Apple 选择 Google 而非 OpenAI：$1B/Year Gemini 交易让 Siri 的大脑降级

Apple，多年来一直坚持自己可以构建AI，现在同意每年支付Google 10亿美元来运行Siri，而OpenAI甚至没有参与竞争。

The Apple Siri Gemini partnership，于2026年1月宣布，并在4月22日的Google Cloud Next大会上重点展示，将一个定制的1.2万亿参数Gemini模型置于Apple重建的语音助手的中心。iOS 26.4已经开始向兼容设备推出升级的第一阶段。完整的对话式Siri，能够处理多步骤任务，跨数周的交互保持上下文，并在每个应用中实时运行，将随iOS 27和iPhone 18在9月到来。

The deal's structure: Apple每年向Google支付约10亿美元。Gemini处理智能。Apple通过其Private Cloud Compute框架路由查询，该架构使用端到端加密和硬件隔离的enclaves，因此，根据Apple的说法，没有用户数据到达Google的服务器。Gemini模型权重在Apple自己的基础设施内运行。

What that arrangement reveals about who is winning the AI race, and how the decision was actually made, is more interesting than most of the coverage has suggested.

What Happened: Apple's $1 Billion Annual Bet on Google Gemini

Apple的AI问题有据可查。这家在2011年用Siri打造了首个真正大众市场语音助手的公司，在接下来的十年里眼看着该产品成为笑柄。在WWDC 2023上，Apple宣布了“Apple Intelligence”，一套设备端AI功能，旨在表明它认真追赶。该公告引发了热情。执行却带来了延迟。

By late 2025，很明显Apple自己的大型语言模型还没有准备好为Tim Cook暗示即将到来的那种上下文对话式Siri体验提供动力。Apple需要一个合作伙伴。

OpenAI was asked and declined. According to reporting from 9to5Mac, OpenAI拒绝了Siri集成，因为两家公司现在在硬件市场上竞争，OpenAI正在通过与设计师Jony Ive的合作开发自己的AI设备。将Gemini引入Siri，而OpenAI正在构建竞争设备，这造成了双方都不想处理的冲突。

Anthropic也被 reportedly考虑过，但因不同原因被拒绝：价格。Anthropic reportedly sought "several billion dollars a year"来为Siri提供动力，Apple认为这个数字对于公司仍打算最终用自己的模型取代的过渡产品来说是不可接受的。

Google提供了性能、Apple可以向用户可信地营销的隐私架构，以及Apple可以接受的价格。它为Apple开发的定制Gemini模型运行在1.2万亿参数上，显著大于大多数公开可用的前沿模型。能力改进是可衡量的：Gemini驱动的Siri reportedly在复杂多步骤任务上达到92%的完成率，高于之前版本的58%，响应时间不到半秒。

The rollout proceeds in two phases. Phase 1，现已随iOS 26.4启动，支持屏幕感知，Siri可以实时查看和解释任何应用中的内容，以及128,000-token上下文窗口，可扩展到100万token，使其能够记住跨数周的对话。Phase 2将于9月随iOS 27到来，带来Apple一直在构建的完整对话式Siri 2.0体验。

Why This Deal Reshapes the Foundation Model Market

Google赢得的不仅仅是一个供应商合同。Apple和Samsung现在都将Google Gemini用作各自语音助手的AI骨干。这两家公司加起来占全球智能手机的80%以上。Gemini不仅仅是一个人们可以选择使用的产品，它是运行在大多数人已经携带的设备上的AI层。

The strategic effect on OpenAI is significant. Before this deal, ChatGPT existed as an optional integration in iOS，用户可以故意调用它来处理Siri无法处理的任务。随着Gemini升级，这个选择性角色进一步缩小。Gemini处理核心Siri体验；ChatGPT处理Gemini无法处理的任何内容。对于绝大多数从不更改默认设置的用户来说，这意味着他们的日常AI交互现在通过Google路由，而不是OpenAI。

Distribution at scale matters more than model quality in consumer markets. The best model in the world，通过故意应用下载访问，只能到达操作系统中内置的第二好模型所到达用户的一小部分。OpenAI的增长建立在ChatGPT的病毒式采用上。Apple deal使这条路径变得更窄。

For Apple，该安排是公司一直小心避免明确做出的承认。Apple Intelligence被营销为原生智能系统。该系统的头条功能Siri，现在运行在竞争对手的模型上。Apple的框架是，这是在其自己的1万亿参数模型（预计2027年准备就绪）开发期间的战略过渡。该模型是一个真实的项目。它是否能够按计划取代Gemini是一个真正的问题。

The deal also validates a market dynamic that AI labs have been racing to capture: OEM integration. The path to scale for foundation models is not convincing individual users to download your app. It is convincing device manufacturers to build your model into the operating system. Samsung chose Gemini for Galaxy AI. Apple chose Gemini for Siri. Both are multi-year commitments. Google now sits inside more than 80% of the world's smartphones, and the agreements run for years.

The Part Apple Isn't Advertising

Apple的整个品牌定位基于隐私 rests on a simple proposition: 你的数据与你同在，而不是与广告商同在。Google的整个商业模式 rests on a different proposition: 关于人们行为的数据是世界上最大广告业务的原材料。这两个命题很难兼容。

Apple的答案是Private Cloud Compute架构。向Gemini的查询通过Apple自己的基础设施处理，使用端到端加密和硬件隔离的计算。Gemini模型权重，即模型本身的数学参数，在Apple的服务器上运行，而不是Google的。根据Apple的说法，没有用户数据传输到Google。在Apple的说法中，在iPhone上使用Siri 2.0的体验不会让Google访问你说了什么或屏幕上显示了什么。

That claim is technically specific and plausibly true. It is also, for some observers, difficult to evaluate from the outside. Privacy architecture is only as trustworthy as the implementation, and users have no independent means of auditing whether the described safeguards are functioning as stated. For users managing sensitive research or work notes who prefer their data to stay on-device, remio's local-first knowledge base offers a privacy-conscious alternative to cloud-routed AI tools.

Elon Musk made this point publicly. When the Apple-Google deal was announced, Musk在X上批评了它， raising concerns about the concentration of AI power in a single provider. His specific concern: Apple和Samsung都使用Google creates a situation where a single company's AI models are processing interactions from more than 80% of smartphones globally. That degree of concentration, Musk argued, creates risks that apply regardless of any individual privacy claim. It echoes concerns regulators raised a decade ago about Google's dominance in search, a domain in which Google also claimed its practices were neutral and beneficial.

The competitive dynamics that led to this deal are also revealing. OpenAI declining the contract is not a footnote, it is evidence that the relationship between Apple and OpenAI is more complicated than it appears. OpenAI，通过与Jony Ive的设计公司io的合作，正在开发专用AI硬件设备。Apple，通过其AI硬件路线图，正在构建使其自己的用户不需要该设备的基础设施。这些公司不再是构建互补产品的合作伙伴。他们在有意义的方面是竞争对手。

Anthropic的拒绝是不同的。Anthropic reportedly priced itself out of the deal. That pricing reflects a real capability premium, Claude is widely considered the enterprise market's strongest model, but it also reflects a B2B strategy that prioritizes professional and enterprise users over mass-market consumer devices. Anthropic has almost no consumer distribution. It didn't get Apple, and it didn't get Samsung. Whether that gap matters to Anthropic's long-term position depends on whether enterprise market share remains more valuable than OEM integration as AI matures.

Samsung, Google, and the Race to Control the AI Layer on Every Phone

The historical parallel that fits most cleanly here is not recent. In the early 2010s, Google paid Apple approximately $1 to $3 billion per year, a figure that eventually grew to an estimated $15 billion annually, to be the default search engine in Safari. That arrangement made Google a verb in the lives of iPhone users while generating enormous revenue for Apple. It was a deal in which Apple's distribution and Google's technology produced mutual benefit, and regulators eventually asked hard questions about whether it constituted an anticompetitive arrangement.

The new deal runs the same basic structure in reverse. Now Apple is paying Google for the AI technology, and Google is paying for distribution through the relationship. The scale is smaller, $1 billion per year versus $15 billion, but the structural logic is similar: Apple controls the device, Google controls the intelligence, and together they reach the world.

For Samsung，决定将Gemini用于Galaxy AI反映了类似的推理。Samsung自己的AI研究，虽然认真，但不在大型语言模型开发的前沿。与其投资数十年和数百亿美元来构建基础模型，Samsung正在许可一个。结果是每个主要Android OEM和Apple现在共享同一个AI大脑。

For OpenAI，设备市场是战略响应。如果OEM集成是获胜的分布路径，构建设备就是你创建自己的OEM的方式。io与Jony Ive的合作是该公司对2022年推出ChatGPT作为网络产品时没有完全预料到的分布问题的回答。

iOS 27, Apple's Own Model, and the $1 Billion Question

The Gemini-powered Siri upgrade is not Apple's final position. The company has been explicit that it is developing its own 1 trillion parameter model, targeted for completion sometime in 2027. If that model performs as Apple hopes, it becomes the basis for renegotiating or exiting the Google arrangement. Apple has strong reasons to want its AI capability in-house: margin, user data (which stays in Apple's ecosystem rather than flowing through an arrangement with a competitor), and the brand coherence of being able to say Siri is genuinely Apple's.

Whether the 2027 model arrives on schedule, and whether it matches Gemini's performance, are open questions. Apple's recent track record on AI timelines, Apple Intelligence was announced in 2023 and delivered in partial, delayed waves, gives analysts reason for caution.

In the near term，2026年9月iOS 27和iPhone 18的推出是第一个真正的考验。完整的Siri 2.0体验，对话式、上下文感知、跨应用主动提供帮助，将成为大多数用户体验到的第一个真正有用的Siri版本。如果它如描述的那样工作，它将是Apple自App Store以来发布的最重要的产品变化。

The broader competitive stakes will clarify over the next 18 months. If Apple's own model is ready and capable in 2027, the Google deal becomes a transitional chapter. If Apple's model is delayed or underperforms, the $1 billion annual payment becomes the baseline for an ongoing dependence, and Google's leverage in any future negotiation grows accordingly.

The apple siri gemini deal is, at one level, a straightforward technology licensing arrangement. Apple needed a better AI engine; Google had one; they agreed on terms. At another level, it is a map of where AI power is consolidating. One company's models now run inside more than 80% of the world's smartphones, with multi-year commitments from the two largest device ecosystems on earth.

That's not a competitive outcome. That's a structural fact about how AI infrastructure is being built, and who gets to sit at the center of it.