Apple outsources Siri to Google's Gemini in $US1b-a-year AI deal

For Apple, the calculation was straightforward enough to write a cheque for roughly $US1 billion ($1.5 billion) a year: two years after promising a smarter Siri at WWDC 2024, the company still did not have an AI model that could deliver it. So it called Google.

Confirming the worst-kept secret in Silicon Valley, Google Cloud chief executive Thomas Kurian stepped on stage at Google Cloud Next in April and described a multi-year deal that will see Apple use Gemini as the foundation for its next-generation Apple Intelligence features. “We’re collaborating with Apple as a preferred cloud provider to develop the next generation of Apple Foundation Models based on Gemini technology,” Kurian told the audience. “These models will help power future Apple Intelligence features, including a more personalised Siri, coming later this year.”

But the timeline he projected is already under pressure. Those same features were initially slated for iOS 26.4, due in March, and have since slipped to at least iOS 26.5 — and possibly to iOS 27 in September, according to Bloomberg’s Mark Gurman, who first reported testing snags in February. Kurian’s “later this year” phrasing leaves the door open to a launch in the September–December window, yet it is a long way from the January enthusiasm that accompanied the original announcement.

No AI partnership in consumer technology carries higher stakes. It pairs the world’s largest phone maker — with more than two billion active devices — with the company that built the Transformer architecture on which every large language model depends. The deal also shows something Apple has been reluctant to acknowledge: the company’s in-house AI effort, after years of poaching researchers and acquiring startups, could not match what Google was already shipping.

The deal that remakes the AI landscape

Apple’s arrangement with Google is not a simple API licence. Valued at about $US1 billion annually, the agreement makes Google Cloud a preferred provider for developing and running Apple Foundation Models. In practice, that means Gemini runs inference for Siri queries that cannot be handled on-device, while Apple retains the ability to fine-tune the models independently.

The scope is broader than Siri alone. Apple Intelligence — the umbrella brand introduced at WWDC 2024 — covers writing tools, image generation, notification summaries, and a more context-aware operating system. Each of these features will lean on Gemini models under the hood.

For Google, the win is distribution. Its AI models will ship on every new iPhone, iPad, and Mac, giving Gemini a reach that no other cloud provider can match. Industry analysts have described the deal as a landmark moment for Google Cloud, which has until now lagged Amazon Web Services and Microsoft Azure in enterprise market share but suddenly has the consumer computing market as its addressable surface.

Apple said in the joint statement it had concluded after “careful evaluation” that Google’s technology “provides the most capable foundation for Apple Foundation Models.” The language is corporate boilerplate, but the subtext is clear: Apple evaluated alternatives — OpenAI, Anthropic, and its own models — and none met the bar. Tim Cook, on an earnings call following the announcement, reiterated the company’s commitment to a privacy-first hybrid architecture, where simple queries run on-device and complex ones are routed to Google’s cloud with end-to-end encryption.

Yet that privacy architecture is one of the reasons the timeline keeps slipping.

What the delay tells us

The Siri features first scheduled for March have now been pushed back to at least September. The problem isn’t random bugs.

Three people familiar with the testing, speaking to Bloomberg and 9to5Mac, described a recurring issue. Siri under Gemini would return accurate, conversational answers to complex queries, but it would also hallucinate personal context drawn from the user’s messages, calendar entries, and email. Under the hybrid architecture, all personal data is supposed to be processed on-device, with only anonymised queries sent to the cloud. In practice, the boundary between on-device and cloud inference keeps leaking.

This is the hardest version of the AI assistant problem. ChatGPT and Google’s own Gemini app can be conversational because they do not need to know you. Siri works differently: it is meant to be the interface to your digital life. So it must know you, and it must not misrepresent what it finds.

Apple’s problem is not that Gemini is bad. The model is too capable — it generates plausible answers to queries with incomplete context rather than declining gracefully. Teaching a system built for fluency to instead err on the side of silence is, by most accounts, where the testing hours are being burned.

The scepticism from the AI reliability community is not unwarranted. Siri has been the butt of jokes for nearly a decade, and users who have tried ChatGPT Voice or Gemini Live will not tolerate a 2026 assistant that still cannot answer follow-up questions or handle multi-step requests.

What Apple gets and what it gives up

The partnership solves an urgent problem — shipping a competitive AI assistant before the iPhone loses its edge as the premium smartphone — but it introduces a dependency that Apple has spent a decade trying to avoid.

Apple designs its own chips. It writes its own operating systems. It built its own maps service rather than rely on Google’s. Yet the decision to outsource the intelligence layer of the platform is a departure from that strategy so sharp that it effectively concedes the AI model race to the hyperscalers.

The $US1 billion annual cost is manageable for a company with $US400 billion in annual revenue. It is not trivial. Roughly equivalent to the research and development budget of a mid-sized AI lab, it will recur every year for the life of the partnership. If Apple ever wants to migrate off Gemini — whether for competitive reasons or because the regulatory environment around Google shifts — the switching cost will be measured in years of re-engineering.

In the near term, though, Apple gets a capable model that its users will actually want to use. And it gets the competitive cover: when Samsung ships Galaxy AI on its next devices, Apple will have a counter that is not built on a second-tier model.

The platform race, reordered

The Apple–Google deal reshuffles the AI platform map. Microsoft, through its OpenAI partnership, has spent two years embedding Copilot across Windows, Office, and Azure. Now Google has struck back, not by winning the enterprise, but by winning the device in every pocket.

For OpenAI, the Apple deal is a strategic miss. According to The Verge, the company had interviewed Apple for a Siri partnership before Apple closed with Google. Losing the iOS distribution channel means OpenAI’s consumer exposure remains gated behind its own app, the ChatGPT website, and whatever integrations Microsoft can drive.

Anthropic, with its Claude models, was not a serious contender, despite its reputation for safety. Apple’s procurement was a capability-first evaluation, and on capability, Gemini tested ahead.

The unanswered question is regulation. Both the ACCC and the European Commission have signalled interest in the competitive dynamics of foundation-model partnerships. A deal that routes the dominant smartphone’s AI queries through the dominant search provider’s cloud is exactly the kind of vertical tie-up that attracts antitrust attention. No regulator has opened a formal investigation, but the quiet phase — information-gathering questionnaires, informal inquiries — is already under way, according to two people familiar with the discussions.

For Australian users, the wait for a genuinely useful Siri stretches toward its third year. The assistant introduced in 2011 was meant to get smarter over time. Instead, it got a new engine and a longer timeline. Whether the Gemini-powered version ships in September or slips again will be the most closely watched software launch of the year.