Argues that Apple's choice to license Gemini is a tacit admission it cannot win the base-model race, but that this doesn't matter because Apple's actual product is everything around the model — on-device routing, server-side orchestration, and especially Private Cloud Compute as a model-agnostic attested substrate. The interesting artifact is that PCC, originally designed for Apple's own weights, now hosts Google-derived weights under the same verifiable privacy guarantees.
Frames the announcement as a major architectural overhaul where Apple's in-house foundation model effort has effectively been replaced by Gemini-derived weights co-developed with Google. The submission's high score (523 points) reflects community read that this is a notable retreat from Apple's year-long push to ship its own models after Siri's contextual upgrades slipped.
Argues cynically that Apple is handicapping itself versus Android by adopting Android's own model vendor as the brain of Apple Intelligence — if both platforms run on Gemini, Apple loses a differentiation lever. The editorial flags this view as 'sharp but probably backwards' since Apple isn't competing on the model itself.
Notes that Apple already pays Google roughly $20B/year for default search placement, so bolting Gemini into Apple Intelligence deepens an existing relationship rather than creating a new dependency. It also conveniently avoids the OpenAI–Microsoft entanglement and sidesteps Anthropic's smaller deployment footprint, making Google the path of least commercial resistance.
Apple today announced a rebuild of Apple Intelligence around a new family of "Apple Foundation Models" that, per Apple's own framing, were co-developed with Google using the technologies behind Gemini. The models are adapted to run in two tiers: a smaller on-device variant on Apple Silicon, and a larger server variant inside Apple's existing Private Cloud Compute (PCC) enclaves. Apple reiterated that user data is used only to execute the immediate request, is not retained by Apple or third parties, and that outside researchers can verify those guarantees against the PCC images.
The announcement is striking less for what shipped than for what it concedes. Apple spent a year shipping a half-finished Apple Intelligence stack built on its own in-house foundation models. Siri's promised contextual upgrades slipped. The on-device 3B model was competent but unremarkable. Today's announcement is Apple admitting, in product form, that it is not going to win the base-model race — and that its actual product is the routing layer.
The choice of Google over Anthropic or OpenAI is the part that's drawing the most discussion. Apple already pays Google roughly $20B/year for default search placement; bolting Gemini into Apple Intelligence deepens an existing commercial relationship rather than creating a new one. It also conveniently sidesteps the OpenAI–Microsoft entanglement and Anthropic's smaller deployment footprint. The cynical read on Hacker News from user `dejawu` — that Apple is disadvantaging itself versus Android by picking Android's own model vendor — is sharp but probably backwards. Apple isn't selling the model. It's selling everything around it.
The interesting technical artifact here is Private Cloud Compute as a model-agnostic substrate. PCC was designed in 2024 as a way to run Apple's own models in attested enclaves where neither Apple nor any operator could inspect the request. The fact that Google-derived weights can now run inside that same envelope, with the same external verifiability claims, validates the architecture as something more general than its original framing. It's a confidential-compute pattern: bring any model, run it under a measured boot, publish the image hash, let researchers audit. The model becomes a swappable component; the trust boundary becomes the product.
Luk212's comment on HN crystallizes the strategy: "Very Apple-ish approach to AI catch up: wrap an external tool in a privacy architecture, embed into the OS and productize the orchestration layer." That is exactly what's happening, and it is not a small thing. Routing — deciding what runs on a 3B local model vs. a 70B+ PCC model vs. punted to a third party — is a real engineering problem. Latency budgets, battery, intent classification, fallback semantics, and PII redaction all live in that routing layer. None of it is glamorous. All of it compounds.
The community skepticism (bensyverson asking what "co-developed" actually means; noobcoder calling it warmed-over Google Assistant) is fair but misses the time horizon. Apple has now decoupled its product roadmap from its model roadmap. If Gemini 3 lands in nine months and crushes benchmarks, Apple can swap weights inside PCC without changing a single user-facing API. If Anthropic ships something dramatically better in 2027, Apple can route there. The orchestration layer is the durable asset; the model underneath is now a commodity input the way S3 is a commodity input to your app.
The strategic cost is real, though. Apple just installed a competitor's model as the default cognitive layer on a billion devices, and bound its own AI product velocity to that competitor's release schedule. Google now has a structural read on what features Apple wants to ship next (because Apple has to coordinate fine-tunes and adaptations), and a structural seat at the table on Apple Silicon's most strategic surface. Tim Cook's bet is that the privacy moat and the OS integration are wider than the model moat. That bet is defensible. It is also a bet.
If you're building on Apple platforms, the actionable surface is not "which model is this." It's the App Intents + Foundation Models framework boundary, where your app declares capabilities and the OS decides where the inference runs. Three concrete implications:
First, design your prompts and tool schemas to be model-agnostic. Apple has now publicly committed to a substitution architecture; assume your inference target will change at least once before iOS 27. Don't depend on quirks of any specific generation — test against the on-device variant *and* the PCC variant, because they will diverge in capability and your fallback paths matter.
Second, take the routing seriously. Apple's framework will give you knobs (or at least heuristics) for when to force on-device, when to allow PCC, and when to call out to a third-party model. That decision is now a product decision with privacy, latency, and cost implications. If you punt it to defaults, you're letting Apple's router make UX choices for you. The teams that win on Apple Intelligence will be the ones who treat routing as a first-class part of their feature design, not an afterthought.
Third, the privacy verifiability story has actual leverage now. If you're in a regulated vertical — health, finance, legal — PCC's attestation model gives you a defensible answer to "where does this prompt go." It's the first time the answer can credibly be "into an enclave whose image you can inspect." That's a meaningfully different posture than calling the OpenAI API from your backend. Whether your compliance team will accept it is the next question, but the architecture is now on the menu.
The interesting question for the next twelve months is whether Apple opens PCC to third-party model providers — letting a developer ship, say, a Claude-powered feature that runs in Apple's attested enclaves rather than Anthropic's. If that happens, Apple has built the iOS equivalent of a confidential-compute marketplace, and the Google deal is just the anchor tenant. If it doesn't, then Apple Intelligence remains a Gemini wrapper with very good packaging, and the long-term question is whether wrapping is enough when the underlying model is someone else's product. Either way, the era of Apple pretending it would build its own frontier model is over. The era of Apple selling the trust boundary has begun.
I would love to learn more about what's actually powering Apple Intelligence now. Are they using flagship Gemini models behind their own prompts? Fine-tuning? Pre-training their own models based on Gemini?Is there a meaningful distinction between the Gemini-powered models and Apple Foundation M
If I can't even trust the results given by ChatGPT and Claude at their highest level of reasoning in my daily life and work, would I be willing to use Siri AI to handle the important scenarios depicted in the livestream?
It's strange to me that Apple would choose to disadvantage themselves by selecting Google as their provider as opposed to, say, Anthropic or even OpenAI. Doesn't this mean they'll struggle more to differentiate themselves from the assistant on Android phones? Thinking more cynically,
Sort of expected this at the first attempt. Use their existing partnership for Google being the default search with Google and leverage Google's models.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
Very Apple-ish approach to AI catch up: wrap an external tool in a privacy architecture, embed into the OS and productize the orchestration layer.It will be interesting to see if the Private Cloud Compute + on-device routing can make third-party model capabilities feel like a first-party system with