Apple didn't lose the AI race. It declined to enter it

Today, John Ternus becomes CEO of Apple. He is not an AI researcher. He is a hardware engineer who has spent twenty-five years building iPhones, iPads, and Macs, and who ran Apple’s hardware engineering organisation for over a decade. At the exact moment the industry consensus says every technology company needs an AI visionary at the top, the world’s most valuable consumer technology company has appointed the man who knows the silicon.

I think that choice tells you more about the economics of AI than most of the commentary written about it this year.

The consensus

The standard account of Apple’s AI position goes like this. Apple fell behind. Siri upgrades slipped from 2024 to 2025 to 2026. Its in-house models were not competitive. In January it signed a multi-year deal to base the next generation of Apple Foundation Models on Google’s Gemini, at a reported cost of around $1bn a year, and analysts read this as a concession. One prominent analyst called it a way to ease short-term pressure rather than a long-term strategic shift. The subtext of most coverage was that Apple tried to build a frontier model, failed, and had to pay a rival to bail it out.

Every fact in that account is accurate. I’d argue the interpretation is wrong.

What Apple actually bought

Look at the structure of the Gemini deal rather than the fact of it. Apple pays a reported $1bn pa. For comparison, the frontier labs and hyperscalers are spending tens of billions annually on training runs and inference infrastructure, and committing hundreds of billions more in capex. Apple secured access to a custom frontier-class model for roughly one percent of what it costs to be in the business of making one.

The deal is non-exclusive. Apple evaluated OpenAI and Anthropic before choosing Google, and nothing in the arrangement prevents it from running that evaluation again in three years. The models run on Apple devices and inside Apple’s Private Cloud Compute, in Apple’s data centres, under Apple’s privacy architecture. Google supplies the weights. Apple keeps the customer, the data, and the interface.

There is a name for this arrangement, and it is not “surrender.” It is procurement.

Apple has decades of practice buying critical components from suppliers it does not fully control, and then squeezing them as the component commoditises. Qualcomm supplied modems while Apple built its own. Samsung supplied displays to the phone competing with its own flagships. Memory, sensors, and for years the very processors in the Mac all came from outside. Apple’s core competence has never been making every component. It has been owning the customer relationship and the integration layer while treating world-class components as inputs to be sourced, benchmarked, and replaced. The Gemini deal slots a frontier model into that same supplier slot. Everyone analysing this as an AI story is missing that it is a procurement story.

Why now is the moment this works

Three things have converged to make the sourced-model strategy viable in a way it was not two years ago.

First, model capability is commoditising faster than almost anyone predicted. Open-weight models are putting genuine pressure on frontier vendors, and the gap between the best model and the fifth-best model keeps narrowing on the tasks that matter for a consumer assistant. I use open-weight models at home on local devices for ~70% of my AI usage. The cost of securing high-quality AI capability has decoupled from the cost of creating it. When an input commoditises, the value migrates to whoever can switch suppliers, not to whoever makes the input. I have made this argument before about the AI stack broadly. Apple is the first company to run it at full scale.

Second, the privacy architecture finally lines up with the technology. Apple’s problem until now was structural: to give users a truly capable assistant, it would have had to let personal data leave the device for third-party servers, and hand third-party agents control over apps and surfaces inside its ecosystem. That breaks the promise Apple has spent fifteen years building its brand on. Running Gemini inside Private Cloud Compute, and increasingly capable models on the device itself, resolves the contradiction. Apple gets frontier capability without the data ever touching a supplier’s infrastructure.

Third, the hardware trajectory favours Apple specifically. The neural engine in Apple silicon is well suited to inference workloads, and each generation of A-series and M-series chips runs materially larger models locally. On-device inference has a marginal cost of zero. No API fees, no per-token economics, no inference bill scaling with usage. For a company with over two billion devices in the field, that is not an efficiency gain. It is a different business model from the one every cloud-centric AI company is built on. Customers get to run AI for free because they bought an Apple device.

Private Cloud Compute is a bridge, not a destination

This gets a little technical, but I think it matters for where this goes next. The custom Gemini model is reported to be around 1.2 trillion parameters. Nothing that size fits on a phone, and it will not for years, even with aggressive quantisation. So the sceptic’s response is that Apple’s on-device story is marketing, and the real intelligence lives in the cloud like everyone else’s.

I’d argue the hybrid architecture is really a solution to a fleet-transition problem. The binding constraint on on-device AI is not what the newest chip can run. It is what the installed base can run. The most advanced on-device features already require the latest Pro-tier hardware. If Apple shipped an AI experience that only worked on the newest devices, it would either strand hundreds of millions of older phones or fragment the platform. Private Cloud Compute lets every device access the same capability today, while the fleet turns over underneath it at Apple’s usual replacement cadence. Each hardware cycle moves the boundary: more tasks migrate on-device, fewer round-trips to the cloud tier, and the cloud model handles a shrinking set of frontier tasks. The frontier tier may stay hybrid indefinitely. The economics still move Apple’s way every year, because every task that migrates on-device is a task with zero marginal cost.

And that is why the choice of CEO is consistent with everything above. If the game were building frontier models, you would appoint a researcher. If the game is silicon, thermals, memory bandwidth, and the neural engine, you appoint the engineer who has spent a career on exactly those problems. I would not claim the succession proves the thesis, since Ternus was the anointed successor well before this strategy came into focus. But a company that believed its AI future depended on catching OpenAI would not have made this appointment. A company that believes its AI future depends on hardware would.

Where I’m betting

The honest caveats first. Apple’s delays were real, its in-house models did fall short, and none of this was executed as a grand plan from day one. Some of it is clearly making a virtue of necessity.

But strategy is partly about which necessities you allow yourself to be forced into. Apple let others spend the hundreds of billions to establish that frontier models exist, waited for the capability to commoditise, then bought it at component prices while keeping the customer, the data, and the device.

My prediction: within three years, Apple treats model suppliers the way it has always treated modem and display suppliers. Multiple vendors benchmarked against each other, contracts renegotiated downward, and an in-house alternative developed in parallel as leverage. The model layer becomes a line item in Apple’s bill of materials. I could be wrong about the timeline, and I could be wrong about whether the frontier tier ever fully migrates on-device. I do not think I am wrong about the direction.

As usual, I look forward to being told where I’m wrong.