Apple iOS 27 Siri rebuilt on Google Gemini, ships after delays

Full analysis

Apple built the new Siri in iOS 27 on Google's Gemini, ripping out the OpenAI ChatGPT hookup that reviewers found flaky. The assistant now chains multi-step tasks: draft and send a message, read what's on screen, query your files and email, look through the camera viewfinder. Apple pays Google for the brains and keeps calling the wrapper Apple Intelligence.

This one's easy to undo in principle and hard to undo in practice. Apple swapped model vendors once already, so a second swap is technically on the table. But 1B+ devices, developer integrations, and a rebuilt Siri app make Gemini sticky fast. The deeper question isn't which model Apple chose. It's whether Apple has conceded that the frontier reasoning layer is a commodity it rents, and whether Google just bought itself the largest consumer inference contract on earth. No hard deadline beyond the iOS 27 ship date, which has passed.

The Skeptic

Apple traded one rented model for another. That's the whole move. Siri's decade of failure was never mostly the model. It was intent disambiguation, shallow app hooks, and users who stopped trusting it. Gemini raises the reasoning floor. It does nothing for the ecosystem friction or the trust rebuild. And notice the structure: Apple pays Google for the intelligence while competing with Google in services. If Gemini is the moat, Apple doesn't have one. If Apple Intelligence is the moat, why does Gemini headline the story? The clean "new model fixes Siri" narrative papers over the messy part, which is whether users who quit Siri years ago even come back to test it.

The Safety Lens

Apple sold privacy for a decade. Now the reasoning engine reading your email threads and camera feed is hosted by Google, the largest ad company alive. Private Cloud Compute was the pitch. That pitch gets harder to make when a third party does the thinking. The on-screen context access is the part that should worry people: a live viewfinder and pulled email are about as sensitive as consumer data gets, and the boundary between what stays on device and what hits Google's servers needs an auditable disclosure, not a settings footnote. EU regulators will treat this as a data-processor relationship needing fresh contracts. Most users will never know their assistant's brain moved to Mountain View.

The Compute Pragmatist

Google just inherited Apple's service-level obligations. Camera viewfinder queries and real-time drafting need sub-second responses at the moment a user asks, across a billion devices. That is a different beast from Gemini's existing developer API traffic, which tolerates slower, batchier calls. Google's TPU capacity planning got a giant new constraint it may not have fully modeled. The asymmetry runs Apple's way: Apple rides Google's compute spend and feeds back real user interactions that are worth a fortune as training signal, while writing a check that's tiny next to building a frontier lab. Google gets scale and data. Apple gets to skip the capex. Both can be happy until the concurrency bill lands.

The Enterprise Buyer

A CIO reads this as a warning about vendor concentration, not a Siri story. Apple, the company most obsessed with owning its stack, decided the reasoning layer was worth renting from a competitor rather than building. If Apple can't justify the cost of a frontier model in-house, the mid-size enterprise betting on its own fine-tuned stack should ask why it thinks it can. The other read: model choice is now a swappable component. Apple went OpenAI, then Google, inside two years. That's the procurement future. Models are interchangeable parts behind your own orchestration layer, with switching cost measured in integration work rather than loyalty.

Tensions

The Skeptic and the Compute Pragmatist disagree on who won. Skeptic says Apple exposed that it has no moat in intelligence. Pragmatist says Apple got the better end: someone else's capex, plus a data feedback loop, for a rental fee. Both can't fully be right about who holds the leverage in three years.

The Safety Lens and the whole "capability leap" framing part ways hard. The reviewers cheer a smarter Siri. The safety read is that the thing making it smarter is the exact thing that breaks Apple's privacy story, and nobody priced that in at launch.

And underneath: is the leap Gemini's reasoning or Apple's improved context plumbing? Nobody can tell yet. That's not a footnote. It decides whether Apple's next vendor swap is trivial or catastrophic.

Synthesis

The decision hinges on one belief: whether the frontier reasoning model is now a commodity part you rent, or a moat you must own. Apple, of all companies, just voted commodity. That's the actual news, bigger than Siri getting good. When the vendor most capable of building in-house chooses to rent from a direct competitor, the "own your model" thesis takes a real hit across the whole market.

The council leans toward the Pragmatist and the Enterprise Buyer read. Apple didn't lose by renting. It offloaded capex and kept the customer relationship, the distribution, and the interface. Google won scale and a data firehose. The loser is OpenAI, quietly dropped from the highest-distribution consumer AI surface on the planet.

What to verify before treating this as settled: whether Apple discloses, in plain and auditable terms, what leaves the device and hits Google's servers. That disclosure decides both the EU regulatory exposure and whether the privacy brand survives contact with the architecture. Second, watch grounding failure rates on the agentic actions in the first 90 days. "Send this message" acting on a misread email thread is a different class of error than a wrong trivia answer.

Prediction: By Apple's WWDC in June 2027, Apple will still be shipping Siri's core reasoning on a third-party frontier model (Gemini or a successor vendor), not a fully in-house replacement, confirming that Apple has conceded the frontier reasoning layer as a rented commodity.

Confidence: Medium. The architecture and vendor swap are set; only timing risk remains.

Why: Apple has now rented its reasoning layer twice in two years, first OpenAI's ChatGPT, then Google's Gemini, which is evidence Apple cannot yet match frontier reasoning in-house and has decided the economics favor renting over the enormous capex of a competitive lab. Building a frontier model to power a billion devices is a multi-year, multi-billion-dollar effort that Apple's own foundation-model results have lagged on, so replacing Gemini with an in-house model inside nine months is implausible. The opposite outcome, Apple announcing a homegrown frontier Siri by mid-2027, would require Apple to have quietly closed a reasoning gap it just publicly conceded by signing Google. The likelier surprise is a further deepening of the Google relationship, not an exit from it.

Revisit by 2027-06-30: We're right if Siri's headline reasoning at WWDC 2027 still runs on Gemini or another external vendor's model. We're wrong if Apple ships Siri's core agentic reasoning on a fully in-house Apple model as the default.

Apple keeping the interface and renting the brain is the same play a smart enterprise should be running: own the orchestration, treat the model as swappable. The company that spent forty years owning every layer just told everyone else it's fine not to.