A two-tier industrial laboratory at night, an amber spotlight illuminating only the lower tier of lit workstations; the upper tier in darkness and abandoned

Every company building the AI era picked a software CEO. Sam Altman, Dario Amodei, Demis Hassabis, Satya Nadella's AI group, Mark Zuckerberg's labs, Elon Musk's xAI — researchers, product managers, platform engineers. On Monday, Apple named John Ternus as Tim Cook's successor. Ternus runs hardware. His signature project was transitioning the Mac off Intel chips and onto silicon Apple designed itself.

The same quarter the succession was being finalized, Apple's software chief was quietly routing the next Siri to other companies' models.

The two moves are not contradictory. They are the same bet.

What the Chips Were For

Apple Silicon began as an escape plan. For a decade after the iPhone, the Mac ran on Intel — chips Apple didn't control, shipped on Intel's roadmap, running hot in laptops Apple's engineers wanted to build thinner. In 2008, Apple bought PA Semi and started hiring chip designers. The first A-series processor shipped in the iPhone 4. The M1 landed in 2020. The two-year Mac transition was handed to Ternus, then a VP of hardware engineering. It finished in 2023.

The project was defensive. Intel's roadmap had slipped for five years. The M1 was Apple finally taking control of the physical layer that everything on a Mac depended on — the layer Intel had defined for the industry since 1978.

Nobody argued the obvious second-order claim at the time: the chips were also an investment in a capability Apple would need for something else.

The Substrate

In May 2024, Bloomberg reported that Apple's first server chips for its upcoming AI features would be the M2 Ultra. Simpler AI tasks would run directly on iPhones, iPads, and Macs; heavier ones would hit data centers running Apple Silicon. The architecture wasn't a pivot. It was the natural extension of a fifteen-year bet: if you own the chips, you own the inference cost, the battery envelope, the privacy model, and the latency profile.

May 2024
Sources: Apple's first server chips for its upcoming AI features will be the M2 Ultra; simpler AI tasks will be processed directly on iPhones, iPads, and Macs
Bloomberg

A month later, Apple detailed Private Cloud Compute — AI servers running custom hardware based on Apple Silicon, with code independent experts could verify. In December 2024, The Information reported that Apple was working with Broadcom on its first AI-specific server chip, codenamed Baltra, for 2026 mass production. In October 2025, the M5 debuted with a Neural Accelerator in every GPU core and over 4x the AI compute of the M4. Apple's own marketing called it "the next big leap in AI performance for Apple silicon."

None of this was a response to Ternus becoming CEO. Ternus became CEO because of it.

Three Apples

Apple has had three distinct identities since 2007.

The first was Jobs's Apple — a hardware-led company that had just shipped a phone, built the vertical it would grow into, and believed it needed to own the important pieces of the stack because every piece was load-bearing.

The second was Cook's Apple. In fifteen years, market cap grew from $350 billion to more than $4 trillion, services revenue crossed $100 billion a year, and the company became the global operational machine — China manufacturing, recurring revenue, the App Store compounding. Cook didn't abandon hardware. He made it predictable. The A-series cycle, the M-series cycle, Apple Park, the supply chain discipline — innovation on a schedule. Investors learned to price it.

The third is the one that started Monday. Ternus isn't an operational CEO. Srouji, the chip chief, was promoted the same day to chief hardware officer — a title Apple had never used. The secret robotics team that used to sit inside Apple's AI division was moved under Ternus in April 2025. Design reported to Ternus by January 2026. The succession wasn't a single choice. It was a stack reorganization, and the CEO pick was its final expression.

Each phase seemed like the identity of the company. Each turned out to be a transition to the next.

What Apple Gave Up

The other half of the reorganization is happening on the software side, and it runs in the opposite direction.

Apple's AI software problem is the best-documented story in tech. Siri, launched in 2011, has been a regression for a decade. The LLM-based rewrite was reported to have slipped to 2027. In January, The Information described how Craig Federighi, Apple's software chief, was leading a push to use other companies' models to ship a revamped Siri. The WWDC "26" teaser confirmed the timeline. On the models themselves, Apple has essentially opted out.

January 2026
A look at Craig Federighi's approach to AI at Apple; sources say he led a push to use other companies' models to help launch a revamped Siri later in 2026
The Information

In practice: the chips are Apple's. The data center chips are Apple's. The robotics are Apple's. The models are whoever is best that quarter.

The chips are Apple's. The models are whoever is best that quarter.

This is not a weakness disguised as a strategy. The model layer is commoditizing in real time. OpenAI's pricing power collapsed as Anthropic, Google, Meta, and four Chinese labs closed the gap. OpenAI and Anthropic have converged on the same use cases. ChatGPT ad rates just fell from $60 CPM. Token prices drop every quarter. For a hardware company, renting the best model this quarter and a different best model next quarter is cheaper than trying to employ the best researchers.

Silicon doesn't commoditize the same way. A chip design cycle is five years. A foundry relationship is a decade. Apple's silicon team — Srouji, Ternus, the hardware engineering group — spent fifteen years building a capability the company's peers could not reproduce if they started today and hit every milestone. That is the asset Apple is betting on.

The Concession

Apple Silicon was designed to escape dependency on a chip company Apple didn't control. The project worked. The Mac doesn't need Intel. The iPhone doesn't need Qualcomm's application processor. The iPad, the Watch, the Vision Pro, the AirPods, the home robot Ternus's team is quietly building — all run silicon Apple designed.

Fifteen years after the project started, Apple is rebuilding exactly the dependency the chips were built to escape. This time the layer is models, not processors. This time Apple designed the dependency on purpose.

The bet is that the model layer — unlike Intel in 2005 — will never consolidate into a single supplier with pricing power over Apple's roadmap. The bet is that there will always be three or four frontier models competing, that Apple can route between them quarterly, that none will be essential, and that the thing Apple owns — the silicon that runs the models, the devices that ship them, the battery and privacy and latency envelope they run inside — will be worth more than the models themselves.

If Apple is right, it just picked the CEO who makes the asset. If Apple is wrong, it picked the CEO who ships the bundle that no longer matters.

Either way, the pick is the answer to a question Apple never stated. Every peer in the AI era picked a software CEO. Apple picked the chip guy because Apple decided the chip was the product, and the model was the component.