Apple rents the frontier: the Gemini-Siri deal and the end of the default assistant slot

Apple is reportedly paying Google around $1 billion a year for a custom 1.2-trillion-parameter Gemini model to run the cloud side of Siri, and is shipping an iOS 27 framework that lets users swap in ChatGPT or Claude as the OS-level default assistant. The licensing fee is interesting. The Extensions framework is the actual story.

Two structural facts moved at once on Sunday, and they point in opposite directions. Apple, the most vertically integrated consumer hardware company in the world, is renting its flagship reasoning layer from a competitor. And Apple, which has held the default assistant slot on 1.4 billion iPhones since 2011, is opening that slot to third parties. Either move on its own would be notable. Together they say something specific about where value is settling in the consumer AI stack.

What was actually reported. Bloomberg, via TechCrunch and MacRumors, has the deal at roughly $1B/year for a bespoke Gemini variant — 1.2 trillion parameters, against the 150-billion-parameter model Apple currently runs in its own Private Cloud Compute (Apple's server-side inference environment, designed so query data stays inside Apple's perimeter).¹ The Gemini weights are reported to run on Apple infrastructure, not Google's. The arrangement is described as temporary while Apple keeps developing in-house. Apple tested Google, OpenAI and Anthropic; Google won.¹

Separately, iOS 27 will ship an Extensions framework letting users designate ChatGPT, Claude or Gemini as the system-level assistant — the slot Siri has monopolised since 2011.² OpenAI keeps an install path on iPhone but loses the WWDC 2024 exclusivity that made the ChatGPT integration valuable in the first place.

The licensing fee is a data point on inference, not a headline. A billion dollars a year, against Apple's installed base of roughly 1.4 billion active devices, is, depending on assumptions about query volume per device, somewhere between very cheap and merely cheap. If even 10% of those devices make one cloud-routed Siri query per day, that is 140 million queries per day, or about 51 billion per year, at frontier-quality reasoning. A billion dollars across 51 billion queries is about two cents per query, at the wholesale rate Google is willing to accept from a single buyer with maximum leverage.

That number tells you something, but only if you decide what it's telling you. The bull reading is that Google can clear margin on frontier inference at roughly two cents a query at this scale, which would make the inference-economics frame — that the binding constraint in the AI market is now the cost of running models, not training them — look less binding than the consensus assumes. The bear reading is that this is a strategic loss-leader: Google is buying Gemini distribution into the largest consumer surface in the world, and the contract is priced as a customer-acquisition cost, not a steady-state inference rate. The deal terms aren't disclosed in enough detail to settle it. Whether the $1B is a floor, a flat fee or a minimum guarantee against per-query metering changes the answer entirely.³

I'd lean toward the loss-leader reading, mainly because Google has every reason to want Gemini's brand inside the default Siri response on 1.4 billion devices, and Apple has every reason to extract a price that reflects that. But the honest answer is: we don't know, and the reporting doesn't tell us.

The Extensions framework is the structural move. The default assistant slot on iPhone has been a closed position since 2011. WWDC 2024's OpenAI deal didn't change that — it gave ChatGPT a privileged secondary surface, not the OS-level default. iOS 27 takes the default itself and makes it programmable. That is a different category of change.

In market-structure terms: the most valuable distribution position in consumer AI just got commoditised by the company that owned it. Apple is choosing to give up a monopoly position on its own hardware. That choice needs explaining, because Apple does not, as a rule, give up monopoly positions on its own hardware.

The most plausible explanation is regulatory. The EU's Digital Markets Act has been pushing toward programmable defaults across browsers, search and now assistants; opening the slot voluntarily, on Apple's own framework, is cheaper than having it opened by a regulator on someone else's terms. The second most plausible is that Apple has read the same charts everyone else has and concluded that the assistant layer is going to be programmable in two years anyway, so it might as well be the company shipping the framework.

Either way, the consequence is the same. The default assistant is now a setting, not a moat.

What this does to OpenAI. The WWDC 2024 deal was, in distribution terms, probably the single most valuable consumer AI agreement ever signed: roughly a billion devices, surfaced inside the OS, default-on. iOS 27 retires it. ChatGPT becomes one of three install options, behind a Gemini-powered Siri that is itself the do-nothing default.

The most defensible distribution position in consumer AI just became an install path.

Default persistence in mobile OS contexts is famously high — most users never change a default once set. So in practice Gemini wins the slot worth winning, OpenAI wins the brand-led active-choice users, and Anthropic gets the Claude-curious. This is a worse position for OpenAI than the previous one in every dimension that matters, and OpenAI's consumer story now depends on the ChatGPT brand pulling users into an active choice against the default, on a device where Apple controls the prompt design.

What this does to Apple's integration thesis. Apple's stated AI strategy in 2024 and 2025 was vertical: Apple silicon, Apple-trained models, Apple-controlled Private Cloud Compute. Renting a 1.2T-parameter Gemini model, eight times the parameter count of the Apple cloud model it replaces, is a concession that the in-house roadmap can't close the frontier-quality gap inside a competitive product cycle. The fact that the Gemini weights reportedly run inside Apple's own server environment preserves the privacy posture but not the model-sovereignty posture. Apple is renting the brain and owning the building.

This is fine, and probably correct on the merits. It is also a structural admission. The set of companies that can produce frontier-quality cloud reasoning models, at the cost and quality Apple needs, has narrowed to roughly three. Apple is not in the three. That is not a story Apple is going to tell at the keynote, but it is the story in the deal.

What to watch.

1. The actual iOS 27 default-assistant selection UI when it ships. Whether Apple presents Siri as the default with alternatives, or presents a chooser on first boot, is the difference between a 90% Gemini-Siri share and a 60% one.
2. Whether Google discloses Gemini inference unit economics at any point in the next four quarters. The Apple contract is now a reference price; analysts will back into the cost curve.
3. Whether OpenAI signs a defensive distribution deal, Samsung, a carrier, a Windows-level integration, to offset losing the iPhone default.
4. Whether Apple's in-house cloud model crosses the 1.2T-parameter threshold in the next eighteen months. If it does, the Gemini deal was bridge financing. If it doesn't, it was a structural decision dressed as a bridge.

Footnotes

¶