Sarah Lin: The word "zero" is doing a lot of heavy lifting in that WWDC keynote. Zero server dependencies. Zero per-token cloud costs. And I keep thinking — who is that zero for?
Dr. Nathan Hayes: For the developer. Core AI eliminates inference costs at the app layer — that's real. It's the successor to Core ML, purpose-built for transformers running on Apple Silicon. The zero is genuine in that narrow frame.
Sarah Lin: But not for the person holding the phone asking Siri something.
Dr. Nathan Hayes: No. Because Siri — which Apple also overhauled at WWDC 2026 — has a Google Gemini integration. Announced in the same keynote. So Apple Intelligence, the thing users touch, is not zero-server.
Sarah Lin: Your phone has its own personal chef, they said. Except — sort of, wait — they also signed a contract with DoorDash. For the meals the chef can't handle. And the food just appears. You never see the app open.
Dr. Nathan Hayes: And the decision about which kitchen your request goes to — that's Apple's. Not yours.
Dr. Nathan Hayes: Now here's where I need to plant a flag. The 70-billion-parameter claim is the one that matters most to me. Full precision, you're looking at roughly 140 gigabytes of model weights. Even at 4-bit quantization, that's still 70 gigabytes. That's a Pro Max configuration, maybe. That's not a standard iPhone. Apple has not published millisecond latency numbers for end-to-end generation at that scale. They showed it runs. That is not the same thing as showing it runs acceptably.
Sarah Lin: Okay but — wait, I'm not sure that's the right flag to plant.
Dr. Nathan Hayes: Why not?
Sarah Lin: Because Ollama already does this. On Apple Silicon. Right now, in the open. So the 70B-on-device thing isn't — um — that's not actually Apple's novel claim to defend. The thing that's real is that Qwen and Mistral ship as pre-converted Swift packages inside Core AI. A developer isn't quantizing anything. They're not managing a conversion pipeline. That's gone. And the discourse around that is developers replacing three to five thousand dollars a year in cloud API costs. That's not a rounding error.
Dr. Nathan Hayes: The cost number — I want to know whose costs. A solo developer with a hundred daily active users, or a startup running a hundred thousand daily inference requests? Effect size matters. Qwen and Mistral are real, I'm not disputing that. But the missing latency benchmarks — that's not a marketing footnote. That's the whole usability question.
Sarah Lin: Mm. But if OpenAI is the implicit comparison, and the cost is genuinely zero per token — I mean, that changes the decision even if the latency is slower. Doesn't it?
Sarah Lin: Actually — wait. I need to back up because there's something I keep almost saying. Private Cloud Compute. Apple's encrypted server tier from 2024. It's still running. Core AI doesn't replace it. So the actual architecture isn't on-device versus cloud — it's three layers of compute, and Apple decides which one you hit.
Dr. Nathan Hayes: That's correct. And it compounds.
Sarah Lin: The three-tier framework — Core ML, Core AI, MLX — that's not a menu. That's Apple's routing logic wearing a developer-friendly label. You as a user never see which tier caught your request.
Dr. Nathan Hayes: Right, and I mean, nobody has asked this clearly yet. The Gemini integration. When Siri routes to Google, does that request pass through Private Cloud Compute first, or does it go raw to Google's servers? Apple hasn't said. That is genuinely unresolved.
Sarah Lin: Wait — they haven't said at all?
Dr. Nathan Hayes: Not clarified. Which means the Foundation Models framework, App Intents expanding into cross-app personal data synthesis, Xcode 27's agentic workflows — all of that deeper OS integration is being layered onto an architecture where the Gemini routing tier is, mechanistically, unknown.
Sarah Lin: And that's — um — that's not a privacy footnote. That's the whole claim. The seam is invisible and we don't even know what's behind it.
Sarah Lin: And that's — I mean, that's the design story I keep almost landing on. Core AI makes the local-versus-cloud decision invisible. Privacy stops being a negotiation you consciously enter. It just becomes the air you breathe. And I think Apple knows that. I think that invisibility is the point.
Dr. Nathan Hayes: Vision Pro, iPhone, iPad, Mac — Core AI runs on all of them. Which of those can actually run 70B versus a smaller quantized tier? Apple hasn't published that. The routing rules are Apple's. The Gemini handoff conditions are Apple's. So the air you're breathing — whether it stays local or goes to Google's servers — that's not yours to measure.
