Apple paid a billion for Gemini, then said its models hold not a drop of Google
Apple rebuilt Siri and Apple Intelligence on Google Gemini at WWDC, yet insists the result is pure Apple — and that careful wording exposes the real shift: stop building the best model, defend distribution and privacy instead.
Summary
At WWDC, Apple rebuilt Apple Intelligence and a new Siri on foundation models it calls “co-developed with Google.” Put plainly, the story Apple told for years about building its own AI is over. Before deciding, it evaluated proposals from OpenAI and Anthropic, then chose Google — reportedly paying around $1 billion a year, with the multi-year deal worth up to $5 billion.
But notice how Apple says it. Not “we use Gemini.” It says these are Apple Foundation Models, built with the help of Google’s technology through distillation and training, with the end result “pure Apple technology and code” — not a drop of Gemini inside. That convoluted phrasing is the most revealing part of the whole thing.
The move
WWDC 2026, June 8. The rebuilt Siri runs on iOS 27 and is essentially a full chatbot: web search, image generation, summarization, coding help, file analysis, multi-step commands. Architecturally, part of the model runs on-device and part runs through Private Cloud Compute, and Apple keeps repeating that user data stays under its control and that Google never touches user requests.
The money and the trade-off are out in the open. A multi-year contract, roughly $1 billion a year. Before signing, Apple weighed OpenAI and Anthropic and concluded that “Google’s technology provides the most capable foundation for Apple Foundation Models.” A company that built its identity on doing things in-house just handed the phrase “most capable foundation” to a rival.
The real motive
The stated reason is the best experience for users. The real motive is that Apple fell behind on frontier models, knows it can’t close the gap soon, and changed tactics: stop building the strongest model, buy it and distill it into your own shell, and defend the one layer you still hold firmly — devices, distribution, and privacy.
“Not a drop of Gemini” is not a technical claim. It’s a political one. It has to reassure three audiences at once: privacy-minded users, whose data never leaves Apple; investors worried about dependence on Google, since the ownership is Apple’s; and Apple’s own engineering culture, which needs to hear that it didn’t surrender, it just borrowed muscle. When a company needs phrasing this contorted to describe a deal, the deal has touched a nerve about who it is.
Who is threatened
The most direct loser is OpenAI. What it lost is Apple as the largest distribution channel there is — the default assistant on billions of devices, and it didn’t get it. Anthropic is out too. Both got stamped with “evaluated, not chosen.”
Google wins, but awkwardly. It proved its model is the most capable foundation available, yet the shape of that win is “distilled into someone else’s product, with your name not even allowed to show.” It got the money and the endorsement, and gave up the brand exposure.
The subtler case is Apple itself. Bought capability can fix the product but not the narrative. The day users realize Siri got smart because Google is behind it, the persuasive power of Apple’s on-device, privacy-first, built-it-ourselves trio gets diluted a notch.
Technical takeaway
“Co-developed, through distillation and training” points to something concrete: use Gemini as a teacher model, have it produce a large volume of high-quality outputs, then train Apple’s own student models on those outputs. The student’s weights belong to Apple, but its capabilities, style, and even its biases are largely learned from Gemini. So “not a drop of Gemini” holds at the level of code and weights, and falls apart at the level of capability lineage.
This also explains how Apple dares to run part of the model on-device. A distilled student is usually far smaller than its teacher, small enough to fit a phone’s compute and memory, with only the heavy requests handed to Private Cloud Compute. The lesson for builders: you rarely need to train a large model yourself. Distilling a small, specialized model out of a strong one is often cheaper, more controllable, and better suited to running at the edge — and Apple just stress-tested that path and showed it ships.
Builder impact
For anyone building AI products, this deal draws a line: the model itself is becoming a part you procure, and the real moat moved to distribution, data, and trust. Apple has no shortage of engineers or money, and it still chose to buy rather than build. For most companies, training a competitive foundation model of their own is no longer where the effort belongs; that effort is better spent on context, data contracts, and the user relationship.
What to ignore
Don’t take “not a drop of Gemini” literally. A model distilled and trained with Gemini carries Gemini in its lineage; the line is mostly legal and PR boundary-drawing, not technical purity.
And don’t rush to call this Apple “giving up.” If your read is that models have commoditized, then declining to burn cash on the best model and spending it on distribution and privacy may be the clear-eyed move. The real test is whether the product is good and whether the privacy promise survives scrutiny — not whose model wears the badge underneath.