Apple Rebuilds Siri on a Secret 1.2-Trillion-Parameter Gemini Model. Here's What That Means for Your Privacy.

The model uses a mixture-of-experts design, meaning it activates only a relevant subset of its parameters for any given query—a setup that preserves the knowledge capacity of a trillion-parameter system while keeping response latency competitive with much smaller models . Apple's own largest previous cloud model was approximately eight times smaller, and the company's Private Cloud Compute servers were reportedly too slow during testing to host a model of this scale efficiently .

How Nvidia and Google technologies fit together

For the heaviest Siri AI queries, the data path runs through Google Cloud running Nvidia Blackwell B200 GPUs . The integration breaks down into three components:

• Google Cloud hosts the licensed Gemini foundation model. Apple does not use Google's serving infrastructure, search, or the consumer-facing Gemini app. Craig Federighi made the line explicit: "The amount of the Google assistant we use, which is none" .
• Nvidia Confidential Computing encrypts data in use during cloud processing on Google's Nvidia-powered servers. Apple worked with both Google and Nvidia to extend its Private Cloud Compute (PCC) security model onto those third-party machines while preserving verifiable privacy guarantees .
• On-device Apple Silicon runs smaller models trained from Gemini through knowledge distillation. This keeps basic AI tasks—text suggestions, simple voice commands, on-screen photo analysis—local and fast .

Apple's own third-generation Apple Foundation Models (AFM) handle the majority of on-device work, with cloud handoff triggered only for queries that exceed local capacity .

The three-tier privacy architecture

The privacy story is the centerpiece of Apple's pitch, and it spans three distinct layers :

1. On-device inference – Small distilled models run locally on Apple Silicon. Data never leaves the device for these tasks, and nothing is logged.
2. Apple Private Cloud Compute (PCC) – Apple's own stateless servers handle moderate-complexity requests. The PCC architecture enforces enforceable guarantees: no persistent storage, no privileged runtime access, and no ability to target a specific user.
3. Extended PCC on Google Cloud with Nvidia GPUs – When a query requires the full 1.2T-parameter model, processing moves to Google Cloud servers equipped with Nvidia GPUs. Apple extended the core PCC requirements—stateless computation, non-targetability, no privileged access, and verifiable transparency—to these third-party nodes . The implementation uses Nvidia Confidential Computing on the GPUs, Intel CPUs with TDX (Trust Domain Extensions), and Google's Titan chip to establish a hardware root of trust . Independent security researchers are permitted to verify the entire trusted computing base.

Apple confirms that data is anonymized at each hop, and all server components and software fall under the same verifiable transparency commitments that apply to Apple's own PCC nodes .

The practical routing logic: simple tasks never leave the device; medium tasks hit Apple's PCC; genuinely hard reasoning requests reach Google Cloud—with the same privacy guarantees enforced at each tier.

Apple's relationship with Google: model licensing, not a consumer product deal

What Apple and Google built is a licensing and engineering partnership, not a consumer-facing collaboration . The distinctions matter:

• Apple licensed a custom Gemini model and collaborated on its tuning. The model runs exclusively for Siri AI queries; it is not a white-label of the Gemini chatbot .
• Gemini frontier models are used to teach Apple's smaller on-device models through distillation—Gemini generates outputs, and Apple's AFM models learn from those outputs to handle local tasks without a cloud round trip .
• Apple Intelligence retains its own architecture, brand, and user experience. Siri AI does not point users to a Google product, does not route queries to Google Search, and does not surface the Google brand in the interface .
• The financial arrangement is reportedly $1 billion per year, but neither Apple nor Google has confirmed the number publicly .

In short, Google provides the foundation model and the cloud silicon; Apple provides the product, the interface, the on-device models, and the privacy architecture. The user never touches a Google surface.

Siri AI launch timeline: September 2026 beta, but not everywhere

Apple confirmed a public beta of Siri AI shipping with iOS 27 in September 2026, alongside the expected release of new iPhone hardware . A July public beta was also mentioned in some reports, though the keynote emphasized the September timeframe . A dedicated standalone Siri app with conversation history synced via iCloud arrives at the same time .

A stable release beyond the beta will follow later in 2026; Apple's official website uses the phrase "later this year" without a firm date, which reportedly rattled some investors who wanted an explicit September general-availability commitment .

The fine print has real teeth:

• No EU launch on iPhone or iPad at launch. Apple cites the Digital Markets Act (DMA), which requires operating systems to be opened to third-party developers. Apple argues that DMA compliance in its current form would force the company to weaken the privacy and security guarantees of Private Cloud Compute. The macOS version also remains unavailable in the EU at launch .
• No China launch at all at launch. Apple is still working through local regulatory requirements for Siri AI and the broader Apple Intelligence feature set .
• English only at launch, with additional languages promised to follow quickly .
• Hardware requirements bite, too. Apple noted that the most powerful on-device model, expressive Siri voices, and advanced dictation features will require the company's most capable devices—which likely means recent iPhone, iPad, and Mac models .

Siri AI is the biggest AI bet Apple has ever made, built on a $1 billion/year handshake with its biggest platform rival. The architecture—three tiers, three companies, one verifiable privacy model—is either the blueprint for privacy-preserving cloud AI or the most elaborate trust exercise in consumer tech. The September beta will start to answer which one it is.