Jordan Hale: You know what keeps nagging at me — Tim Cook's last big swing, and it's Siri. Of all things.
Ryan Castillo: Farewell keynote, June 8th, Apple Park. And the centerpiece is a three-tier AI routing architecture that ends — ends — on Google Cloud infrastructure.
Jordan Hale: Wait — like, that's literally where the queries land? Google's servers?
Ryan Castillo: Tier 3, yeah. The most demanding Siri AI requests hit a 1.2-trillion-parameter Gemini-derived model running on Nvidia Blackwell GPUs in Google's data centers. Apple Intelligence is the umbrella brand, Siri AI is the product, and the whole thing costs Apple a reported billion dollars a year.
Jordan Hale: A billion a year. And then — okay, this is the part that, like, I can't get past — Federighi's out there saying Apple uses none of Google's models. Publicly. After signing a billion-dollar deal.
Ryan Castillo: That's the load-bearing claim. Does it hold? That's what we need to figure out.
Jordan Hale: Okay but wait — because the headline is 'Apple is using Google's AI' and that's... I mean, that's not wrong, but it's also not really the specific weird thing that's happening. Like, what Macworld actually confirmed is five third-generation models — two on-device, three cloud-based — and the Tier 3 cloud model used Gemini foundation weights and then got retrained on Apple-proprietary data. That's a different product than Gemini.
Ryan Castillo: Right, but — what's the actual delta between 'trained on Gemini weights' and 'is Gemini'? That's Federighi's distinction. Does it survive scrutiny?
Jordan Hale: The 1.2 trillion parameters actually blew my mind — Apple's internal Apple Foundation Models peaked at around 150 billion. That's eight times smaller. So the build-versus-buy math was just... done.
Ryan Castillo: That's the ceiling number. Yeah.
Jordan Hale: And — okay, this is the part I want you to react to — the 1.2 trillion model runs mixture-of-experts architecture. MoE. Meaning Apple is not activating all 1.2 trillion parameters per query. Only the relevant slices fire. Which means the cost picture and the latency picture are actually... smarter than the headline suggests?
Ryan Castillo: That changes the margin math significantly. But it doesn't change the Switzerland problem — users can set Claude or ChatGPT as default at the UI layer, but Tier 3 routes through Google Cloud regardless. That's not neutral. That's a choice baked in below the choice users think they're making.
Ryan Castillo: The reassuring read — and I want to name it explicitly — is that Private Cloud Compute solves this. That's what's circulating. I don't think that holds.
Jordan Hale: Yeah, and like — that's exactly what the X sentiment was after WWDC. Neutral-to-reassuring. Nobody really pushed back that week.
Ryan Castillo: Which is the problem. Because LavX News ran a cryptographer analysis on June 14th — six days post-WWDC — and the argument is that PCC handles data in transit and at rest fine. That's not the exposure. The exposure is data being *acted on*. An AI agent executing a task inside someone else's hardware is a different threat model entirely.
Jordan Hale: Okay, make it concrete for me.
Ryan Castillo: Lawyer. Tuesday morning. She asks Siri AI to summarize case precedents and flag contradictions. That query — the legal domain, the content, the fact pattern — that's Tier 3. It lands on Google Cloud hardware. She cannot opt out even if Claude is set as her default.
Jordan Hale: Wait — even with Claude set as default? That's — I mean, that's the Switzerland thing completely breaking down, right? Like, the user made a choice and it... didn't actually apply to the layer that matters.
Ryan Castillo: Exactly. You can't be neutral when Google is already in the infrastructure basement. Federighi's 'none of the models that Google built' framing — technically he's pointing at the retraining. But that nuance is doing load-bearing work that nobody reading a headline will ever see.
Jordan Hale: And that's — I mean, that's the thing that actually keeps me up about this. If users love Siri AI. If the privacy holds, if the queries feel fast and smart and nobody's data gets exposed — does that mean the Google dependency was just... the right call? Or does it mean Apple successfully buried it deep enough that nobody looked? Like, I genuinely don't know which of those is true.
Ryan Castillo: The number I'd watch is what happens to that billion-dollar-a-year deal when antitrust pressure on the Apple-Google relationship hits a breaking point. Or when a single high-profile breach makes the PCC-on-Google-Cloud story front page. Cook framed WWDC 2026 as the redemption arc after two years of underdelivering on Apple Intelligence. That architecture survives goodwill. Does it survive a forced unwind?
Jordan Hale: Yeah. And I don't think anyone has that answer yet. That's — that's actually the question, right? What happens to Apple's AI credibility if the thing holding it together is something Apple can't control?
