John Ternus made a decision this week, and he made it in the open. In an interview with Steven Levy published by Wired on October 31, Apple’s chief executive said the company would use artificial intelligence across its products, and would not build AI for its own sake. Done properly, he said, a customer should never have to work out which part of an experience came from a model and which from ordinary software. The only question that matters is whether the product is good.
That is a decision about where Apple competes, not whether it competes. The company already ships AI in its phones, laptops and watches, and has since it introduced Apple Intelligence in 2024. What Ternus ruled out is the contest his two largest rivals have made the center of their corporate identity: the race to publish the strongest model on a fixed calendar, to top a benchmark table, and to announce the largest cluster of chips and power contracts in the same quarter.
His motive is the argument Apple has made about itself for four decades. The company sells finished objects, not components, and the value of an assistant inside a device depends on it disappearing into the task. A phone that announces it is reasoning about your calendar has failed at the only thing a phone has to do. With more than two billion devices in active use, Apple can treat the model as one input among several, alongside silicon it designs itself, an operating system it controls, and a support network no rival can reproduce. Google and Microsoft have to convince buyers that their models are the best available. Apple only has to convince them that the thing in their pocket works.
The comparison cuts both ways. Microsoft has tied Copilot to the software most offices already run and has spent heavily on data centers to do it. Google owns both a frontier model and the distribution to put it in front of billions of users. Both have accepted the capital intensity that comes with training at the frontier, and both have told investors to expect the bills to keep arriving. Apple’s capital spending remains a fraction of theirs, because it buys access to the model layer from partners instead of building all of it. That arrangement keeps its costs down and its dependence up.
Ternus was asked whether agents, which complete tasks across applications without a person opening each one, would make apps themselves redundant. He answered that making products is a human business, and did not elaborate. The answer is narrower than it sounds. Apple’s advantage has always been the last inch between a capability and a customer’s hand, and an agent that books a flight still has to be trusted with a credit card, a calendar and a face. That trust is an operating system problem, which is the problem Apple is best equipped to solve.
The cost of the decision is visible in the parts of the business Apple does not control. If the frontier moves to systems that reason across many steps and act without instruction, the companies that own those systems will set the terms on which everyone else integrates. Apple has been here before, with maps and with search, and in both cases it settled into the role of distributor for someone else’s capability, a position that paid well and left it unable to change the terms. A voice assistant leaning on a partner’s model has the same shape.
There is a talent cost too. The researchers who build frontier systems generally want to build them, not to fit them into a shipping product with a battery constraint and a privacy rule attached. Apple has lost senior people to competitors with looser mandates, and it will keep losing them while internal prestige sits with the device rather than the model.
And there is a timing cost, the hardest to price. Rivals can ship a capability the week it works and fix it in public. Apple’s method requires the capability to work on the oldest supported phone, in a dozen languages, without sending the user’s data anywhere it should not go. That is a slower road, and it has served the company well for twenty years. It assumes, though, that the capability in question is stable enough to be finished. If the useful version of AI turns out to be the one that changes every quarter, the discipline that made Apple’s products reliable becomes a tax on its relevance.
None of this makes the decision wrong. It is a bet that the model layer commoditizes and the integration layer does not, which is the same bet Apple made when it designed its own chips and let others write the applications. The question is what it costs to be right late. Apple has spent a decade demonstrating that it will pay for control, from silicon to retail to the cut it takes on other companies’ software. The bill for this decision will not arrive in the next product cycle. It will arrive in the one after that, when the company either has an assistant people trust with their money and their messages, or has a very good phone that someone else’s software knows how to use.
