# 蘋果使用 Google 的 Gemini 模型來訓練可在 iPhone、iPad 和 Mac 本地運行的較小蒸餾模型。

*genai, semiconductor · news · 2026-05-30 · Gizchina.com*

## Key points

- 蘋果使用 Google 的 Gemini 模型輸出來訓練較小的本地 AI 模型以支援 Siri。
- 複雜的 Siri 查詢在 Google Cloud 上使用授權的 Gemini 模型處理，而非蘋果自有基礎設施。
- Nvidia 的機密運算技術加密在 Google Cloud 處理的資料，確保雲端推論期間的用戶隱私。
- 儘管部分 Siri 處理發生在非蘋果自有伺服器上，蘋果仍將使用 Private Cloud Compute 品牌。
- 蘋果可能收購較小的 AI 公司，以降低對 Google 的依賴並提升本地 AI 能力。

The Information's Aaron Tilley broke the technical architecture behind Apple's biggest Siri overhaul in years — and it's more complex than Apple's usual privacy-first messaging suggests. Gemini is being used to train smaller on-device models through distillation. Complex Siri queries route to Google Cloud running a licensed Gemini model. Nvidia's confidential computing encrypts data during cloud processing. WWDC on June 8 is where Apple reveals the features built on top of this infrastructure. Key Points Apple is using Google's Gemini model to train smaller distilled models that run locally on iPhones, iPads, and Macs — a process that keeps basic AI tasks fast and private on-device Complex Siri queries route to Google Cloud running a licensed Gemini model — Apple's own Private Cloud Compute infrastructure can't efficiently host the full trillion-parameter Gemini model Apple recently approved Nvidia's Confidential Computing technology — encrypting user data and AI models while being processed inside Nvidia GPUs on Google Cloud servers Apple will continue using "Private Cloud Compute" branding even though some processing runs on Google Cloud infrastructure — a framing gap that critics have noted Full Gemini-powered Siri reveal is expected at WWDC June 8, with broader rollout alongside iOS 27 in September What Distillation Actually Means Apple isn't putting the full Gemini model on your iPhone. It can't — a trillion-parameter frontier model requires data center-scale hardware. Instead, Apple uses Gemini's outputs to train a much smaller, more efficient model that captures Gemini's reasoning capabilities in a form factor that fits on-device memory. The result is a local model that handles everyday Siri requests — timers, text editing, basic questions — with no cloud connection required. Only complex queries that exceed the local model's capabilities get routed outward. That architecture is genuinely privacy-preserving for routine interactions, even if the cloud component complicates the narrative. Why Google Cloud — Not Just Apple's Own Servers Apple's Private Cloud Compute runs on M-series Mac chips in Apple-owned data centers. Ars Technica reported that this infrastructure has struggled to host even a distilled version of Gemini efficiently. The compute demands of frontier AI inference at scale require GPU clusters that Apple simply doesn't have at sufficient scale yet. Routing through Google Cloud fills that gap — while the Nvidia Confidential Computing approval gives Apple a privacy story for those external queries. The system encrypts data and AI models while Nvidia GPUs process them, meaning Google's servers handle the request without accessing identifiable user data. Google is contractually barred from using iOS query streams to train its own models. The privacy trade-off is real but mitigated. The speed hit from confidential computing — which slightly slows cloud AI processing — is the cost Apple pays for maintaining its privacy claims even when using external infrastructure. WWDC June 8 — What to Expect Bloomberg confirmed a redesigned Siri interface debuts at WWDC alongside new Apple Intelligence features. The Gemini-powered backend is the engine. The front-end changes — a new visual design dropping the legacy glow overlay in favour of tighter hardware integration — are what users will actually see. September's iOS 27 rollout is the broader deployment target. Apple is also reportedly exploring acquiring smaller AI companies to accelerate the on-device model compression work — suggesting the Google dependency is a bridge strategy while Apple builds longer-term capability.

**Companies:** Apple, Google, Nvidia
**Countries:** United States

[Read the full story on Gizchina.com](https://www.gizchina.com/apple/apples-siri-overhaul-for-ios-27-relies-on-google-gemini-google-cloud-and-nvidia-chips)

---

Canonical: https://newsio.io/zh-TW/n/ad605e20-74a0-40ff-8d6b-f727e99c294c/google-gemini-iphoneipad-mac-private-cloud-compute-gemini
Summarized by Newsio from Gizchina.com. https://newsio.io/how-it-works