Apple spent approximately one billion dollars a year licensing Gemini to power Siri. That number is not a secret: it has been reported publicly and Apple confirmed the relationship at WWDC. The brain running inside Apple's flagship AI product is a rented commodity from a competitor.
This is not a criticism of Apple. It is a structural observation about where the durable advantage in personal AI actually lives.
Why Apple rented rather than built
Building frontier reasoning capability requires sustained investment at a scale that does not make strategic sense for Apple. Google, Anthropic, and OpenAI have spent years and billions of dollars on this. Apple has the integration story, the hardware, the distribution, and the privacy architecture. It does not need to win the model race. It needs to win the product race.
The Extensions Framework Apple shipped at WWDC is the logical conclusion of this: the AI brain is user-swappable. Gemini by default. Claude and ChatGPT as user choices. One active at a time. Apple built the socket; the model is a plug.
The brain is swappable. The history is not.
When you switch from Gemini to Claude inside Siri, the reasoning engine changes. The context it operates over does not. Your 20 years of email, your iMessage history, your resolved understanding of who Alex is across every thread, your AI conversation history from the last two years: none of that travels with the brain swap. It stays with Apple.
What the corpus actually is
The corpus is not a database of messages. It is a resolved, calibrated, accumulating record of your life's interactions.
Every email you have sent or received is a data point. But a data point is not knowledge. The gap between raw messages and useful context requires four things that no reasoning engine alone can provide: deduplication (resolving that Alex, A. Rivera, and the person in your iMessage threads are the same individual), entity resolution (knowing which Jordan Chen you mean when there are forty emails and a widely-indexed public figure with the same name), inference calibration (scoring what the corpus implies about a relationship versus what it confirms), and longitudinal accumulation (weighting the full arc of a relationship, not just the most recent context).
A corpus that has done this work for two years is not a two-year-old database. It is a knowledge graph with two years of resolution and calibration applied. The signal-to-noise ratio has improved with every source added and every inference corrected. It compounds in a way that a fresh installation does not replicate.
The corpus moat is structural, not just temporal
It is tempting to describe the corpus advantage as a head-start: you have been building yours for two years; a new entrant starts from zero. That is true but incomplete.
The more important point is structural. A reasoning engine can be retrained or replaced in weeks. The same is not true of a corpus built from your specific correspondence, your specific entity graph, and your specific AI conversation history. There is no corpus that generalises to your life the way yours does. The Gemini model Apple licensed is the same Gemini model every Gemini user gets. Your corpus of 20-year relationships, resolved and calibrated across six communication sources, is not available in any model and cannot be rebuilt from scratch.
This is not a lock-in argument. You should own your corpus. It should be portable. The argument is not that the corpus is hard to move; it is that the corpus is the thing worth building, and no amount of reasoning capability substitutes for it when you do not have it.
What this means for which product you build on
If the brain is a commodity input and the corpus is the durable advantage, the product question is: who holds the corpus, and who builds the infrastructure around it?
Apple holds the Apple corpus. It does not hold your Google conversation history, your Anthropic session history, or your 20 years of email from before iCloud existed. Google holds the Google corpus. It does not hold your iMessage history or your Claude sessions. Anthropic holds your Claude corpus. It does not hold any of the rest.
The cross-provider corpus is the one no individual AI provider can build. Every email, every message, every AI conversation you have ever had, across every walled garden, resolved into a single entity graph, held on a server you control. That is what compounds when you add a new source. That is what persists when you switch AI providers. That is the thing WWDC proved matters, because WWDC proved the brain does not.
Mandaire is the cross-provider corpus layer. It ingests every source, resolves the entity graph, and connects to whichever AI brain you bring. The brain is your choice. The corpus is yours.
Request early access