On Tuesday April 22, 2026, Google Cloud CEO Thomas Kurian opened Cloud Next ‘26 in Las Vegas with a keynote that buried the most important consumer-AI announcement of the month inside a corporate-customer-roundup slide. Apple, Kurian said, is now a Google Cloud customer of strategic importance, and Apple’s revamped Siri — coming “later this year” — will run on Apple Foundation Models that are themselves built on Gemini. MacRumors and 9to5Mac ran the news within the hour. By Friday, Bloomberg’s prior reporting that Apple is paying Google approximately $1 billion per year for a custom 1.2-trillion-parameter Gemini variant had been re-quoted in every English-language tech publication and most non-English ones.
This guide is the read-out for the iPhone owner who wants to know what’s actually changing in the device they paid $1,200 for. What was confirmed at Cloud Next versus what’s still Bloomberg-sourced rumor. How Siri will route requests in the new architecture. What data goes where, and to whom. How the Apr 27 Microsoft-OpenAI restructure changes the calculus. And the practical question every privacy-conscious user is asking: should I disable Gemini-on-Siri when iOS 27 ships, and how would I even do that?
What Was Actually Announced
Two distinct news points worth keeping separate.
Confirmed at Cloud Next on April 22. Thomas Kurian named Apple as a strategic Google Cloud customer, described Google Cloud as Apple’s “preferred cloud provider” for the new Apple Foundation Models, and stated that Gemini-based Apple Foundation Models will power “a more personalized Siri coming later this year.” That’s the on-stage, on-the-record material. Kurian did not state a parameter count, did not state a price, did not give a specific iOS version number, and did not commit to a launch date beyond “later this year.”
Bloomberg-sourced, widely echoed but not re-confirmed at Cloud Next. The 1.2-trillion-parameter custom Gemini variant. The ~$1 billion per year payment. The multi-year contract. These numbers come from Bloomberg’s November 2025 reporting and Reuters’ summary of the same. They have been repeated in SiliconANGLE, 9to5Mac, MacRumors, and most subsequent analyses, but they were not explicitly re-confirmed by Kurian on stage. They are widely accepted as accurate; they are not yet on Apple’s or Google’s official record.
Already in motion before Cloud Next. Apple and Google issued a joint statement on January 11-12, 2026 announcing a multi-year collaboration in which “the next generation of Apple Foundation Models will be based on Google’s Gemini models and cloud infrastructure.” Cloud Next is best understood as the first public co-marketing of that already-announced partnership, not the original announcement.
The honest synthesis: Gemini-Siri is real, the deal terms are credible enough to assume, but the hardest specifics (which queries route to Gemini, what privacy guarantees apply when, what the iOS 26.4 versus iOS 27 split is) remain underspecified by Apple itself. This guide is honest about what’s verified and what’s analytical inference.
How Siri Will Route Your Requests
The architecture, as best as can be reconstructed from converging technical reporting, is a three-tier model.
Tier 1 — On-device Apple models for local tasks. When you say “Set a timer for ten minutes,” “Turn off Wi-Fi,” “Open Notes,” “Play the album I was listening to last night,” your iPhone handles the request entirely on-device using Apple-built models running on the Neural Engine. Nothing leaves your phone. This is how Siri has worked for the past two iOS cycles, and it doesn’t change with the Gemini integration.
Tier 2 — Gemini-based Apple Foundation Models in the cloud for heavy queries. When you say “Summarize my last three emails from my boss,” “Plan a weekend trip to Lisbon and put the flights on my calendar,” “What’s a good gift for my dad’s 65th birthday based on what I’ve shared with him in iMessage this year,” “Look at this picture and tell me what kind of plant it is,” the request is heavy enough that on-device inference can’t satisfy it. The query escalates to the cloud — and this is where the Gemini integration kicks in.
Tier 3 — Third-party providers for opt-in extensions. Apple’s multi-provider Siri architecture, which has been live in some form since the ChatGPT-on-Siri integration shipped in late 2024, lets a user grant explicit permission for Siri to route specific queries to a third-party provider (currently ChatGPT; soon, presumably, Gemini and likely Claude as well). This is the tier where the user sees a prompt — “Send this question to ChatGPT?” — before any data leaves the Apple-controlled stack.
The genuinely new part of the Apr 22 announcement is what’s happening in Tier 2. Apple’s previous architecture for cloud-tier Siri queries used Apple-built foundation models running on Apple’s Private Cloud Compute (PCC). Going forward, those Apple Foundation Models are themselves derived from Gemini — either as fine-tuned variants of Google’s 1.2T-parameter custom model, or (depending on the routing) directly invoking Gemini at runtime. The user-facing surface is the same; the underlying intelligence shifts from “Apple-built” to “Apple-built on top of Google’s foundation.”
The Privacy Architecture (And the Honest Gaps)
This is where the public reporting diverges, and where you have to be careful what you accept as a verified fact.
Apple’s Private Cloud Compute design is genuinely strong. The published PCC architecture has four pillars: hardware running on Apple silicon and Apple-controlled servers; requests encrypted in transit and processed statelessly with session data deleted after inference; cryptographic attestation that lets the device verify only Apple-approved software is running on the PCC node; and PCC binaries that are made available for inspection by independent researchers in a Virtual Research Environment. None of this is marketing fluff — Apple’s PCC documentation is one of the most rigorous publicly-released privacy architectures any major tech company has ever shipped.
How Gemini fits into PCC is where the reporting splits. Two narratives are competing in the coverage.
Narrative A — Gemini runs inside PCC. Several technical explainers (HappyCapyGuide, Kersai, AI CERTs, WindowsForum) describe a design where Apple has integrated Gemini’s runtime or the model weights themselves into PCC nodes. In this version, when Siri escalates a query to the cloud, the inference happens on Apple-controlled hardware running an Apple-attested software stack. Google has no runtime visibility into your prompts. The security properties of PCC apply unchanged.
Narrative B — Some Gemini workloads run in Google data centers. MacRumors’ Cloud Next coverage notes that “it is unclear if the new Siri and Gemini-powered Apple Intelligence features will use Private Cloud Compute or will run on Google’s servers,” adding that Apple has reportedly asked Google to explore running Siri workloads in Google data centers because of expected high-volume cloud usage. In this version, at least some queries leave Apple’s control and run on Google infrastructure, with whatever de-identification and encryption Apple negotiated as part of the contract.
The conservative read. Until Apple publishes an official Siri-on-Gemini architecture paper, treat Narrative A as the design goal and Narrative B as the realistic possibility. Strong “Google can’t see anything” claims from third-party explainers are technically possible but not yet audited. Apple’s track record on PCC inspires trust; the practical engineering realities of running 1.2-trillion-parameter inference at iPhone-scale volume may have required compromises that haven’t been publicly documented.
What this means for your behavior depends on your threat model. If you’re a typical user concerned about general data minimization, the worst-case Narrative B is still substantially better than how most cloud-AI services handle data today. If you’re a journalist, lawyer, or activist with adversarial-state threat concerns, the absence of an audited end-to-end architecture is a real reason to use the on-device-only Siri configuration when iOS 27 ships.
Multi-Provider Siri: How Gemini Coexists With ChatGPT (and What’s Coming Next)
The architecture Apple chose isn’t “replace Siri with Gemini” — it’s “make Siri itself smarter, then let users opt into specific third-party providers for the queries they specifically want a different model to handle.” The two layers operate independently.
Gemini-based Apple Foundation Models = the default Siri. When you talk to Siri after iOS 27, you’re getting Gemini-derived intelligence by default, mediated through Apple’s PCC architecture (or, per Narrative B, partially through Google data centers under contractual privacy terms). You don’t choose this; Apple chose it for you when you bought an iPhone.
ChatGPT-on-Siri is a specific opt-in. Distinct from the underlying Siri model. When Siri can’t answer something or the user explicitly asks “use ChatGPT,” a prompt asks permission to send the query to OpenAI. The data leaves Apple’s control. The user knows this, because the prompt says so.
Gemini-on-Siri (as a specific opt-in) is rumored to follow. Apple has filed patents for a multi-provider Siri framework that exposes the same opt-in pattern for additional providers. Industry coverage strongly implies Gemini will be added as an explicit opt-in choice (separate from the underlying model), and Claude is rumored as a third option in the same framework. Timing for these explicit opt-ins is likely iOS 27.x rather than iOS 27.0.
The practical implication. If you currently use ChatGPT-on-Siri because the default Siri couldn’t handle your queries, the new default Siri will probably handle most of them — eliminating the round-trip to OpenAI. If you specifically prefer ChatGPT’s writing voice, ChatGPT’s news recency, or ChatGPT’s specific tooling, you’ll keep using the explicit opt-in. The two layers don’t conflict; they compose.
Has Apple Effectively Conceded the Frontier Race?
This is the honest question every Apple-watcher is wrestling with, and the honest answer is “yes, in one specific way, and no in another.”
The yes. Apple does not currently ship its own 1.2T-parameter frontier model. Apple’s internal Foundation Models (the on-device tier and the smaller PCC-tier models) are competitive at narrow tasks but trail the frontier. The decision to build Apple Foundation Models on top of Gemini, rather than scaling Apple’s own foundation work to frontier compute and training budgets, is an admission that Apple has chosen not to compete at the frontier-model layer. It’s an “infrastructure-as-leverage” play — Apple owns the device, the OS, the chip, the privacy architecture, and the consumer relationship; Google provides the underlying intelligence at a wholesale price. That’s a credible strategic choice, but it is structurally different from “Apple is going to win at AI.”
The no. Apple Intelligence as a product strategy was never premised on Apple shipping the world’s most capable model. The premise was that Apple would ship the world’s most integrated and private AI experience on premium consumer hardware. That premise is unaffected by who trains the underlying weights — and is arguably strengthened by the Gemini deal, because Apple now ships frontier-quality intelligence on PCC-class privacy infrastructure without taking on the operational cost and risk of training the model. The user-facing product gets better; the strategic position is more defensible than trying to outspend OpenAI and Google on training compute.
The market read. Apple stock has been roughly flat through the announcement and the subsequent Cloud Next coverage. Investors appear to have largely accepted the trade — Apple is paying Google to ship a better Siri faster than Apple alone could, and the math works for Apple’s gross margins as long as the Siri experience generates services revenue and lock-in commensurate with the $1B/year deal cost.
How the Apr 27 Microsoft-OpenAI Deal Changes the Picture
Three days after Cloud Next, Microsoft and OpenAI rewrote their partnership — exclusivity ended, OpenAI is free to ship on AWS and Google Cloud, the AGI clause was deleted. For Apple-Gemini-Siri, three implications worth tracking.
Apple’s hedging position improves. Now that OpenAI can ship on Google Cloud, Apple has a credible path to expanding ChatGPT-on-Siri (or even routing some of the underlying Apple Foundation Model layer through OpenAI in the future) without conflicting with the existing Gemini deal. Apple becomes a multi-vendor AI consumer with leverage on each.
Google’s exclusivity in the Apple deal looks more valuable. Until Apr 27, “Apple uses Gemini for Siri” was one of several frontier deployments. Post-Apr 27, with OpenAI shipping multi-cloud, Google’s Apple deal is one of the few remaining genuinely strategic enterprise wins for any single AI provider. Watch for Google to leverage this more aggressively in its enterprise sales motion.
Consumer choice expands. The likely 12-24 month trajectory is: Gemini becomes the default Siri intelligence (via Apple Foundation Models), ChatGPT and Gemini both exist as opt-in third-party Siri providers, Claude joins the opt-in list as Anthropic builds out its consumer presence. The user gets to pick the model whose voice, factual accuracy, and writing style they prefer. That’s a measurably better consumer-AI ecosystem than where we were a year ago.
Should You Disable Gemini-on-Siri? (Privacy-Conscious Users)
A practical answer for the user who’s worried.
For most users, the right answer is: no, don’t disable it. Apple’s PCC architecture, even in the conservative Narrative B reading, is meaningfully more privacy-respecting than the typical cloud-AI service you’re already using daily. The on-device tier handles sensitive personal queries entirely locally. The cloud tier escalates only what the on-device model can’t satisfy. The data minimization, encryption, and statelessness guarantees apply even if the inference partially runs in Google data centers under contractual privacy terms.
For users with elevated threat models (journalists, lawyers, activists, healthcare professionals discussing patient information), the right answer is: configure for on-device-only Siri until you’ve reviewed Apple’s official architecture documentation post-iOS 27. The path will likely be Settings → Apple Intelligence → “Use cloud-tier inference” → Off. (Exact menu names will be set by Apple at iOS 27 release; the configuration option is widely expected to exist based on Apple’s previous PCC opt-out controls.) You’ll lose the heavy-query capabilities — long summarization, multi-step planning, multimodal queries — but the sensitive parts of your workflow stay strictly on your phone.
For everyone, regardless of threat model: turn off Siri’s “Improve Siri & Dictation” setting if you haven’t already. This is the long-standing Apple control that opts you out of having voice samples reviewed by humans for model improvement. It’s distinct from the Gemini integration but it’s the single most impactful privacy lever Apple has shipped in the past five years, and most users haven’t toggled it.
Bottom Line
The Cloud Next confirmation is real but underspecified. Gemini will power Apple’s revamped Siri later this year. The 1.2-trillion-parameter and ~$1B/year numbers are credible but Bloomberg-sourced. The privacy architecture is best-in-class on paper; some open questions remain on whether all inference runs inside PCC versus partially in Google data centers. The multi-provider Siri model means ChatGPT-on-Siri continues, Gemini-on-Siri arrives as both a default and (likely) an explicit opt-in, Claude probably follows.
For the typical iPhone owner, the practical effect is “Siri gets meaningfully smarter when iOS 27 ships, with privacy guarantees that are imperfect but unusually strong by industry standards.” For the privacy-conscious user, the practical action is to keep an eye on Apple’s iOS 27 documentation for the on-device-only configuration option, and to toggle “Improve Siri & Dictation” off in the meantime.
The strategic story — Apple paying Google a billion dollars a year because the cost of training a 1.2T-parameter model is greater than the cost of buying access to one — is the most honest acknowledgment any of us are likely to get this year about how the frontier-model race is consolidating. Two of the three named providers in the consumer-AI conversation now route through the same underlying Google infrastructure: Apple’s Siri, and (effectively) any device on Google Workspace. That’s not a bad outcome for users. It’s a more concentrated outcome than it looked like a year ago.
