John Gruber on Apple's AI dilemma: Siri AI is fine for normal users, but Apple has no path to the frontier

Apple's AI dilemma: good enough for everyone, frontier for no one

John Gruber, founder of Daring Fireball, makes a sharp distinction that cuts through most Apple AI commentary: Siri AI is not trying to compete at the frontier, and that is a deliberate choice, not a failure of ambition.

The new Siri, currently in iOS 27 public betas ahead of a fall release, handles basic LLM queries well. It finds things in Apple Notes, searches iMessages as promised, and works with a simple squeeze of the side button. For the hundreds of millions of Apple users who have never opened ChatGPT or Claude, this will be their first real exposure to generative AI. Gruber says it performs well for that use case. Nobody following AI closely is impressed — that is beside the point.

The timeline, though, is damaging. Apple announced Apple Intelligence at WWDC 2024, promised meaningful features in the first half of 2025, and is now delivering them at the end of 2026. In a market moving as fast as this one, that is a significant slip.

“When people look at the new Siri AI that's in the OS 27 betas right now, is anybody who really is juiced into the whole AI system saying, wow, they've really taken the lead here in any way? No, absolutely not. This is very basic stuff. But it is going to be the intro to LLM-based generative AI for hundreds of millions of Apple customers who've never used any of this stuff. Is anybody seriously afraid that four years from now Apple is going to be the producer of the leading edge models? Nobody.”

No path to the frontier

Apple has no credible route to frontier model development, and Gruber argues the company is not trying to build one. Nobody excited about AI research is looking at Apple as the destination. The talent, the ambition, and the CapEx are all elsewhere — and Apple is not chasing them.

What Apple is doing instead is the classic Apple move: take something cutting-edge, strip it to the useful core, and push it to a billion users through hardware they already own. Gruber compares it to HBO in the 1980s — a better experience, no ads, funded by a premium device price. The formula has worked for decades.

The Google partnership and its contradictions

The financial logic underneath Siri AI is genuinely murky. Apple's existing Google search deal pays roughly $20–25 billion a year in near-pure-margin revenue — Google serves the ads, Apple's hands look clean, and Safari users never notice the arrangement. Siri has no equivalent model. Queries cost money to serve, and there are no ads to offset it.

Gruber believes Google cut Apple a favorable deal on the LLM backend specifically to blunt OpenAI and Anthropic. A world where a billion iPhone users default to a Google-powered Siri is a better outcome for Google than one where Apple had stuck with the original OpenAI partnership — which two years ago had ChatGPT branding embedded in Siri responses. That deal has visibly deteriorated.

The contradiction is real: Google is helping Apple commoditize the AI assistant layer to slow OpenAI and Anthropic, but a successful Siri eventually threatens Google's own search revenue. Gruber's read is that Google is betting on the "no moat" thesis — that no single company will own frontier AI, that the technology commoditizes, and that Google's scale survives commoditization the way it has in search, where Bing is technically comparable and largely irrelevant.

Apple's position is structurally similar. If there is no durable moat in frontier models, Apple does not need to own one. It needs optimized silicon for inference and a distribution advantage measured in billions of devices. That is what it already has.

Acquisition outlook

On whether Apple makes a significant AI acquisition, Gruber's answer is essentially no on the model side and maybe on silicon. The PA Semi acquisition — the hardware bet that became Apple Silicon — is his reference point. Apple is more likely hunting for a team with a breakthrough inference or chip architecture idea than for an LLM lab. At current market prices, even $1 billion may be a modest floor for anything meaningful.

The inference cost argument is where this lands concretely: if Apple can run slightly older, well-optimized models cheaply on its own silicon, the economics of a free, no-ad Siri become defensible — and the entire conversation about whether Apple eventually runs ads becomes less urgent. That is the bet Apple appears to be making, even if it has not said so publicly.