Apple rebuilt Siri on Google's brain. The device is still Apple's. The intelligence isn't.

Here is the sentence Apple would prefer you read past. The new Siri it unveiled at WWDC on Monday — conversational, context-aware, finally able to do the things it has promised since 2024 — does not, at its hardest moments, run on anything Apple made. The heaviest reasoning is handled by a custom 1.2-trillion-parameter version of Google's Gemini, executed on Google's cloud, on Nvidia's Blackwell chips, under a multi-year license reported to cost Apple around a billion dollars a year. The glass is Apple's. The customer is Apple's. The intelligence underneath is rented.

That is the story of WWDC 2026, and almost everything else in the keynote was arranged so you would not state it that plainly. Tim Cook's farewell as chief executive — he hands the role to John Ternus on September 1 and moves to executive chairman — gave the day a valedictory glow. The homeOS preview gave the developers something to build for. The redesigned Siri, with its own app and its conversation history, gave the demo its applause line. All of it is real. None of it is the power question. The power question is this: who now owns the part of Apple's product that actually thinks?

The answer, for the first time in the company's modern history, is not Apple.

What Apple actually conceded

Start with what the deal is, because the structure is the argument. Apple and Google announced the partnership back in January. At WWDC, Apple shipped it. Siri now routes a request through three tiers: simple tasks stay on the device, on Apple's own models; moderately hard ones go to Apple's Private Cloud Compute; and the genuinely demanding reasoning is sent out to Google Cloud, to a Gemini model tuned for Apple, running on Nvidia B200 GPUs. Apple says queries are anonymized and tokenized at each hop so that neither its staff nor Google can tie a request to a person. Take that privacy engineering at face value — it is genuinely good work, and it matters.

But notice what the architecture concedes. Apple kept the tier it could build and outsourced the tier it could not. The on-device layer is Apple silicon doing what Apple silicon is famously good at. The frontier layer — the part that makes the new Siri feel like a 2026 assistant rather than a 2021 one — is Google's. Apple did not license a feature. It licensed the capability that the entire keynote was built around, and it is paying a rival roughly a billion dollars a year for the privilege of putting that rival's mind behind its own brand.

Apple kept the tier it could build and outsourced the tier it could not. That is not a partnership. That is a dependency with a contract stapled to it.

I want to be precise about why this is a power story and not just a procurement story, because the difference is the whole column. Companies buy inputs all the time; Apple buys modems, glass, and memory from suppliers it could replace. Those are commodities. A frontier model that defines how your flagship product behaves is not a commodity, and the company you are buying it from is not a neutral supplier. It is the one firm on earth that competes with you for the same future — the assistant that becomes the front door to everything, the layer that sits above the operating system and decides what the device is even for. Apple just made that competitor its landlord.

Follow the money, both directions

Now hold this new billion-dollar payment next to the one that already runs the other way, because the pair of them is the real map of who has leverage.

Google pays Apple to be the default search engine in Safari — more than twenty billion dollars in 2022 alone, a sum a federal court found amounted to roughly a third of the search-ad revenue Google earned through Apple's browser. In August 2024, Judge Amit Mehta ruled that those default deals were how Google illegally maintained a search monopoly. The remedy, finalized in 2025, was milder than the government wanted: Google can keep paying Apple, but the deals can no longer be exclusive, they are capped at one year, and Apple must stay free to promote rival search engines and AI products. The Justice Department and a coalition of states filed to appeal in February. The case that defined the last decade of these two companies is still, technically, open.

So look at the flows. Google pays Apple twenty-plus billion a year to own the default question. Apple now pays Google about a billion a year to own the default answer. The numbers are lopsided, and the lazy read is that Apple still has the upper hand because the bigger check flows toward Cupertino. That read has it backwards.

The twenty billion Google pays is rent on distribution — access to Apple's users, a position Google has to keep buying because it does not own the device. The billion Apple pays is rent on capability — access to intelligence Apple has not been able to build, a position Apple now has to keep buying because it does not own the model. Distribution you can re-auction every year; that is exactly what Mehta's remedy forces. Capability compounds. Every quarter Siri runs on Gemini is a quarter Apple's own frontier effort is the fallback, not the product, and a quarter Google's model learns what a billion-device assistant needs to do. One of these dependencies gets easier to exit over time. The other gets harder.

The counter-argument that nearly persuaded me

Let me make the strongest case against my own thesis, because it is a good one and it deserves better than a strawman.

Apple, on this view, did the disciplined thing. It refused to torch tens of billions and several years chasing a frontier model it was clearly behind on, and instead bought the best one available on commercial terms while keeping the assets that actually defend a business: the silicon, the operating system, the install base, and the customer relationship. The Gemini deal is non-exclusive in spirit and replaceable in practice — if Apple's own models catch up, or if Anthropic or OpenAI offer a better deal, Apple swaps the engine and the user never notices, because the user was never talking to Google. They were talking to Siri. The privacy routing means Apple even denies Google the one thing Google most wants, which is the data. By this logic Apple has the leverage: it is the buyer, it owns the demand, and it has turned a weakness into a line item. Pay a rival a rounding-error billion, keep the trillion-dollar relationship, move on.

That argument nearly changed my mind, and I have changed it in this column before — I spent years insisting Apple's silicon independence made it the one big platform immune to the compute-concentration story everyone else was trapped in. I was wrong about the scope of that immunity. Owning the chip in the phone turns out not to be the same as owning the chip the answer is computed on, and in 2026 the second one is where the assistant lives.

Why the optimistic case still loses

The replaceability argument fails on a point Apple's own engineers would concede in private: you do not casually swap the model that your entire assistant has been tuned, evaluated, and shipped against to a billion devices. Switching a frontier provider is not changing a setting; it is re-validating behavior, safety, latency, and tone across every language and every feature that now assumes a particular model's quirks. The longer Siri runs on Gemini, the more the product is shaped to Gemini, and the more 'we could switch' becomes the kind of thing companies say right up until the moment they can't. Optionality you never exercise quietly stops being optionality.

And the 'we keep the customer' argument mistakes the performance of control for control itself. Apple still owns the relationship, yes — the brand, the bill, the trust. But the relationship is increasingly a relationship with Google's intelligence wearing an Apple interface. When the assistant becomes the primary way people use the device — and that is the explicit bet of a standalone Siri app — the company that supplies the assistant's mind is supplying the most important thing the device does, no matter whose name is on the back. Apple has spent fifteen years arguing that the integration of hardware and software is what makes its products magic. It just told you the most important software layer of the next fifteen years is one it buys.

Owning the chip in the phone is not the same as owning the chip the answer is computed on. In 2026, the second one is where the assistant lives.

None of this means Apple made an irrational choice. Faced with being years behind, renting the frontier while you try to close the gap is defensible, maybe even correct. But 'rational given the position' is not the same as 'a position of strength,' and Apple's marketing depends on you confusing the two. The company that built its identity on not needing anyone — on doing the hard, expensive, vertically integrated thing precisely so it would never be at a supplier's mercy — has, for the defining technology of the decade, put itself at a supplier's mercy. That the supplier is Google makes it sharper, not softer.

Who decides, when the model is someone else's

Here is the stake, and it is bigger than one keynote. The assistant is becoming the layer where the terms get set — what your phone will and won't do, what it surfaces, what it refuses, how it reasons about you. For a generation, Apple set those terms because Apple built the whole stack. Going forward, on the questions that matter most, the terms are negotiated with a model trained in Mountain View, run in Google's data centers, on chips neither company makes. Apple can still say no, can still wrap the thing in privacy guarantees and a beautiful app. But it is now saying no inside a structure it does not fully own, and structures, in the end, decide more than slogans do.

So watch the next two years for one number that no keynote will put on a slide: how much of Siri's frontier reasoning Apple has managed to bring back in-house. If that share climbs, the Gemini deal was the bridge Apple says it is. If it stalls — if a billion a year becomes two, if 'custom Gemini' becomes simply Gemini, if the in-house model stays the fallback nobody routes to — then Monday was not a product launch. It was the day the most powerful company in technology quietly conceded that, for the part that thinks, it had become a tenant.

Tim Cook's last keynote will be remembered for the farewell. It should be remembered for the lease.