This week Google Cloud CEO Thomas Kurian publicly confirmed what Bloomberg first reported in November: the next generation of Siri runs on a custom Google Gemini model, not on Apple's own AI. Apple is paying roughly $1 billion a year for it. The model is a 1.2 trillion parameter version of Gemini built specifically for Siri, about eight times larger than the 150 billion parameter system behind today's Apple Intelligence. First features are rolling out now in iOS 26.4. The full conversational redesign gets its reveal at WWDC on June 8.

The headline most outlets are running is some version of "Apple admits defeat." That read is lazy, and for an audience that has actually run businesses, it misses what is interesting. The defeat framing is real but incomplete. What Apple actually did here is a more instructive piece of capital allocation than the coverage suggests, and it is worth understanding properly because it is a template you will see again.

Strip the emotion out and here is the decision Apple faced. Building a competitive frontier model in-house is a multi-year, multi-billion-dollar effort with an uncertain completion date, and Apple was visibly behind. Its 2024 Apple Intelligence launch underwhelmed. Its smart Siri slipped. Meanwhile the gap between Apple's 150 billion parameter model and a true frontier model was widening every quarter, and customers were forming the habit of reaching for ChatGPT instead of Siri. Habits are the asset Apple cannot afford to lose. The entire value of the iPhone is that it is the default.

So Apple made the rational move that founders make when speed matters more than pride. It bought the capability instead of building it. A custom 1.2 trillion parameter Gemini model, eight times its current size, benchmarked at a 92% success rate on complex multi-app queries versus 58% on the old architecture. Available now, not in 2028. The price, about $1 billion a year, is real money but trivial against Apple's roughly $100 billion in annual net income. It is rounding error for a company this size to stop the slow erosion of its most important product.

The privacy engineering is the part that took real work. The Gemini model runs on Apple's own Private Cloud Compute servers, not Google's. Apple says queries are anonymized, Google sees the processed query rather than raw user data, and nothing is retained or used to train Google's models. Security researchers correctly note that Private Cloud Compute is a new attack surface worth watching. But the architecture is the whole point: Apple found a way to outsource the intelligence without outsourcing the customer relationship or the data. That is the hard part, and it is why this took until 2026.

And critically, this is a bridge, not a marriage. Apple is building its own roughly 1 trillion parameter model in parallel, with custom server silicon codenamed Baltra targeted for late 2026 production. The Google deal is explicitly designed to be temporary, non-exclusive, and replaceable. Apple is renting frontier capability while it builds its own, and it structured the deal so it can walk.

Three things worth holding onto, because they generalize beyond Apple.

First, the "build versus buy" line for frontier AI has moved decisively toward buy, even at the very top. If Apple, with the largest cash position and engineering bench in the world, concluded that renting was smarter than building on its own timeline, that is the clearest signal yet that frontier model development has become too capital-intensive and too fast-moving for all but a handful of labs to lead. Expect more large, capable companies to license rather than build. The model layer is consolidating into a few suppliers, and everyone else becomes a distribution and product layer on top.

Second, distribution is becoming the durable moat, not the model. Google lost the chatbot mindshare war to ChatGPT and is winning a quieter, more important war by embedding Gemini into roughly 2 billion Apple devices on top of its own Android base. The lesson for anyone allocating capital or building a company: the model is increasingly a commodity input, and the defensible position is owning the channel to the user. Apple owns the channel. Google is paying to rent space in it while also supplying its brain. Both are playing the distribution game. Neither is betting the company on model supremacy alone.

Third, the privacy-as-architecture approach Apple used here, run someone else's model on your own controlled infrastructure, is going to become a standard enterprise pattern. Any regulated business that wants frontier capability without handing data to a model provider is going to want exactly what Apple built. If you are evaluating AI vendors inside a serious organization, that architecture, model access decoupled from data exposure, is the thing to ask for. Apple just productized the template at planetary scale.

Some Cool Stuff Worth Noting

While Apple was renting Google's brain, Alibaba made a quieter but telling move. It launched Qwen 3.7 Max, and for the first time, it kept the weights closed.

That matters because Alibaba built its reputation on open weights. Qwen 3.5 and 3.6 were among the most-used open models in Western developer workflows, the cheap, downloadable, self-hostable workhorses that helped Chinese models hit 60% of usage on OpenRouter. Qwen 3.7 Max breaks that pattern. It is API-only, proprietary, available through Alibaba Cloud and aggregators like OpenRouter, with no downloadable weights as of now. The smaller open variants are expected in June or July, but the flagship stays closed.

The model itself is strong. It scores 56.6 on the Artificial Analysis Intelligence Index, ranking fifth overall and the highest-placed Chinese model on that leaderboard at launch. It runs a 1 million token context window and tops Claude Opus 4.6 on several agentic coding benchmarks. Pricing is roughly half of Claude Opus 4.7, at $2.50 and $7.50 per million tokens. Alibaba claims it can run autonomously for 35 hours and chain over 1,000 tool calls in a single session, though that number is Alibaba's own and has not been independently reproduced yet.

The strategic read ties back to the top story. Apple decided the frontier was too expensive to build alone and rented it. Alibaba decided the frontier was valuable enough to stop giving away and closed it. Both moves point at the same conclusion: frontier-grade AI has become too costly to treat casually. The era of labs giving away their best models to win goodwill is ending. The best models are now the product, not the marketing.

Two companies made opposite bets on the same question today. Apple is paying Google about $1 billion a year for a custom 1.2 trillion parameter Gemini model to power a rebuilt Siri, confirmed by Google's cloud chief this week and fully revealed at WWDC on June 8. The lazy read is that Apple lost the AI race. The truer read is that Apple bought speed instead of building it, used its distribution leverage to get favorable terms, and still nets roughly $19 billion a year from Google even after paying for the model. Meanwhile Alibaba launched Qwen 3.7 Max and, for the first time, kept the weights closed, abandoning the open-source approach that made it the most-used Chinese model family in the West, in favor of a proprietary frontier flagship priced at half of Claude Opus. Apple decided the frontier was too expensive to build alone and rented it. Alibaba decided the frontier was too valuable to give away and closed it. Both point at the same conclusion: frontier-grade AI has stopped being something companies treat casually. It is too costly to build, too valuable to gift, and now the central asset everyone is reorganizing around. The build-it-yourself era and the give-it-away era are both ending in the same week.

Stay building. 🤖