Apple Just Handed the Keys to Its AI Future to Google — And Called It 'Siri'
Apple Just Handed the Keys to Its AI Future to Google — And Called It "Siri"
Here's the version of this story that dropped at WWDC 2026: Apple has finally shipped a new Siri. It understands context. It takes actions across apps. It can summarize, plan, and remember. It comes with a shiny dedicated Siri app, a Liquid Glass UI that bubbles out of the Dynamic Island, and a waitlist in the iOS 27 beta because Apple doesn't have the server capacity to give it to everyone at once. Cue the think pieces about Apple catching up.
And yeah — the new Siri is real, and for most users it will feel like a step change.
But that's the boring story. The more interesting one is the one Craig Federighi accidentally told in his post-keynote tech talk on Monday: Apple's new Siri runs on Google.
Not on Google's consumer Gemini app. Not on Google's search. Not on Google's Assistant. But the foundation models underneath the new Siri — the third-generation Apple Foundation Models, the AFM Cloud, AFM Cloud Pro, the AFM Fusion model, the image model — were developed in collaboration with Google, using technologies behind the Gemini family. Apple is essentially running a custom-forked Gemini on its own Private Cloud Compute servers, with its own on-device orchestrator model on top.
That is the most consequential admission Apple's made about AI since it killed the Project Titan car, and almost nobody is processing what it actually means.
The Architecture Apple Is Admitting To
Federighi walked through the stack in unusually direct terms. Strip away the keynote demo and the new Siri is built like this:
- On-device models handle speech recognition, voice synthesis, on-screen context understanding, and a handful of small tasks locally.
- A System Orchestrator sits above them, coordinating requests against an "App Toolbox" (for actions inside apps), a Spotlight Semantic Index (for personal content), and on-screen context. The orchestrator decides whether a request can be handled locally or needs to go to the cloud.
- When the orchestrator escalates, the request hits Private Cloud Compute — Apple's privacy-promise cloud stack. PCC runs Apple's AFM Cloud and AFM Cloud Pro models. These are the third-generation Apple Foundation Models, co-developed with Google, and "architected to run on our deployment architecture," in Federighi's words.
- For world-knowledge questions, the answer is grounded in Apple's World Knowledge Service — which Apple has been building for years, distinct from Google Search.
Federighi was very clear about what is and isn't Google. "The amount of the Google Assistant we use is none." Apple doesn't use Gemini's client code, doesn't use the Gemini models Google ships to its own customers, doesn't use Google's deployment infrastructure. What Apple took from Google, by Federighi's telling, is the base architecture and training techniques of the Gemini family — and then ran it on Apple's own silicon, in Apple's own data centers, behind Apple's privacy model.
That's a meaningful distinction. It is also, in production terms, the same thing as saying: we needed someone who knows how to build a frontier model, and we picked Google.
The Story Apple Is Telling vs. The Story It Just Lived
If you've been following the Apple Intelligence story since June 2024, this looks less like a strategic pivot and more like the landing of a long, quiet, embarrassing arc. Two years ago, Apple pitched Apple Intelligence as the moment its on-device models would outpace the cloud-bound competition. The original Siri overhaul — the one announced at WWDC 2024 — was supposed to deliver personal context, deep app actions, and on-screen awareness. It was delayed. The features that did ship were widely panned. The Apple Intelligence team went through a leadership shakeup that, by multiple reports, included John Giannandrea being moved aside and Mike Rockwell taking over Siri.
What got announced Monday is essentially the company that emerged from that shakeup choosing not to fight the last war. Apple's bet two years ago — that a focused, privacy-first on-device model could compete with the hyperscalers' parameter-count arms race — has been quietly abandoned. The third-generation AFM is, by Apple's own description, a cloud model. The on-device layer is now a thin orchestrator on top of a Gemini-derived engine.
In any other industry, this would be a "we tried, they won, we're partnering" press release. In Apple's telling, it is "the most powerful privacy architecture ever shipped in a consumer product." Both can be true, but only one of them explains why Google's name kept coming up at a company event where Google has historically been a punching bag.
The Real Tension: Compatibility, Capacity, and the EU
The new Siri has three visible frictions worth flagging, and they're the parts that will determine whether this lands as a hit or a slow-burn embarrassment.
One — The device cutoff is brutal. Apple confirmed the new Siri will only run on the iPhone 16 family, iPhone 17 family, iPhone 18 family, M2-or-newer iPads, and M2-or-newer Macs. Everyone with an iPhone 15 or older is permanently excluded, not from Apple Intelligence broadly, but from the new Siri. That's roughly four years of installed base cut off from the headline feature of the OS upgrade. Compare that to iOS 18 supporting devices back to the iPhone XR. Apple is willing to leave tens of millions of devices behind to make sure the on-device layer can actually handle the orchestrator's work.
Two — There's a waitlist in the beta. This is the most telling operational detail of the launch. iOS 27 beta users are reporting that accessing the new Siri requires joining a waitlist. Apple is, in effect, manually throttling access to a feature it announced as the centerpiece of the OS. Federighi addressed this in his tech talk by emphasizing that the System Orchestrator dynamically decides whether to run a request locally or escalate to Private Cloud Compute — and the cloud side has finite capacity. Apple built the waitlist because Private Cloud Compute for AFM Cloud and AFM Cloud Pro is not infinitely scalable. Google handles this by spinning up TPUs. Apple handles this by queueing users.
Three — The EU is locked out. The new Siri AI experiences will not be available in the European Union, due to the Digital Markets Act. Apple has spent three years blaming the DMA for delaying features from the App Store to iPhone Mirroring, and now the highest-profile AI feature ever shipped on a phone joins the list. The DMA is a real regulatory constraint, but so is the political reality that a major chunk of Apple's most engaged users will not get the new Siri at all in 2026.
What This Means If You Care About the AI Race
Zoom out. Three things just happened at once, and the strategic read is in the collision between them.
The foundation model race is now a four-player game at most. OpenAI, Anthropic, Google, and Meta are the only companies with the capital, data, and infrastructure to train frontier models at scale. Apple's announcement is the second time in a year that a major consumer-tech company has effectively conceded it can't go it alone — Samsung, of course, is the most visible example on the Android side, leaning hard on Galaxy AI to wrap Google's models. Apple is the most strategically significant concession yet, because it is the only one of the four biggest consumer-tech companies in the world to have staked its identity on building its own silicon and its own software stack. The handshake with Google is, quietly, a re-acknowledgment that silicon is not the same problem as intelligence, and Apple is great at the first and was not, by its own actions, winning the second.
The "personal context" moat is now the only thing Apple has. If the foundation model underneath Siri is essentially a forked Gemini, then the long-term defensible advantage for Apple is the integration with the system: the App Toolbox, the Spotlight Semantic Index, the on-screen context, the personal data graph. This is also exactly the part of the architecture Apple is going to make impossible to replicate on any other platform. The new Siri is the most deeply OS-integrated AI assistant ever shipped, and that integration is, in fact, a moat. It is just a much smaller moat than the one Apple was describing two years ago.
Privacy as a product feature is about to be tested harder than ever. Federighi spent a long time on the "we use no Google client code, we use no Gemini-deployed-to-Google-customers models, we run them in our own PCC" argument. That is technically true and operationally important, but the trust the company is asking users to extend is enormous: every Siri AI request that needs cloud compute will go to a model that was built in a Google research lab, trained on Google infrastructure, and adapted by Apple engineers. The fact that Apple can credibly say "but it runs in our data centers" is going to be enough for a lot of users. It is not going to be enough for everyone, and the next 12 months of security research on AFM Cloud and Private Cloud Compute is going to be the most consequential product review cycle Apple has ever gone through.
The Industry Read
The simple way to say this: Apple just built the most powerful privacy-preserving consumer AI product in the world, and the foundation underneath it is a Google model.
The more honest way: Apple just stopped pretending it could win the foundation model race on its own, and the thing it is shipping under its own brand is, by a wide margin, the best-integrated consumer AI on any platform. Both are true. Both are uncomfortable. The product is genuinely impressive, and the strategic concession is genuinely significant, and there is no version of this story where Apple comes out looking like it ran the table.
What Apple did not do, and this is the part most analysts will miss, is hand over user data to Google. The data path stays inside Apple. The model lineage goes back to Google. That distinction is going to be the single most argued point in the AI press for the rest of 2026, and Apple's ability to hold that line is going to determine whether the new Siri is remembered as a turning point or a fallback.
The bottom line: the new Siri is a real product, a real engineering achievement, and a real strategic retreat — all at the same time. The fact that Apple can ship all three in a single keynote is itself the story. Whether it is the start of a new era or the most polished end of the old one is the question every other AI company in the world is now quietly trying to answer.
The story of Apple's new Siri is not the polished keynote moment — it's the architecture diagram Craig Federighi drew on the stage at his post-keynote tech talk. That diagram is a confession, an admission, and a strategy reset all in one. DMC has been tracking this shift since the first Apple Intelligence delay, and we work with hardware and platform companies navigating exactly the same kind of strategic reset: when the model race moves faster than your roadmap, and the right answer stops being "build it ourselves." If your AI product strategy is being pulled apart by the same forces — let us know. We've been here before. Contact DMC here.