twenty-five years in hardware
Apple just handed the company to the man who builds the enclosures, in the middle of the model era. That is not a failure of nerve. It is a claim about where the scarcity went.
Tim Cook stepped down as Apple’s chief executive on 1 September. John Ternus has the job. Cook stays on as executive chairman, which is the arrangement you make when nobody is being pushed.
Ternus joined Apple in 2001 and became vice president of hardware engineering in 2013. Twenty-five years, all of them on the side of the building where things are made of aluminum and glass and silicon. He is the person who has been responsible for the physical object.
Now read the other half of the news. Apple does not have a frontier model. In January it signed a multi-year deal to run Siri on Google’s Gemini — reported at roughly a billion dollars a year, never confirmed by either company at that figure. The most valuable manufacturer on earth rents its assistant’s brain from a competitor.
So: the company without a model just promoted the hardware guy.
The obvious read is that Apple has given up on the model layer and is retreating to what it knows. I think that read is exactly backwards, and the evidence is in how the Gemini deal is built.
Gemini does not run on Google’s machines when it touches your data. It runs on Apple Silicon servers inside Private Cloud Compute — Apple’s own racks, Apple’s own chips, Apple’s own attestation story. Apple licensed the weights and kept the substrate.
That is not a company conceding the layer. That is a company deciding which layer it intends to own, and then paying rent on the other one while it builds. Apple’s model team is reportedly working on something around a trillion parameters aimed at replacing Gemini as early as 2027. The lease has an end date.
Which reframes the succession entirely. The question a board answers when it picks a CEO is not “who is the best executive.” It is “what is scarce for the next decade, and who is fluent in it.”
In 2011 the answer was manufacturing at scale, and Apple picked the man who had spent his career making a supply chain behave. That looked, at the time, like the operations guy inheriting a design company. Fifteen years later it is obvious the board had read the constraint correctly: the hard part of the 2010s was not imagining the phone, it was building two hundred million of them a year without the margin collapsing.
The 2026 pick makes the same shape of claim about the 2030s. Not “design is scarce.” Not “models are scarce.” Substrate is scarce — chips, thermals, memory bandwidth, power envelope, the physical body a model has to live inside if it is going to run next to you instead of in a building in Nueces County.
Simondon has the concept that makes this legible. In Du mode d’existence des objets techniques he describes technical evolution as concretization: an object starts as an assembly of separate parts, each solving one problem, and matures as those parts collapse into each other until a single element performs several functions at once and can no longer be cleanly removed. The early engine has a cooling system bolted on. The mature engine’s fins are structural. The primitive object is a collection; the concretized object is an individual.
Every part of the AI stack right now is the primitive kind. The model is a component. The inference silicon is a component. The memory hierarchy, the assistant, the privacy boundary — separate parts, bolted together, purchasable, swappable. That is precisely why Apple can rent one of them for a billion a year without the product falling apart, and precisely why everyone can rent the same one.
Concretization is the bet. Apple’s entire institutional competence — the thing it does that nobody else does at all — is taking a pile of separately-purchasable components and fusing them into one object with no seams. It did this to the phone. It did it to the laptop when it stopped buying Intel. A hardware engineer running the company during the concretization phase of a technology is not a retreat to the past. It is the correct specialist for the exact moment when the bolted-together thing becomes the fused thing.
And it explains the sequencing. You cannot concretize around a component you do not have. So you rent the component, keep it on your own silicon, learn its shape from the inside for two years, and build the replacement to fit a body you already control.
The steelman against this is strong and I want to state it properly, because if it holds, the succession is a mistake.
It goes: model capability is not a component, it is the product. The gap between a frontier model and a good-enough one is not closing, it is compounding, and it compounds on training compute — which is a function of capital and data-center access, not of industrial design. On that view Apple is doing what BlackBerry did, optimizing the object while the value migrates into the service running on it, and the fact that Gemini executes on Apple Silicon is a detail that will read, in ten years, like insisting the modem was yours.
That argument has one genuine weakness. It assumes the frontier keeps being the thing people buy. Most of what a person actually asks an assistant to do does not require the frontier, and the part that does can be dispatched to a building somewhere. If the useful envelope is “good enough, private, instant, on the device in your hand,” then the binding constraint really is thermals and memory bandwidth, and the company that controls the body wins by controlling what the body permits.
Nobody knows which of those is true. But the two positions are not equally hedged, and that asymmetry is the interesting part. Apple can lose the model bet and still ship the object. Google cannot lose the object and still ship the model — it has to get onto somebody’s hardware, and it just agreed to get onto Apple’s.
MacIntyre’s distinction between a practice and an institution is doing quiet work here. A practice has goods internal to it — things you can only get by doing the thing well, which practitioners recognize and outsiders cannot fake. An institution is the structure that keeps the practice alive, and it traffics in external goods: money, status, market position. The institution always threatens to consume the practice it exists to sustain.
Apple’s practice is integration. Its external goods are so enormous now that the practice can be, and sometimes is, replaced with procurement — buy the model, buy the modem, buy the maps data. Promoting a twenty-five-year hardware engineer is the institution reaching back for the practice, deliberately, at the moment it would be easiest to just keep buying.
It might be sentimentality. It might be the last correct read of a constraint before the constraint moves again.
The foldable ships this month, which means Ternus’s first public act as chief executive is a hinge. There is something almost too neat about that, and I do not want to lean on it.
The real test date is 2027, when the trillion-parameter model either arrives and is good, or arrives and is not, or does not arrive. Until then Apple is paying a competitor to hold a place in its own product, on its own silicon, in a building it owns.
That is either the most patient thing in the industry or the most expensive way to learn you were late.
Sources: ABC News ↗ · MacRumors ↗ · WBUR Here & Now ↗ · Reuters via MarketScreener ↗ · Simondon, On the Mode of Existence of Technical Objects ↗ · MacIntyre, After Virtue ↗