Apple announces Siri AI with on-device and server inference at iPhone event

On Monday, September 14, 2026, Apple released its 2027 software releases — iOS 27, iPadOS 27, macOS 27, watchOS 27, visionOS 27 and tvOS 27 — led by the next generation of Apple Intelligence, whose flagship is Siri AI, described by Apple as "an entirely new version of Siri" and "a profoundly more capable and conversational assistant with personal context understanding, broad world knowledge, onscreen awareness, and even more systemwide app actions."

Key rollout facts (FACT — release-day newsroom and support pages, independently corroborated):

• Siri AI began rolling out today in beta in English, with French, Japanese, Korean, Portuguese and Spanish "coming next month" (October 2026); it requires an Apple Intelligence-enabled device set to a supported language, and is not available to users under 13.
• The rollout is waitlist-gated: per Apple's support documentation, users go to Settings → Siri → "Try Siri AI (Beta)" to join a waitlist, and Siri AI downloads automatically once the wait ends. The Next Web independently confirmed the waitlist mechanism and that wait times "can vary."
• A dedicated Siri app was added that syncs conversational history privately over iCloud across iPhone, iPad, Mac, Apple Watch and Apple Vision Pro.
• New capabilities: personal-context search across Messages, Mail, Photos and more; onscreen awareness (asking about and acting on what is on screen); broad world knowledge pulling up-to-date answers from the web; expanded systemwide app actions (draft an email, edit and share photos, conversational follow-ups); Visual Intelligence extended to iPhone Camera ("Siri mode"), iPad screenshot experience, Mac keyboard shortcut and Vision Pro; Write with Siri writing tools systemwide; expressive customizable voices and improved dictation powered by the most advanced on-device model.
• Third-party app actions: Siri AI works with WhatsApp and Audible at launch; Microsoft Outlook, Notability and Tripsy integrations are coming.
• Architecture (Apple's description): the new capabilities are "powered by the next generation of Apple Foundation Models, custom-built in collaboration with Google and its Gemini models for deeply integrated Apple Intelligence experiences. These latest models run on device and on servers using Private Cloud Compute." A System Orchestrator coordinates requests against the on-device Spotlight Semantic Index and App Toolbox; "for some requests, models are capable of processing your Siri requests entirely locally on the device," and the orchestrator sends more sophisticated requests to models running in Private Cloud Compute. Apple restates its privacy promise: "When Private Cloud Compute is handling users' requests, their personal data is not stored nor made accessible to Apple or anyone else. Outside experts can continue to verify this privacy promise at any time."
• Daily usage limits were introduced on server-side-model features ("including but not limited to Siri AI, Intelligent Photo Editing Tools, Image Playground, and AFM 3 Cloud models in Shortcuts"), with limits varying by feature, request complexity and system demand — and Apple disclosed that "expanded access to such features will be available for a fee in the future."
• Regional carve-outs: Siri AI will not be available initially in the EU on iOS, iPadOS or watchOS (DMA dispute, see §14), but is available to EU users on macOS 27 and visionOS 27 in a supported language; and Siri AI and other new Apple Intelligence features will not be available in China while Apple works through regulatory requirements.
• Device eligibility: Apple Intelligence and Siri AI run on iPhone 16 models or later, iPhone 15 Pro / Pro Max, iPad mini (A17 Pro), iPads with M1 or later, MacBook Neo (A18 Pro), Macs with M1 or later, Apple Vision Pro, Apple Watch Series 9 or later / Ultra 2 or later / SE 3 (paired with a supported iPhone). The most advanced on-device model features (expressive-vs-standard voices, advanced dictation) are limited to iPhone 18 Pro/Pro Max, iPhone Duo, iPhone 17 Pro/Pro Max, iPhone Air, iPads with M4 + 12GB, Macs with M3 + 12GB, and Vision Pro with M5 — Apple's silicon-tiered inference story in one list.
• The release landed the same week as the hardware wave it showcases: iPhone 18 Pro/Pro Max pre-orders opened September 12 and ship September 18; the iPhone Duo launches October 23; Apple Watch Series 12 / Ultra 4 and AirPods 5 go on sale September 18.

The release completed a journey that began at WWDC26 (June 8, 2026), when Siri AI was unveiled with the same architecture described above; a January 12, 2026 joint Google statement had already said the next Apple Foundation Models would be based on Gemini models and cloud technology; and Bloomberg's Mark Gurman had reported the Siri chatbot reboot (code-named "Campos," January 2026) and the November-style opening of Siri to rival assistants (March 2026).

• Siri stopped being a voice command shell and became a full conversational assistant ("Siri AI"), with personal-context retrieval across Messages/Mail/Photos, onscreen awareness, web-grounded answers and cross-app actions, plus a dedicated chat app with iCloud-synced history — positioning it directly against ChatGPT, Gemini and Claude as an OS-level assistant.
• Apple's AI architecture changed definitively from "mostly on-device + Apple-only cloud" to a hybrid with a Google component: the third-generation Apple Foundation Models (AFM) family is "custom-built in collaboration with Google and its Gemini models," with the top server model (AFM Cloud Pro) reportedly running on NVIDIA GPUs in Google Cloud wrapped in Private Cloud Compute. This is the first time Apple has publicly anchored frontier-grade inference partly in a partner's cloud.
• The privacy architecture gained a new, more complex promise: Apple extended its Private Cloud Compute claim ("not stored nor made accessible to Apple or anyone else, verifiable by outside experts") to a deployment that includes NVIDIA hardware inside Google Cloud, with a reported contractual bar on Google training on Apple user data — a company claim that no independent audit has yet verified (INDEPENDENT EVIDENCE of the claim's audited status is absent).
• Apple introduced metering for the first time on core assistant features: daily usage limits on server-side-model features, plus an explicit statement that expanded access "will be available for a fee in the future" — the first public confirmation of a paid tier for Apple Intelligence features (no price or date published).
• The assistant market's default changed: the world's largest installed consumer base (billions of active devices, iPhone 15 Pro-era hardware and newer) gained an LLM-grade assistant inside the OS, not bolted on — a scale shift in who uses conversational AI daily.
• In-window change (Sep 10–17): public release (Sep 14) and the first full week of real-world usage/feedback, immediately followed by hands-on reviews and the first wave of independent assessments; EU availability of iOS 27 shipped without Siri AI; Apple Watch Series 12/Ultra 4 + iPhone 18 hardware events bracketing the release.

• Before: Siri was a command-and-query assistant (timers, calls, media, simple questions) that had fallen years behind LLM chatbots; Apple Intelligence (2024/2025) had delivered only a delayed, modest rewrite; Apple's only frontier-adjacent capability was a ChatGPT integration that handed off queries it couldn't answer. Apple's AI runs were on-device or on Apple's own servers; the company was publicly criticized for missing the LLM assistant era ("Apple's AI crisis" — Gurman, June 2026).
• Change: At WWDC26 and then in the September releases, Apple rebuilt Siri from the ground up on the third-gen AFM family (Gemini-derived, Apple-customized), introduced the System Orchestrator routing between on-device models, Private Cloud Compute on Apple silicon, and — for the heaviest reasoning — NVIDIA GPUs in Google Cloud; shipped a dedicated Siri app, personal context, onscreen awareness, Visual Intelligence and Writing Tools; added waitlist beta rollout, usage limits and a future paid tier; and split regional availability (EU iPhone/iPad/watch excluded over DMA; China held pending regulation).
• After: An LLM-grade, context-aware assistant ships on-device-and-in-cloud privacy architecture to hundreds of millions of supported devices in English beta, with five more languages in October; the default assistant experience on iPhone/iPad/Mac/Watch/Vision Pro is now a genuine ChatGPT-class competitor; Apple has committed publicly to "frontier-level intelligence" delivered through a verifiable private-cloud model — the definitional stake in the private-inference versus cloud-AI debate; and the competitive landscape (OpenAI ChatGPT, Google Gemini, Anthropic Claude, xAI Grok) gained a new distribution channel at unprecedented consumer scale.

The model family (FACT for Apple's own description; COMPANY CLAIM for capability levels and quality equivalences): Apple's machine-learning group described a family of five third-generation Apple Foundation Models at WWDC26 (details consolidated from Apple's newsroom, Federighi's post-keynote tech talk as reported by 9to5Mac, and The Next Web/The Decoder/otontechnology breakdowns):

Apple states the on-device and Apple-silicon models are "custom built for Apple Silicon, trained using proprietary data and refined using outputs from Gemini frontier models"; Apple does not ship the Gemini app, does not run Google's client code, and does not use Google Search as the knowledge base (Federighi tech talk; INTERPRETATION: Apple uses Google as distillation target and cloud substrate, not as the user-facing product). The Next Web reported the Google deal at roughly $1B per year (COMPANY CLAIM/unverified — not confirmed by either company).

The System Orchestrator (Apple's description, corroborated by Federighi's tech talk): the orchestrator is the routing and privacy core. It coordinates requests against:

• the Spotlight Semantic Index (on-device index of personal content),
• the App Toolbox (access to actions within apps),
• on-screen context (what the user is currently looking at),
• and decides when a request is answered by an on-device model ("entirely locally") vs. escalated to Private Cloud Compute ("when it realizes it's a more sophisticated question").

Private Cloud Compute (Apple's published claim): PCC "extends the privacy and security of iPhone into the cloud"; requests are "never stored," "never accessible to anyone, including Apple," processed only as part of the request, and "third-party researchers can continuously verify" these properties. For AFM Cloud Pro specifically, Apple, Google and NVIDIA extended the same PCC wrapper onto NVIDIA GPUs in Google Cloud, using NVIDIA confidential computing to encrypt data in process (INDEPENDENT EVIDENCE — reported by The Next Web and The Decoder from the WWDC26 tech talk). No independent audit of the Google-Cloud tier has been published (FACT — absence of evidence).

Rollout mechanics (FACT): Siri AI is a beta, opt-in via a waitlist (Settings → Siri → Try Siri AI); conversation history syncs via iCloud in the dedicated Siri app; server-side features are metered with unpublished daily caps that "may vary by feature, request complexity, system demand"; paid expanded access is confirmed to come later; users under 13 are excluded; the assistant is English-only at launch with six languages supported by October.

• Scale: Siri AI puts an LLM-grade, context-aware assistant in the operating system of the world's largest consumer hardware base. No other assistant runs at this installed-device scale with OS-level personal-context access — this is the biggest single distribution event for conversational AI since ChatGPT.
• The private-inference stance becomes a product: Apple is the only major player arguing that frontier-grade assistant AI can be delivered without the provider (or a third party) storing or seeing personal data. PCC's "verifiable privacy" claim — extended onto NVIDIA hardware inside Google's cloud — is now the industry's reference architecture for "frontier intelligence without the data grab," and a direct counter to the cloud-first model of OpenAI/Google (the-decoder: Apple's position as "counterpoint" to labs "racing forward without regard for people").
• It legitimizes the hybrid model-sourcing strategy: Apple outsourced model quality (Gemini-derived AFM) and part of its inference substrate (NVIDIA in Google Cloud) while owning the OS, the orchestrator, the on-device index and the privacy claims — evidence that at frontier scale, "own the experience, license the model, verify the cloud" can ship. Every other OEM now has a template for catching up without building a frontier lab.
• Competitive re-alignment of the assistant market: ChatGPT, Gemini and Claude face a first-party, OS-integrated rival with personal-context permissions no third party can match (Six Colors: to get the same results from a competing chatbot "you'd need to hand it access to all your data"). Gurman's verdict — Siri AI is "just good enough to ease Apple's AI crisis" — frames the strategic stakes: Apple had to ship and it shipped credibly.
• The EU fight is a defining governance moment of the quarter: Apple withheld its flagship AI feature from iPhone/iPad/Watch in the EU over the DMA, publicly blamed the Commission, and the Commission publicly blamed Apple — a high-stakes test of whether DMA interoperability obligations and Apple's privacy architecture can coexist (Reuters/AP; see §14).

• Conversational AI with OS-level personal context at consumer scale: users can ask questions over their own messages, email, photos and calendar ("what did my family member suggest for the visit?") and get grounded answers plus actions — previously impossible for any assistant without bulk data-sharing to a third party.
• Onscreen- and camera-grounded action: asking about what's on screen or in the camera viewfinder and having Siri act (add calendar events, split a bill with Apple Cash in the U.S., add a pass to Wallet, brew instructions from a coffee-bag photo) became native on iPhone/iPad/Mac/Vision Pro.
• A private-cloud "frontier" tier for consumers: for the first time, consumers get access to a top-tier reasoning model (AFM Cloud Pro, agentic tool use) under a published, auditor-verifiable no-retention promise rather than an opaque cloud backend.
• A metered, monetizable Apple AI service: usage limits plus a confirmed future paid tier create the infrastructure for Apple Intelligence monetization (the first explicit step beyond hardware-embedded AI).
• Cross-platform conversational continuity: a dedicated Siri app with iCloud-synced history makes chat-with-Siri a persistent, multi-device activity (start on iPhone, continue on Mac).
• Developer-delivered assistant experiences: the Foundation Models framework and App Intents entities/intents let third-party apps plug into Siri AI; the small-business program (fewer than 2M first-time downloads) gets Apple's cloud models on Private Cloud Compute at no cloud API cost — a genuine indie-agent channel.

• What are the actual daily usage limits, and how will the paid tier be priced? Apple has published neither numbers nor a date; "including but not limited to" leaves the feature list open.
• Can independent researchers verify the PCC no-retention claim for the Google-Cloud/NVIDIA tier — and will Apple publish statutory transparency/audit material for it comparable to its original PCC program materials?
• What exactly did Apple propose ("Trusted System Agent," 18-month rollout) and what did the Commission reject? Neither side has published the technical details; Reuters/AP confirm only the dispute's public shape. Does EU iPhone Siri AI ever ship — and what does the resolution look like (DMA interop obligations vs. Apple's architecture)?
• When do the October languages arrive, and does the EU Mac/Vision Pro version demonstrate the feature set for the EU iPhone version later?
• Will third-party depth arrive? The App Intents entities/intents work (Outlook, Notability, Tripsy; Telegram class absent) determines whether Siri AI is genuinely systemic or Apple-apps-only in practice.
• Are the private Model Delegation / Inference Providing hooks (pdfu's Claude/GPT-5.6-Terra-as-Siri demos) officially supported, and do they imply a future "bring your own model" Siri? (Bloomberg's Gurman reported Apple plans to open Siri to rival assistants beyond ChatGPT — consistent with the hooks.) Unverified as of window close.
• Does the Google deal (reported ~$1B/year, no-training clause) withstand renegotiation, and what happens to AFM if it does not?
• How does on-device AFM Core Advanced behave at scale (battery, thermals, sparse-router efficiency on A20 Pro/dual 16-core Neural Engine vs iPhone 15 Pro)?
• Will the China regulatory path clear, and what changes are demanded?

• October 2026: French, Japanese, Korean, Portuguese and Spanish Siri AI support ("next month"); watch whether EU Mac/Vision Pro users get them simultaneously, and whether the language expansion signals EU iPhone progress.
• Later this year: Audio Intelligence betas on Apple Watch Series 12/Ultra 4 (Live Rewind, Siri Recap, Sound Recognition, Secure Exclave on S11) and the redesigned Health app with Apple Intelligence Insights arrive; SynthID watermark support for AI-generated images comes to Apple Intelligence tools; watchOS 27 Siri AI developer testing.
• Hardware wave lands: iPhone 18 Pro/Pro Max ship September 18; iPhone Duo (foldable, A20 Pro, premier on-device AI) pre-orders October 16, launch October 23 — the first full fleet running AFM Core Advanced.
• Watch items (3–8 weeks): first independent privacy-audit attempts of the PCC Google/NVIDIA tier; publication (or silence) on usage-limit numbers and paid-tier pricing; developer SDK maturity surfacing third-party App Intents depth (Outlook, Notability, Tripsy); pdfu-style Model Delegation hooks either becoming official or being patched out; EU–Apple DMA engagement updates; first reports of assistant-injection incidents against App Toolbox actions.
• Structural (quarter+): Siri AI becomes the default LLM assistant for the Apple installed base; expect OpenAI/Google to respond on distribution (deeper OS partnerships) and Apple to fold paid-AI-tier into services reporting; the DMA dispute outcome shapes EU assistant regulation; China remains gated by regulatory requirements.

The story of "Apple ships Siri AI" is not really about a voice assistant catching up with chatbots — it is about the two bets Apple is now making in public. First, that the industry's frontier race can be entered without being run: license the frontier from Google, re-optimize it as the AFM family, and compete on distribution, OS integration and privacy instead of benchmark scores. Second, that consumers will choose "frontier intelligence without retention" when offered honestly — a direct challenge to the data-for-AI bargain every cloud-first assistant rests on, and a claim whose credibility now depends entirely on independent verification of a cloud that runs, in part, on NVIDIA hardware inside Apple's biggest rival's datacenter. The EU missed the launch, China is gated, the beta is waitlisted, the caps are unpublished and the paid tier looms — but for the first time in years, the default iPhone assistant is genuinely, verifiably useful. That is why this release — not the foldable phone, not the new silicon — is the week's defining Apple story, and why the verification, the pricing and the EU fight will matter more than the demo videos.