In Focus: Siri AI

Codelooru mythos

Information in this post reflects publicly available sources as of June 10, 2026.


For years, the running joke about Siri was that it could not reliably set a timer without mishearing you. While OpenAI shipped ChatGPT to 900 million weekly users and Google folded Gemini into every Android surface, Siri answered questions about the weather and occasionally called the wrong contact. The gap between what Apple's assistant could do and what the competition offered had become impossible to ignore.

On June 8, 2026, at WWDC in Cupertino, Apple stopped pretending the gap did not exist.


What Apple announced

Siri AI is a ground-up rebuild of Apple's voice assistant, announced alongside iOS 27, iPadOS 27, and macOS 27 Golden Gate. It ships this fall as a free software update and arrives on iPhone, iPad, Mac, Apple Watch, and Apple Vision Pro. A standalone Siri app is included. The new assistant is conversational, context-aware, and capable of multi-step tasks across apps.

The structural story underneath the product announcement is a deal. Apple signed a multi-year licensing arrangement with Google to run a custom version of Gemini for Siri's cloud intelligence. The model is reported at 1.2 trillion parameters, roughly eight times larger than the largest cloud model Apple had previously built in-house, at a cost of approximately $1 billion per year. Complex queries route to Nvidia Blackwell B200 GPUs on Google Cloud infrastructure.

Craig Federighi, Apple's SVP of Software Engineering, was unusually specific about the architecture at the WWDC developer session. Apple's operating system now includes a system orchestrator: a piece of software that sits between the user's request and the model that handles it. The orchestrator decides, per query, whether to handle the request entirely on-device, route it to Apple's own Private Cloud Compute servers, or pass it to Gemini. The routing decision is based on how much compute the query requires and how much personal context it needs.

For requests that stay on-device, Apple's own Foundation Models handle everything from speech recognition to on-screen context understanding. For requests that go to the cloud, Apple says personal data is not retained or made accessible to Google. Federighi described the system orchestrator as "key to the privacy architecture of our entire system."

User Request "Hey Siri..." System Orchestrator Routes by compute need and personal context On-Device Models Speech, vision, context No network needed Private Cloud Compute Apple servers Data not retained Google Gemini Complex queries Nvidia Blackwell / GCP Simple / personal Moderate complexity High complexity Siri AI routing architecture as described by Craig Federighi at WWDC 2026

The deal was confirmed publicly by Google Cloud CEO Thomas Kurian at Google Cloud Next '26 earlier this year. The partnership was first announced in January 2026; WWDC was the product reveal that put flesh on the announcement.


How it came to public attention

Coverage of the Apple-Google AI deal began months before WWDC, driven largely by Bloomberg's Mark Gurman, who reported the commercial terms and the Gemini integration well in advance of the keynote. By the time Federighi took the stage, most of the product details had already leaked. What the keynote added was the technical architecture, the developer implications, and the framing Apple chose for its own pivot.

That framing was notably pointed. Federighi described other companies as "seemingly pursuing AI for the sake of AI, without clear regard to the people it's ultimately meant to serve." The subtext was not subtle: Apple is positioning this as a considered, privacy-respecting deployment rather than a race to ship. Whether the audience read that as principled restraint or defensive spin depended on how charitably they viewed the last two years of Siri falling behind.

The announcement also carried personal significance. WWDC 2026 was Tim Cook's final keynote as CEO before handing over to John Ternus on September 1. Siri AI was presented as a capstone. Whether it reads as a legacy achievement or a problem deferred to the next administration is a question that will take at least a year to answer.


What it actually is

Technically, Siri AI is a hybrid inference system. Apple builds and operates the orchestration layer, the on-device models, and the Private Cloud Compute infrastructure. Google supplies the frontier model for the hard queries. The two systems do not share user data: the orchestrator is designed to strip personal context before anything reaches Gemini.

The 1.2-trillion-parameter Gemini model Apple licensed uses a mixture-of-experts architecture. Rather than activating all parameters for every query, it routes each request to a relevant subset of specialised sub-networks. This keeps latency competitive despite the model's size. The inference hardware is Nvidia's Blackwell B200, running on Google Cloud.

For developers, the practical changes are significant. App Intents is now the mandatory integration surface for Siri. SiriKit received a formal deprecation notice at WWDC; it will continue to work for now but carries compile-time warnings, and Apple has signalled a two-to-three-year support window. Any third-party app that wants Siri integration going forward must migrate to App Intents.

APPLE-BUILT LAYER Foundation Models Speech, vision, text On-device System Orchestrator Routing logic Privacy gating Private Cloud Compute Apple-run servers Verifiable privacy App Intents Developer API (SiriKit deprecated) complex queries only, personal data stripped GOOGLE-SUPPLIED LAYER Gemini (custom, 1.2T params, MoE) Nvidia Blackwell B200 on Google Cloud Apple controls the orchestration layer; Google supplies the frontier model

One developer-facing implication worth noting: Apple also announced MCP (Model Context Protocol) support in the new Siri architecture, meaning enterprise apps can expose intent libraries to Siri through standardised MCP servers. For organisations already standardised on Apple hardware, this is a meaningful shift in what agentic workflows become possible without introducing new infrastructure.


Why people are reacting the way they are

Apple users and the general press are largely positive. The product demos at WWDC showed a Siri that could handle multi-step tasks, understand on-screen context, and give directions to a landmark seen in an Instagram post. After years of Siri lagging visibly, the reaction is relief as much as excitement. Investor sentiment has reflected this: Apple stock approached record highs in the weeks leading up to WWDC as the Gemini deal became public.

Antitrust observers are more cautious. The Google-Apple relationship is already under legal scrutiny. Google pays Apple approximately $20 billion per year for search default placement on iOS and Safari, a deal at the centre of the US Justice Department's antitrust case against Google. Layering an AI licensing deal on top of that relationship raises obvious questions about whether the two companies are reinforcing each other's market positions in ways regulators will eventually challenge.

EU regulators and Apple are in open conflict. Apple announced at WWDC that Siri AI will not launch on iPhone or iPad in the EU at all. The reason is the Digital Markets Act, which requires Apple to offer the same deep system-level access it gives Siri to rival AI assistants. Apple proposed an interim solution it called the Trusted System Agent, giving rivals limited access over an 18-month transition period. The EU rejected it. Apple then applied for a blanket 18-month exemption. The EU rejected that too. The result is an indefinite freeze: EU iPhone users will not get Siri AI on launch. The EU's position is that nothing in the DMA prevented Apple from launching; Apple's position is that compliance would require it to expose user data to competitors it does not trust. Both sides have accused the other of refusing to negotiate in good faith.

Developers are split. The deprecation of SiriKit is disruptive for anyone who has built integrations against it over the past decade. The App Intents migration is not trivial. On the other hand, the new integration surface is more capable, and the MCP support opens up workflows that were not previously possible. The net reaction in developer communities is cautious pragmatism: this is better for the long term, but the migration cost is real.

Apple's AI competitors have been notably quiet publicly but have reasons to pay attention. Apple's framing of its approach as privacy-first and user-centred is a direct shot at OpenAI and Google's own assistant products. More concretely, Apple described Siri AI as built to compete head-on with ChatGPT, Claude, and Gemini. For a company that historically avoided naming rivals on stage, that is a shift in posture.


Where things stand

Siri AI ships with iOS 27 this fall, initially in English only. The EU and China are excluded at launch, for different regulatory reasons. The EU situation is unresolved and may remain so for an extended period: Apple has stopped assigning engineering resources to the EU compliance problem, which suggests the company is treating the standoff as the EU's problem to solve rather than its own.

The antitrust question in the US is live but slow. Any challenge to the Apple-Google AI deal would likely proceed through the existing Google antitrust case, the remedies phase of which is already considering whether Google's exclusivity arrangements with Apple should be unwound. Adding an AI licensing deal to that picture complicates the remedy calculus considerably.

On the product side, the open question is whether Apple's bet pays off. The company has chosen not to build a frontier model of its own, instead licensing one and concentrating its engineering on the orchestration layer, on-device models, and hardware integration. If that strategy works, Apple becomes a kind of neutral infrastructure layer in the AI wars, controlling the interface and the privacy architecture without having to win the model race. If it fails, the concern is that Apple has made itself structurally dependent on a competitor's roadmap for one of the most important features on its platform.

Apple Orchestration + on-device models Google Gemini model (~$1B/yr) ~2B iPhone users Excl. EU + China at launch licenses model ships Siri AI EU / DMA Blocked at launch. Standoff ongoing. + existing $20B/yr search default deal (under DOJ scrutiny) The Apple-Google Siri AI deal and its regulatory dimensions, June 2026

Summary

The simplest way to read the Siri AI announcement is as a concession: Apple could not win the model race on its own and chose a capable partner over continued delay. The more generous read is that Apple made a deliberate architectural bet, keeping the layer it considers strategically important (the interface, the privacy architecture, the hardware integration) while licensing the layer it would have needed years to catch up on.

Both readings can be true at the same time. What neither reading disputes is that the deal changes the map. Apple is now structurally dependent on Google for the intelligence behind its most consumer-facing AI feature. Google is now embedded in the iPhone at a level that goes well beyond search defaults. Regulators in the US and EU are watching a relationship that was already under scrutiny become significantly deeper.

The product may be great. The question is what it costs Apple over the next five years to find out.


This is a standalone post. Future posts covering AI news and releases will appear under the In Focus label.



×