The new CEO is John Ternus. A 25-year hardware engineer at Apple, responsible for the Mac's transition from Intel to Apple Silicon. Directly below him: Johny Srouji, head of chip development for a decade, now Chief Hardware Officer. The two top leaders don't come from software, not from services, not from AI. They come from silicon. That's no coincidence, that's a statement.

Apple has been functionally organized for 15 years. No iPhone team, no Mac team. A hardware team, a software team, a services team. Nobody owns the product; everyone converges on the iPhone and argues. That made the iPhone great. It also sank Apple Intelligence.

Generative AI is not an integration product. It's a speed race. The frontier labs ship a new model every quarter, sometimes monthly. Not because their people are smarter, but because their org chart allows one person to decide and push it through. At Apple, consensus has to be built horizontally. That works precisely for iPhones. For AI features, it means falling behind by a year. Or two.

Apple's board had two options. Install a software leader and try to force the organization into a frontier-lab cadence. Or change the game. They chose option two. Hardware to the top. Apple is structurally admitting that it cannot win the software velocity race and is placing its bet on a different playing field.

And this other playing field exists because the cloud AI business in its current form doesn't scale. Every major frontier lab is losing money on its top consumer tier. Sam Altman publicly stated that OpenAI loses money on ChatGPT Pro, even at 200 dollars a month. Not because of abuse. Because a capable model for a serious user costs more than any consumer subscription brings in.

The math is upside down. It is masked by investor capital, GPU supply growing roughly alongside demand, and the assumption that prices per token fall faster than capabilities grow. All three assumptions are wearing thin. Investors eventually want to see returns. GPU supply is constrained by power and fab capacity, not by Nvidia's willingness to ship. And frontier capabilities are currently scaling faster than prices are falling.

If nothing changes, this ends in a two-tier system. Large corporations with seven-figure contracts get real AI. Everyone else gets throttled consumer access. What you can do with AI will depend on which tier you can afford. That's not a forecast; it's happening right now. Every tightening of rate limits in recent months is unit economics making itself heard.

For Apple, this is threatening. Their customers' AI experience would be tethered to whatever labs can afford to deliver for 20 francs a month. You don't build a ten-year product strategy on a third party's loss-making business with built-in price hikes. Apple needs an alternative. There is only one.

The alternative is called: on-device inference. Running on the device itself. This is mostly sold on privacy grounds. But the real leverage lies in the cost structure.

Apple Silicon is the emergency exit from this meter. That's why Mac Minis sold out, why the popularity of open models exploded. Apple won't beat the best cloud model, and they probably won't try. They are betting on the long tail of what most people actually use AI for. Summarizing documents, drafting emails, transcribing meetings, translating, searching personal data. When that runs locally, it runs outside the meter.

Apple has made this bet before. In the 1970s, computing was a service. You rented time on a mainframe. The Apple II didn't beat the mainframe on raw performance. It brought a viable amount of compute power to a device you actually owned. Once bought, additional use cost nothing. VisiCalc could only emerge there. Same company, same structural move 50 years later.

Now comes the piece missing from public debate. There is a very specific buyer segment with a problem the industry has no clean solution for. Law firms. Medical practices. Fiduciary firms. Tax advisors. Wealth managers. Therapists. Every profession bound by strict confidentiality obligations, attorney-client privilege, medical privacy, fiduciary duty of care.

These firms watch their competition pull ahead with cloud AI and are not allowed to join in. Running client work through a cloud model is often a liability issue, a regulatory issue, or at best a massive technical headache. Even if it were formally compliant: clients could legitimately terminate the relationship the moment they learn their confidential data was processed by a third party's cloud model two tiers and two countries away.

What do these firms do? Many are converging on the same answer. They buy Mac Minis. A handful of M-series Mac Minis in a cluster is enough to run viable generative models locally. A few thousand francs of hardware in a closet, a private network, no outbound connection. The data never leaves the building. Professional secrecy holds. The compliance story is solid.

And before anyone mentions Private Cloud Compute: yes, Apple has that. Cryptographically attested, even Apple's admins cannot read the data. That is genuine progress compared to normal cloud AI. It is not the solution for this segment. A law firm's problem is not whether an admin can read along. The problem is the question: Can I attest to clients, regulators, and malpractice insurers that this data never left my physical control? No cloud service permits that statement.

So these firms improvise. Retail Mac Minis. Custom orchestration glue. An acquaintance who knows their way around. Open-weight models, fine-tuned for the domain. And hope the whole thing holds together. They do this because Apple hasn't built the product they need. And neither has anyone else.

There is no rack-mountable enterprise form factor for Apple Silicon. No clustering software. No admin tools for IT teams managing local inference. No identity layer that mirrors iCloud but stays on-prem. No HIPAA BAA agreements. No curated model ecosystem for regulated workflows. None of the infrastructure a law firm's IT director expects from an enterprise vendor.

The US professional services economy alone is measured in trillions of dollars and tens of millions of employees. A significant share of it has a structural need for AI that never touches the cloud. They know it. They are trying to buy. Nobody is selling it cleanly. Either Apple builds this enterprise stack, or a startup wraps Apple hardware in the enterprise layer Apple won't ship, just as third parties once wrapped IBM hardware in service layers. The window is open.

What does this mean for you in practice? Three perspectives.

An additional point for developers: The Valley has shipped iOS first in every new consumer category over the past decade. Instagram was iOS-only for 18 months. ChatGPT's mobile app debuted on iPhone. Threads, Bluesky, every premium consumer app. If local AI becomes a category, developer momentum is already pointing toward Apple Silicon. Apple doesn't need to convince anyone to build. They just need not to mess up the platform conditions.

And a note on upgrade dynamics: In the smartphone's first decade, the difference between a two-year-old device and the latest model was minor. That era is ending. If the on-device thesis holds, the generation of the Neural Engine starts to matter. The jump from M2 to M5 is palpable. The case for the flagship and more frequent upgrades hasn't been this compelling in a decade. Apple shareholders will love that.