Quick recap from part 1: Siri is a mess, and it doesn't matter, because the ecosystem locks everyone in. 2.5 billion devices, ~89% loyalty, a household you can't really leave.
So here's the obvious question. If Apple already owns the lock-in, why pay Google a billion dollars a year for a model?
This is where I disagree with the "Apple got humbled by Google" crowd. I think paying Google a billion dollars a year is one of the more clear-headed calls Apple has made in the AI era.
The model is a deflating asset. Frontier quality is converging and inference prices tend to collapse across the board. Look at the trend: the capability you pay a premium for this year shows up in a cheaper, smaller, often open model within a year, sometimes months. The frontier keeps moving, but the floor keeps rising, and the gap between "best" and "good enough" keeps shrinking.
Internally, Apple almost certainly evaluated OpenAI and Anthropic before picking Gemini, and it can renegotiate or swap providers later.
When the thing you'd spend tens of billions to "win" gets better and cheaper and more interchangeable every single quarter, it isn't a moat, It's a commodity input.
Spending to win that race is overpaying for something that depreciates the moment you ship it. The best models of 2026 will be the cheap, boring middleware of 2028. (HOT TAKE).
"So rent the commodity, own the appreciating asset."
This is just good capital allocation. Pay ~$1B a year for a model you can replace (and use the traffic to train your own in-house), then pour the difference into the thing that compounds and that no amount of money can source: 2.5 billion devices, a twenty-year developer ecosystem and the trust people put in the computer in their pocket that money can't buy. Period.
It's not easy to build an ecosystem, and it's even harder to maintain one. You can buy a model. You cannot buy a billion-device install base or two decades of switching costs. Failing at the model while owning the surface is survivable. Apple is living proof. Owning the best model while renting the surface (exactly the spot OpenAI and Anthropic are in) is the far more exposed bet. They have to earn every relationship from scratch. Apple already has them all.
Renting is a bridge, not a surrender. Apple didn't just plug Gemini into Siri; it built an architecture where the model is a swappable part piece ,look at the automotive industry - it's the same thing!
Now let's focus on another great solution coming from Apple again to support the ecosystem not the model, App Intents, Apple's answer to MCP, turns apps into things the operating system and other apps can call using LLM. The Foundation Models framework (introduced at WWDC 2025, expanded in 2026) lets Apple own the interface developers code against, with a model-choice protocol behind it. Core AI lets developers run models locally on Apple silicon. Put it together and Gemini isn't Siri's brain so much as a brain plugged into a socket Apple controls.
Think about what that means in practice. A developer codes against App Intents and the Foundation Models framework, not against Gemini. They never touch the model directly. So the day Apple's own model is good enough, Apple swaps it in behind the same interface and nobody, not the developer, not you, has to change a thing, or even no one should notice.
The model is a tenant, rented from google now, but apple owns the building.
Not only that, this way apple is building the biggest integration layer ever, a whole app can be installed on the computer and never appear on the screen, just get triggered by LLM.
And this isn't vaporware, it's the whole point of App Intents.
The framework itself shipped back in iOS 16 (2022), but what Apple wired into it recently is the interesting part. An app exposes its actions and its data to the system through schemas, a typed contract the OS actually understands. "Send the message." "Log the workout." "Add it to the cart." Apple Intelligence reads that contract and calls the action directly. No tapping. No opening the app. Sometimes nothing on the screen at all. The data side feeds Spotlight's semantic index too, which is how Siri gets your personal context without shipping your life off to some server.
So when a developer adopts App Intents, they're not bolting on AI. They're registering their app as a set of tools the OS can call for them.
And here's the part almost nobody is talking about (reported, not yet official): Apple is building system-level MCP support on top of App Intents. If that ships, the same actions you expose to Siri become callable by Claude, by ChatGPT, by any agent that speaks MCP, and the developer never writes a line of MCP themselves. Apple becomes the layer every app plugs into AND the layer every AI has to go through to touch your phone. Your apps on one side, every model on the other, Apple sitting in the middle collecting the toll.
Gemini today. Apple's own model tomorrow. Claude the week after. They all knock on the same door.
So no, Apple didn't lose the AI race! It looked at the race everyone's running, who builds the smartest model, and decided that's not the one worth winning. It's renting that part. What it's keeping is the thing nobody can buy.
So that's the bull case. But I'm not going to pretend it has no holes. Part 3 covers where this bet could break, and what the whole thing means for you.