The new Siri, powered by Gemini, went into service on Sunday. Apple, the only major tech company without a frontier model of its own, is handing its assistant over to a model built by a direct competitor. It is an admission of a delay we have been documenting for months, and it is probably not a bad decision.
Two weeks ago, writing about the change of leadership, we said the question was not about catching up on a score but about deciding whether the privacy constraint remained a principle or had become negotiable. The answer has arrived.
The original bet, and why it failed
Let us recall the position Apple held alone: process as much as possible on-device, do not send personal data to servers, use the cloud only within a controlled framework.
That constraint carried a technical cost that we explained. A model that fits in a phone's memory is far smaller than a model served from a data centre. You can shrink a great deal, but not indefinitely.
The result was a functional gap visible in an area where comparison is brutal: you ask two assistants the same thing, and one answers better. On a consumer product, that gap cannot be closed with an argument of principle.
Why it is probably the right decision
Three reasons, and they deserve to be set out before the criticism.
The delay was structural, not cyclical. Building a frontier model requires computing infrastructure that Apple does not have and that would take years to catch up on. Meanwhile, the product degrades against the competition.
A commercial agreement can be renegotiated. Unlike a technical choice, it is not a permanent commitment. Apple can change supplier, or return to an in-house solution once its models are ready.
Privacy can be contractual. An agreement can provide for processing in a dedicated enclave, without using the data for training. That is not the same guarantee as local processing, and it is not nothing.
A technical guarantee and a contractual guarantee are not the same in nature. When processing happens on the device, there is nothing to get out: the protection rests on physics. When it rests on a contract, it rests on what a company undertakes to do, and on the ability to verify that it does it. We wrote this about the shift of large companies towards internal hosting: what does not leave your premises cannot be read. That is precisely the argument Apple was selling, and which it can no longer sell in the same way.
What it changes for the market
This decision has a consequence that goes beyond Apple.
Google now finds itself powering its main competitor's assistant on mobile hardware, on top of its own ecosystem. We noted that Gemini had passed a billion users thanks to distribution rather than the performance of its flagship model. This agreement considerably amplifies that advantage.
It also raises a competition question that regulators will probably examine. We covered the obligation on Google to open Android to rival assistants. An agreement in which the same player powers the two main mobile systems moves in the opposite direction.
What to take away
There is a general lesson in this case, and it holds beyond Apple.
A company that accepts a constraint its competitors refuse looks far-sighted on the day the constraint becomes a criterion, and behind the times as long as it is not. Two weeks ago we wrote that we did not know which of the two would happen.
The answer arrived sooner than expected, and it is nuanced: the privacy concern has become the majority view — half of large American companies now favour internal hosting — but not fast enough for Apple to hold its position on a product judged day to day.
Being right too early in a consumer market amounts to being wrong. That is probably the fairest summary of this episode.