谷歌AI开始过收入关:云业务翻倍之后,企业该学什么

10月28日，Alphabet交出了一份容易被“25%收入增长”掩盖的财报。

真正值得看的是另外三组数字：

Google Cloud收入307亿美元，同比增长103%；第一方模型API每分钟处理接近320亿个Token；Gemini应用用户达到10亿。公司还明确说，TPU系统销售已经对云业务收入形成了有意义的贡献。blog.google

把这几条放在一起，谷歌正在完成一件AI公司普遍想做、却很难做成的事：让模型能力穿过产品、基础设施和销售体系，变成可以出现在财报里的收入。

这份财报回答的，不是“Gemini强不强”

模型榜单更新很快。企业真正关心的问题通常更朴素：

客户为什么付钱？按什么计价？用得越多，毛利会不会被算力成本吃掉？这项产品能否进入原来的采购流程？

谷歌给出的答案分成几层。

搜索业务利用AI增加查询量，尤其是带有商业意图的查询。这里的收入仍主要来自广告，AI负责提高产品使用频率。

Gemini应用承担用户入口。10亿用户很醒目，但用户数不等于收入。它的作用更接近分发渠道：建立使用习惯，再把用户带向订阅、办公软件和云服务。

模型API按Token和调用量收费。每分钟320亿个Token说明使用规模已经扩大，但这项业务能不能持续赚钱，还要看推理成本、折扣和客户结构。

再往下是TPU。过去，自研芯片更多被理解为谷歌降低内部成本的工具。现在TPU开始直接对外销售并贡献云收入，它从成本中心的一部分变成了商品。

这四层组合起来，才是谷歌AI业务的真实形态。

谷歌卖的不是一个模型，而是四种计价单位

从产品角度看，谷歌把同一套AI能力拆成了四种商品：

广告按流量和转化计价；Gemini按用户和订阅计价；模型API按调用量计价；云基础设施按算力、存储和芯片系统计价。

这比“做一个AI助手，然后收月费”难得多，却也稳得多。

单一AI产品会遇到两个麻烦。客户容易比较价格，模型能力也可能很快被同行追上。谷歌的做法是把模型放进搜索、Workspace、Cloud和TPU，让客户的切换成本来自整套工作过程，而不是某个聊天窗口。

投行视角下，这种收入结构还有一个好处：不同业务的现金流周期不一样。广告提供现金，云业务承接企业长期合同，API随使用量增长，芯片和基础设施吸收大额采购预算。它们可以互相支撑。

风险也很清楚。Alphabet第三季度资本开支上升，季度自由现金流转为负值。云收入翻倍，并不代表AI投入已经全面进入轻资产盈利期。路透对同期大型科技公司的比较也显示，高额资本投入正在拉开企业之间的融资能力差距。reuters.com

企业不能照抄谷歌，但可以照着问问题

大多数企业没有必要训练基础模型，更不需要自建芯片。

可以学的是它的收入设计顺序。

先找原有收入里，AI能改善的环节。可能是提高销售转化、缩短交付时间，或者让客户多使用一次产品。这个阶段不要急着给AI单独定价。

等使用频率稳定，再把其中可重复的能力做成收费模块。例如自动生成报告、审核合同、整理售后工单。客户付费的理由应该是任务完成得更快、更稳定，而不是因为界面上出现了AI按钮。

使用量继续增长后，企业才需要考虑按席位、调用量还是结果收费。若一次任务会消耗大量模型调用，按固定月费可能越卖越亏；若客户很少使用，按调用收费又难以形成稳定收入。

还有一项经常被忽略：内部能力能不能外卖。

TPU的变化很有启发。企业为自己建设的知识检索、数据清洗、质量检测或供应链预测系统，如果能够标准化，未来也可能变成面向供应商或同行的服务。当然，前提是数据权限清楚，产品确实可以脱离内部团队运行。

一张更实用的AI项目检查表

准备投入AI项目时，可以先回答五个问题：

• 客户原本为什么付钱？
• AI改变了哪个可测量指标？
• 收入按用户、使用量还是结果计算？
• 使用量增加时，单位成本会上升还是下降？
• 如果底层模型更换，客户是否仍愿意留下？

这五个问题答不清，项目可能只是演示。

答得清楚，也别急着扩大范围。先选一个高频任务，跑满一个真实业务周期。记录人工耗时、模型费用、错误率、复核时间和客户续用率。数据不好看时及时停下，比继续采购更多算力更有价值。

谷歌跨过的是收入门，不是终点线

Google Cloud同比增长103%，说明企业愿意为AI基础设施和解决方案付费。TPU开始贡献收入，则说明谷歌正在把自研能力变成外部商品。

这仍不是“AI投资已经全部兑现”的证明。资本开支、自由现金流和模型成本还会继续受到市场追问。

但财报至少给出了一个更可靠的观察方法：少看发布会里有多少新功能，多看客户按什么单位付钱，以及这些收入能否承担下一轮投入。

这比任何一张模型排行榜都更接近商业现实。

本文基于截至2026年10月29日可核验的公开信息撰写。财报中的云收入不能全部归因于生成式AI，TPU、Gemini和模型API的独立收入尚未完整披露。文中商业分析不构成证券、投资、采购或技术选型建议。企业部署AI时应单独评估数据安全、知识产权、模型错误、供应商依赖和实际成本。

English Version

Google AI Is Starting to Pass the Revenue Test

Alphabet’s October 28 earnings report contained three numbers that matter more than the headline growth rate.

Google Cloud revenue reached $30.7 billion, up 103% year over year. Google’s first-party model APIs were processing nearly 32 billion tokens per minute. The Gemini app reached one billion users. Alphabet also said TPU system sales had become a meaningful contributor to cloud revenue.

Together, these figures show how Google is turning AI capability into several distinct products—and several different billing units.

This is more useful than asking whether Gemini leads a particular benchmark. Enterprise customers care about practical questions: Why will a customer pay? How will usage be measured? Does heavier usage improve margins, or does it simply increase inference costs? Can the product fit into an existing procurement process?

Google’s answer has four layers.

Search uses AI to increase query activity, including commercially relevant queries. The revenue still comes mainly from advertising; AI improves product usage.

Gemini provides a consumer distribution channel. One billion users is a powerful reach metric, but it is not the same as revenue. Gemini can support subscriptions and direct users toward Google Workspace and Cloud services.

Model APIs create usage-based revenue. The token volume shows adoption, although profitability still depends on inference costs, discounts and customer mix.

TPUs add another layer. Google originally developed its chips largely to support internal workloads and control infrastructure costs. TPU systems are now becoming products sold to outside customers. An internal cost capability is turning into an external revenue line.

Google is therefore monetizing AI through advertising, subscriptions, API consumption and infrastructure sales. That structure is harder to build than a standalone chatbot subscription, but it is also more resilient.

A single AI product is easy to compare with alternatives. Model performance changes quickly, and price competition can be severe. Google is placing its models inside Search, Workspace, Cloud and TPU systems. The customer relationship depends on the surrounding workflow, data and infrastructure—not only on a chat interface.

There is a financial advantage as well. Advertising generates cash. Cloud contracts can provide longer-term commitments. APIs grow with usage. Chips and infrastructure capture large enterprise budgets. Different revenue cycles can support one another.

The risks remain substantial. Alphabet’s capital expenditure continued to rise, and quarterly free cash flow was negative. Cloud revenue doubling does not mean the company has reached a low-capital, high-margin phase of AI deployment. The market will continue to ask whether the cash generated by AI-related products can support the next round of infrastructure spending.

Most companies cannot copy Google’s investment scale. They can copy its sequence of questions.

Start with an existing revenue stream. Identify where AI can improve a measurable outcome: conversion, delivery time, service quality or customer usage. At this stage, the company may not need a separate “AI price.”

When usage becomes consistent, convert a repeatable capability into a paid feature. Examples include report preparation, contract review or service-ticket classification. Customers should pay because a real task is completed more reliably—not because the interface contains an AI label.

Only then should the company decide whether to charge per seat, per action or per outcome. Fixed subscriptions may be dangerous when each customer action consumes expensive model calls. Usage pricing may be unattractive when customers use the service infrequently.

Companies should also ask whether an internal AI capability can become an external product. A system built for internal data cleaning, quality inspection or supply-chain forecasting may eventually be useful to suppliers or industry peers. This works only when permissions are clear and the system can operate without constant help from the original team.

Before funding an AI project, management should answer five questions:

• What did the customer already pay for?
• Which measurable result does AI change?
• Is revenue tied to users, usage or outcomes?
• What happens to unit cost as usage increases?
• If the underlying model changes, will the customer stay?

If those questions cannot be answered, the project may still be a demonstration rather than a business.

Google Cloud’s 103% growth suggests that customers are willing to pay for AI infrastructure and enterprise solutions. TPU sales show that Google can also commercialize technology originally built for its own operations.

This does not mean the AI investment cycle has been fully proven. It does mean that investors and corporate leaders now have a better test: look past model announcements and examine what customers are paying for, how they are charged and whether the resulting cash can fund the next stage of growth.