is that Apple is probably the only major company for which the second option is commercially viable, because it sells hardware rather than advertising.
What to take away
There is a broader lesson in this story, and it applies beyond Apple.
A company that accepts a constraint its competitors refuse looks behind as long as no one cares about the constraint. It looks prescient the day everyone starts caring. We do not know which of the two will happen here.
What we do know is that the question of where data goes has moved, in eighteen months, from a specialist concern to a purchasing criterion for half of America's largest companies. That rather favours those who anticipated it.