While OpenAI, Google, and Anthropic race to build the most powerful AI models in history, Apple has been quietly doing something different — and arguably smarter. Two developments from the past week reveal a strategy that may be more commercially durable than the frontier AI arms race: use Google's own Gemini model as a teacher to train smaller, more efficient Apple-exclusive AI models, while simultaneously building an AI App Store inside Siri. Apple isn't trying to out-model the model-builders. It's trying to own the platform where AI models get used.
The Deal No One Fully Explained: Apple Has "Complete Access" to Gemini
When Apple and Google announced an expanded AI partnership in January 2026, most coverage framed it as a follow-up to the pre-existing Google Search deal — Apple was simply adding Gemini to Siri, much as it had added ChatGPT in iOS 18. That framing dramatically undersold what was actually agreed to.
Reporting from The Verge on March 25, 2026, confirming an original scoop by The Information, reveals that Apple reportedly has "complete access" to Gemini in its own data centers — not merely API access, but access at a level that allows Apple to use the model as a training resource. Specifically, Apple is using Gemini to train what are described as "student" AI models via a technique called model distillation, optimized for running on Apple's own devices. These student models are smaller, faster, require significantly less computing power, and are specially tuned for on-device inference on Apple's hardware.
Neither Apple nor Google has officially confirmed this specific use of the partnership. Apple declined to comment when The Verge asked directly. But the detail, if accurate, fundamentally reframes what the Apple-Google AI deal is actually about.
What Is Model Distillation — and Why Does It Change the Calculus?
Model distillation is a machine learning technique in which a large, high-capability model — the "teacher" — trains a smaller, faster model called the "student." The student doesn't memorize the teacher's training data from scratch; instead, it learns to reproduce the teacher's behavior and output patterns, essentially inheriting compressed knowledge from a much more powerful system.
The technique was formally described in a 2015 paper by Geoffrey Hinton and colleagues titled "Distilling the Knowledge in a Neural Network" and has since become a standard tool across the AI industry. Google's own Gemini Nano — the version of Gemini built for on-device deployment on Pixel phones — is itself a distilled descendant of larger Gemini models. Meta, Mistral, and virtually every organization trying to run capable AI on constrained hardware uses some form of this approach.
For Apple, distillation from Gemini is a strategically elegant solution to a genuinely difficult problem. Building a frontier AI model from scratch — the kind capable of powering the next generation of Siri — costs hundreds of millions to over a billion dollars in training compute alone, plus years of research talent and infrastructure investment. The results are not guaranteed to be competitive. Apple's Core ML team is excellent, but it is not OpenAI's research division or Google DeepMind.
Distillation sidesteps that problem. Apple uses Gemini — a frontier model that Google spent several years and billions of dollars building — as the teacher. Apple's student models end up smaller and faster, but they inherit performance characteristics that would have been prohibitively expensive to develop independently. Critically, those student models can be optimized specifically for Apple Silicon: the Neural Engine in Apple's A-series and M-series chips, which on the M5 Pro delivers up to 38 TOPS of AI processing power. The resulting models run inference more efficiently on an iPhone than any general-purpose model designed for cross-platform deployment.
The competitive implication is striking: billions of dollars Google invested in building Gemini effectively become, in part, an Apple R&D subsidy. Apple extracts the intelligence, discards the compute overhead, and ships a model no one outside Apple can replicate on Apple hardware. It's a structural advantage that no other consumer device platform has access to.
Siri Extensions: The AI App Store Inside Your Phone
The second piece of the strategy arrived via Bloomberg's Mark Gurman in his Power On newsletter on March 29, 2026. iOS 27, expected to be unveiled at WWDC this June, will include a major new feature called Siri Extensions — a system that allows users to install third-party AI chatbots directly into Siri.
The scope is broader than the existing ChatGPT integration in iOS 18 and iOS 26. Rather than a single approved partner, Siri Extensions opens the door to a marketplace of AI integrations. As Gurman described it, Apple is "opening Siri and Apple Intelligence to third-party services." These extensions will get their own dedicated section in the App Store — creating, in effect, an AI marketplace embedded within the platform that already controls distribution to approximately 1.5 billion active Apple devices globally.
The Verge noted the obvious implication: given the number of AI companies competing for consumer attention, "it's hard to believe Apple will stop at just a couple of chatbots." ChatGPT was the proof of concept. Siri Extensions is the infrastructure for a full AI distribution platform.
The strategic logic is clean. Siri becomes an AI aggregator — not competing with ChatGPT, Claude, Perplexity, or Gemini in the model quality arms race, but packaging and distributing them through Apple's existing platform rails. The same App Store mechanics that drove the smartphone software economy for 18 years now apply to AI: Apple takes its commission, sets its content policies, and controls the interface through which AI reaches iPhone users.
Why This Strategy Is Brilliant — And Its Real Risks
The two developments — Gemini distillation and Siri Extensions — point in the same direction: Apple has concluded that it doesn't need to win at model-building to win at AI. It needs to win at two narrower things: on-device AI experience (where its hardware moat applies) and AI distribution (where its 1.5 billion device installed base applies). Both, conveniently, are things Apple already does better than anyone else.
The cost math is compelling. Training a frontier model on Apple's compute timeline would cost $500M–$1B+ and produce a model that would still trail OpenAI and Google at launch. Distilling from Gemini likely costs a fraction of that — and Apple can continuously update its distilled models as Gemini improves, getting performance upgrades without paying for frontier training costs each cycle. Apple Silicon's Neural Engine efficiency compounds this advantage: the student model runs faster and more privately on-device than a comparable cloud-inference product from a competitor without Apple's hardware integration.
The platform play is similarly well-positioned. Every AI lab that wants to reach iPhone users at scale will eventually need to appear in the Siri Extensions marketplace, just as every app developer ultimately needs to be in the App Store. Apple becomes the tollgate, the standards body, and the default interface for AI on the device most premium consumers already own.
But there are real risks worth naming:
The Google dependency: Apple's distillation pipeline runs through Gemini. If Google renegotiates or restricts the terms of the deal — perhaps as Gemini becomes more commercially valuable as a standalone product — Apple's student model training timeline is disrupted. This is a single point of failure for Apple's on-device AI roadmap.
Regulatory exposure on the AI App Store: The European Commission's Digital Markets Act and the DOJ's ongoing App Store scrutiny have already put Apple's distribution monopoly under intense pressure. An AI-specific marketplace, embedded within Siri and the App Store, is likely to draw even faster regulatory attention. The Epic v. Apple case established the legal playbook — expect every major AI lab to explore the same arguments the moment Siri Extensions goes live with a 30% commission structure.
Siri's brand credibility gap: Years of widely-mocked underperformance have made "Siri is bad at AI" a cultural meme durable enough to persist even after genuine improvements. The Siri Extensions strategy implicitly acknowledges the problem: by inviting better AI in, Apple is admitting Siri alone cannot satisfy users. Whether users will discover and trust Siri Extensions — rather than simply switching to Claude or ChatGPT directly — depends on execution Apple has not yet proven.
Model quality floor: Distillation produces capable models, but the student doesn't fully match the teacher. If Apple's on-device AI is noticeably inferior to what users can get from third-party apps, they'll route around it, reducing the stickiness of Apple's native intelligence layer and commoditizing the on-device AI advantage Apple is betting on.
The Q2 2026 Roadmap: What to Watch
WWDC 2026 (June): The expected venue for official iOS 27 announcements. Siri Extensions and Apple Intelligence updates will almost certainly headline the developer conference. How Apple frames the Siri Extensions developer opportunity — particularly pricing and commission terms — will be the first real signal of how aggressive the AI App Store play will be.
On-device model benchmarks: As independent developers and researchers get access to iOS 27 betas, watch for benchmarks comparing Apple's distilled on-device models against Google Gemini Nano (on Pixel), Samsung Galaxy AI, and Microsoft Copilot+ on ARM PCs. The competitive gap — or lack of one — will define how credible Apple's on-device AI differentiation actually is.
Google's response: Google has not commented publicly on the distillation use of the deal. But Google is simultaneously competing with Apple in the premium Android market and supplying the AI infrastructure Apple is using to train competing on-device models. That tension will not stay quiet indefinitely. Watch for any renegotiation signals in the next Gemini product cycle.
The foldable iPhone wildcard: Gurman's same March 29 newsletter described the foldable iPhone — expected in H2 2026 — as "the most significant overhaul in the iPhone's history." A larger, flexible display combined with distilled on-device AI models creates new surface area for AI-native interaction patterns Apple hasn't publicly detailed yet. The foldable and the AI infrastructure may be designed to debut together.
The prevailing narrative of the 2026 AI race is about who builds the biggest model. OpenAI is racing toward GPT-6. Google is iterating on Gemini 3. Anthropic is fighting for the right to sell frontier AI to the Pentagon. Apple is watching all of it — and building the platform those models will run on. Whether that's wisdom or a failure to understand what the AI race is actually about will probably be settled by the time iOS 27 ships in September.
