Who Owns the Front Door?

4:59

I followed Apple's WWDC keynote this year, and then most of a week of writing about it, waiting for someone to say the thing that seemed obvious to me by the end. No one did. So here it is.

Apple has stopped trying to build the best AI. It has decided to own the place where AI meets you instead. That is the real story of this keynote, and once you see it, the announcements stop looking like a pile of features and line up behind a single decision.

Here is why it matters. When AI starts doing real work for you all day, the question underneath everything is where that work runs, and who controls the place it runs. In a chat app you open on purpose? On a server farm that meters every request? Or inside the device you already own, next to your files and photos and messages, with a cloud behind it only when the phone is not enough? Apple spent ninety minutes building toward that last answer, and it did it by renting the parts of AI that are turning into commodities and keeping the one part that is not.

You can see it in what they were willing to rent. The new Siri runs on a version of Google's Gemini that Apple pays roughly a billion dollars a year to use. For a company that sells itself on building its own silicon and its own software, paying a rival for the brain of its assistant is a remarkable concession. Apple made it without apparent hesitation, because the raw model is becoming a commodity. Whoever has the best one this quarter will not have it next quarter. There is no point bleeding capital to win a race that resets every few months.

You can see it again in the cloud. Apple's "Private Cloud Compute" is the system that handles requests too heavy for your phone. It used to run only on Apple's own chips in Apple's own data centers. At WWDC, Apple said it now also runs on Nvidia chips inside Google's data centers. On its face that sounds like Apple handing over the sensitive part, the part that touches your data. It is the opposite.

Apple built the arrangement so that it keeps cryptographic control of that hardware even though Google owns it, down to a public, tamper-evident ledger of every machine in the fleet, with the software left open for outside researchers to inspect. Apple rents the data center. Apple keeps the keys. You are still trusting Apple, not Google.

Then there is the part almost nobody wrote about, which is the part that matters most if you build software. It is a framework called App Intents, and Apple just made it mandatory for any app that wants to work with the new Siri. App Intents is how an app tells the operating system what it contains and what it can do. In the old world, you opened an app, looked at its screen, and tapped the buttons yourself. In the world Apple is building, you ask the system, and the system reaches into the app and does it for you. The app becomes something the operating system can drive. Apple spent twenty years teaching every company on earth to build an app. It is now teaching the operating system to operate the app on your behalf. It cannot push this too far, because it still needs the App Store and the developers who fill it, so apps stay apps. But they have to become legible to the system, or the system cannot use them, and an app the system cannot use is about to become invisible.

Put the three pieces together. Apple rented the model. Apple rented the cloud. Apple kept the interface, the permissions, the data sitting on the device, and the trust of the person holding it. It gave away the parts that are becoming cheap and held onto the part that is becoming scarce.

I notice this pattern because I am building a clinical AI in medicine, where the same law is brutal and obvious. What kills most clinical models is not that they are dumb. It is that they cannot be trusted in a new place. A model trained at one hospital meets a second hospital's data and falls apart. The intelligence was real and it did not matter, because the thing that decides whether an AI is allowed to do real work was never the intelligence alone. It is the intelligence, plus the data it can reach, plus the trust to let it act for you. The model is necessary and it is nowhere near sufficient. Apple looks to have reached the same conclusion about consumer AI, and built for the sufficient part while renting the necessary one.

This is why the usual way of scoring the AI race is about to mislead you. The leaderboard measures the wrong thing. There are two real bottlenecks in this industry. The first is raw compute, the chips and data centers and power, and one company mostly owns that and is doing extraordinarily well off it. The second is the trusted surface, the place where AI actually meets a person, sees their context, and is handed permission to act. Almost no one talks about the second one, and it is the one Apple just spent a keynote trying to take. If the future of AI is mostly bigger models in bigger clouds, the compute owner keeps winning. If most of the AI that ordinary people actually use runs through the device in their pocket, then the device becomes the default and the giant clouds become specialists it calls only for the hard jobs.

Apple is not there yet. The model is borrowed. The trust has to be re-earned every day and can be lost in a week. But the plan is now legible, the same way Apple wants every app to be legible. So the next time you read that some lab has posted the new best model, note it, but know that it is not the thing that decides this. The thing that decides it is quieter. Whose device, whose operating system, whose front door does a billion people walk through when they finally let AI touch their real work? That is the last mile of AI. It is built out of trust. And Apple just told us it means to own it.

((Z))

.................................

June 13, 2026

.