Apple's AI absence: strategy or stumble?

While OpenAI, Google, Anthropic, and xAI have spent the last two years shipping models on a weekly cadence, Apple has looked, from the outside, like a company sitting out the biggest platform shift of the smartphone era. No chatbot app. A delayed Siri that became a punchline. A keynote strategy that whispered "Apple Intelligence" while rivals demoed agents.

But the quiet is not the whole story. Two years after Apple Intelligence launched, the record splits cleanly in two: a set of shipped, on-device features that quietly work — and a flagship assistant failure bad enough that Apple agreed to pay a reported $250 million to settle a class-action lawsuit over it, and is now reportedly paying Google roughly $1 billion a year to bail Siri out. The question isn't whether Apple fumbled. It did. The more interesting question is what its on-device, privacy-first posture actually buys — and whether it's still a strategy, or just a comforting story.

The stumble: an honest timeline#

Strip out the commentary and the failures are stark:

At WWDC 2025, software chief Craig Federighi was unusually candid about the root cause: the company had a working first-generation version of the new Siri, scrapped it, and restarted on a second-generation architecture. "We had V1 working, but it just wasn't good enough when we looked at the whole picture — reliability, context understanding, app integration," he said. Marketing chief Greg Joswiak added: "People want an assistant who is always reliable rather than one who is 80% right."

That's a defensible engineering standard. It's also an admission that Apple spent two years and billions on an internal assistant architecture it couldn't ship. When an assistant can send messages, move files, and take actions across apps, an 80%-right assistant is genuinely dangerous — but rivals shipped ambitious assistants and iterated in public while Apple restarted in private.

The strategy: what "on-device" actually means#

Here's the part the absence narrative misses. Apple never set out to build a better chatbot. It built a two-tier inference architecture with a clear idea of where intelligence should live:

• On-device: Simple, fast tasks — rewriting a sentence, summarizing a note — run on a roughly 3-billion-parameter model executed locally on the Neural Engine in Apple silicon. No network, no latency, no data leaving the phone.
• Private Cloud Compute (PCC): Complex requests route to Apple-designed servers running Apple silicon, with cryptographic guarantees that data is processed but never stored, and that even Apple employees can't see it. Independent security researchers can audit those guarantees.

And the features built on this architecture are, by most accounts, genuinely good. Writing Tools across Mail, Notes, Messages, and Safari quietly became one of the most-used AI features on any phone. Visual Intelligence — point the camera at something, identify it, act on it — is the kind of ambient capability that makes a phone feel different. iOS 26 added over 20 new features, including Live Translation in Messages, FaceTime, and Phone calls. A Morgan Stanley survey cited in industry reporting put Apple Intelligence usage at roughly 80% of eligible US iPhone users.

The developer story matters too. At WWDC 2025 Apple opened its Foundation Models framework, giving third-party developers free access to the ~3-billion-parameter on-device model — no API costs, no internet dependency, no privacy liability to manage. That's a meaningful moat-builder: if Siri becomes the primary interaction layer, App Intents is how apps stay visible, and Apple's model is the cheapest inference on the platform by definition.

What the posture actually buys Apple#

So what does the privacy-first, on-device posture purchase, in hard terms?

Differentiation that survives contact with reality. Every other frontier AI company is monetizing inference in the cloud, which means user data flows through someone's server. Apple's PCC architecture — verifiable, not merely asserted — is the one design in the industry where "we don't see your data" is an engineering property, not a policy promise. In regulated industries and enterprise deployments, that's a real, durable edge, and it's why enterprise IT scrutiny of the new Siri's data routing will be the story that matters in 2026–27.

Distribution no lab can match. Reuters noted Apple's installed base of more than 2 billion active devices. When Apple Intelligence features work, they arrive on hundreds of millions of phones without a download, a signup, or a subscription. OpenAI has to win users one at a time; Apple starts with the phone in your hand.

A quality bar that protects the brand. Joswiak's "always reliable rather than 80% right" framing is easy to mock until you remember the notification-summaries episode: Apple shipped a feature that hallucinated headlines for news apps, disabled it, and took the hit. The stumble cost real money and credibility. But the restraint around shipping half-working AI into a system that acts on your behalf — messages, calendars, payments — is the same instinct that made people trust Apple with health data and biometrics. Trust compounds; demos don't.

The bill comes due#

None of that erases the invoice. The Gemini partnership is the largest third-party AI infrastructure deal Apple has ever signed, and it's an admission dressed as a partnership. Bloomberg's Mark Gurman reported the license runs roughly $1 billion per year for a custom Gemini model — neither company has confirmed terms — and Google Cloud CEO Thomas Kurian publicly called Google Apple's "preferred cloud provider," which sits awkwardly next to Apple's insistence that user data never reaches Google's infrastructure. Reports that advanced Siri queries in iOS 27 may route to Google's own cloud remain unaddressed by Apple, and enterprise IT teams have noticed.

Then there's the hardware gatekeeping. Apple Intelligence requires an iPhone 15 Pro or newer (the standard iPhone 15 is excluded), M1-or-later iPads and Macs, and a recent Watch paired to a compatible iPhone. With well over a billion iPhones in active use, the eligible base is a minority of the total. On-device-first is genuinely superior on privacy; it's also genuinely exclusive — and rivals deliver capable AI to far older, cheaper hardware via the cloud.

The talent question lingers too. Apple's secrecy culture has long been cited as a deterrent to research talent, creating a structural gap that a licensing deal can't close — you're renting someone else's research organization.

And the timing is pointed: Apple announced on April 20, 2026 that Tim Cook would become executive chairman and hardware chief John Ternus would take over as CEO on September 1, 2026. The AI record of Cook's final years — the Siri delay, the settlement, the Gemini pivot — is the unfinished business his successor inherits.

Takeaway#

The honest answer to "strategy or stumble?" is both — and they're entangled.

The on-device, privacy-first posture is a real strategy with real payoffs: verifiable privacy engineering, distribution at planetary scale, and a developer platform where inference is free. The shipped features — Writing Tools, Visual Intelligence, Live Translation — show what Apple means when it says it's "enhancing what people already do" rather than chasing novelty. That part of the story is underrated.

The Siri saga, though, was a stumble in the plain sense: two years of promised features, a scrapped internal architecture, a quarter-billion-dollar settlement, and a bailout deal with the company's most direct competitor in AI. The $1-billion-a-year Gemini license is the price of the strategy's blind spot — Apple optimized for reliability and privacy while underestimating how fast frontier models would redefine what an assistant must do.

What to watch now: whether the Gemini-powered Siri shipping this fall is genuinely competitive with GPT-class assistants or just competent enough to stop the bleeding; whether Apple publicly reconciles the PCC privacy claims with Google's cloud involvement before enterprise buyers force the issue; and whether Ternus, an engineer and hardware man, treats AI as Apple's next silicon-scale investment or keeps renting it. The silence is over either way. What replaces it will define Apple's next decade.