Start with what actually shipped. Apple did not build a frontier model. It licensed one. The cloud intelligence behind the new Siri is a custom Gemini model from Google, paid for annually, while a smaller model runs on the device for the requests it can answer without leaving the phone, and anything larger routes to Apple's own private compute. On top of that, Extensions lets the user choose a different provider entirely. The intelligence is sourced, swappable, and increasingly interchangeable. What Apple kept under its own control is everything that is not the model: the privacy boundary, the on-device fallback, the way the assistant reads the screen and reaches across apps, the feel of the interaction.
That division is the whole point. For three years the industry treated the model as the product. The race was for the highest benchmark, the biggest context window, the newest release, and the assumption was that whoever held the best model held the advantage. Apple, with more buying power to commission a custom model than almost anyone, looked at that race and declined to run it. It treated the model as a component to be sourced on the best terms and spent its design effort on the surfaces a customer actually touches. When the company with the deepest pockets rents the engine, the engine has stopped being the moat.
This is not the same claim as choosing the right model per task. Routing assumes the model still carries the advantage and the skill is selecting it well. The shift here is one level up. The models are converging on capability, the price of renting them is falling, and the gap between the best and the second-best keeps narrowing for ordinary work. When the inputs become interchangeable, the thing that differentiates a product is no longer which input you picked. It is what you composed on top of it. The model becomes substrate, the same way electricity or a database became substrate, essential and assumed and no longer a story you tell customers.
Rent the model. Own the design. When everyone can call the same API, the experience drawn around it is the only thing a competitor cannot buy.
So the advantage moves to the designed layer, which is the layer most teams underinvest in. The workflow that turns a capable model into a task someone trusts. The trust surface that tells a user what the system did and lets them correct it. The judgment about what the product should refuse to do. The data and context only you can supply. The interface that makes a generic capability feel native to one specific job. None of these come from the model. All of them come from design decisions, and all of them are hard for a competitor to copy precisely because they are not for sale by an API.
The practical consequence for a business is a change in where the effort goes. The cheap, replaceable half of the build is the model integration; assume from the start that you will swap providers, and design the seam so swapping is a configuration change, not a rebuild. The expensive, compounding half is the experience around it, and that is where the budget belongs. A team that spends its quarter chasing the newest model and bolts a thin wrapper around it ships something a competitor reproduces in a week. A team that assumes model parity and spends the same quarter on workflow, trust, and fit ships something that stays ahead even after the competitor rents the identical model. The model is a line item. The design is the asset.
The pattern is the one this journal keeps finding. When AI design tools learned to generate, the systems that held up were the ones where the design system stayed the ground the tool deferred to, not a suggestion it overrode. When agents became useful, the unit of architecture turned out to be the contract around the agent, not the raw capability inside it. The rented model is the same lesson at the level of the whole product. The intelligence is borrowed and getting cheaper. The architecture composed around it is the part you own. Apple just demonstrated the move at the largest possible scale, and the businesses paying attention will stop buying the model as a strategy and start designing the only thing a competitor cannot rent.
Field note prompted by Apple's WWDC 2026 keynote on June 8: a rebuilt Siri running on a licensed custom Google Gemini model, a small on-device model with private compute for escalations, and an Extensions system that lets users select Claude or ChatGPT behind Apple Intelligence. Reporting via MacRumors. The principle outlasts any single keynote: the model is rented, the design is owned.