On May 25, Microsoft released its fiscal third-quarter earnings report for fiscal year 2026. Total revenue reached $82.9 billion, a year-over-year increase of 18%. This figure itself is not surprising; what is surprising are two data points hidden within it.
First: AI business annualized revenue run rate exceeds $37 billion, up 123% year-over-year.
Second: Copilot paid seats reached 20 million.
微软AI业务年化收入运行率370亿美元 · 同比增长123% · 2026财年第三季度
$370亿美元意味着什么？它已经超过了许多上市科技公司的年收入。而123%的同比增长率又说明了什么——一年前，这个数字还不到170亿美元，如今已经翻了一番多。
Where does the money come from? Just look at RPO to find out.
RPO（剩余履约义务）指标比当前季度收入更具说服力。微软本季度的RPO达到6270亿美元，同比增长99%，几乎翻了一番。
RPO is the amount of revenue that customers have signed contracts for but have not yet recognized. The 99% growth indicates two things: first, large enterprises have a very strong willingness to sign AI contracts; second, these contracts are typically multi-year, large-scale deals. Azure cloud services grew 39%-40% at constant currency, with AI-related computing power demand being the main driver.
AI业务年收入370亿美元，同比增长123%。剩余履约义务（RPO）为6270亿美元，同比增长99%。
招商证券（香港）维持买入评级，目标价616.4美元。Wedbush重申跑赢大盘评级，目标价575美元。两家机构的核心逻辑一致：Azure的AI需求尚未见顶，Copilot的企业渗透仍在加速。
But here is a noteworthy signal: the biggest concern among institutions has shifted from "can it rise" to "how long can high growth be sustained". At the Evercore Global TMT Conference on June 2, analysts will likely press for answers on the return on capital expenditure.
On the same day, the domestic policy floodgates opened
On the same day Microsoft released its data (May 26), Xinhua Net published a lengthy article titled "AI Agents: More Than Just Chatting, They Can Really Get Work Done." The article contained a wealth of information, but the most interesting part was not the technical details—it was the timing.
5月，国家互联网信息办公室、国家发展和改革委员会、工业和信息化部联合印发了《关于智能体标准化应用和创新发展的实施意见》。
This is the first time at the national level to give an official definition of "intelligent agent": an intelligent system with autonomous perception, memory, decision-making, interaction, and execution capabilities, which is an important form of artificial intelligence products and services.
The document lists 19 typical application scenarios, covering industries such as manufacturing, finance, and government affairs, requiring the establishment of full-chain safety norms. In other words, previously, enterprises using intelligent agents were in a "gray area"—usable, but without rules. Now the rules are here.
在最近的一次公开演讲中，李彦宏表示：“第一次，AI的主角不是模型，而是应用。”他提出用DAA（每日活跃智能体）取代DAU，作为衡量平台生态的新指标——即有多少智能体在为人类工作并交付成果。
This statement is particularly interesting when viewed against the backdrop of Microsoft's $37 billion. Essentially, the money Microsoft earns comes from companies paying for "AI that can work." Behind each of Copilot's 20 million paid seats is an enterprise user using AI to perform actual work tasks.
Three signals worth noting
Reading Microsoft's financial report together with China's domestic intelligent agent policies reveals several trends:
First, enterprise AI procurement is shifting from "testing the waters" to "budgeting." Microsoft's RPO has nearly doubled, indicating that companies are signing long-term contracts rather than one-time pilot projects. AI budgets are beginning to enter annual IT spending plans, much like cloud services and office software.
其次，“执行能力”取代“模型参数”成为竞争焦点。新华社的报道反复强调一个词——“行动AI”。百度、阿里巴巴、腾讯和字节跳动都在推广自己的Agent平台。Kimi帮助证券公司将研究报告初稿的撰写时间从2-3天压缩到2-3小时。千问×淘宝秒杀实现了“一键下单”+自动劝阻不合理购买。这些不是聊天机器人，而是直接介入业务流程的执行者。
第三，安全边界成为硬约束。三部门《实施意见》专门提到安全治理。360人工智能安全研究院指出，风险正从"生成风险"转向"执行风险"——人工智能不再只是说错话，还可能在没有人类监督的情况下自主执行越界操作。阿里巴巴的通义千问推出了"引用"功能，自动将信息不明确的来源标红，而深势科技的科研智能体则通过可复现计算验证关键结论。这些都是在为智能体安装"安全阀"。
这对CIO和数字领导者意味着什么
If you are planning your company's AI strategy, here are a few practical suggestions:
- Stop just doing PoCs. Microsoft's data proves that enterprises that truly generate value have already entered the stage of large-scale deployment. If your AI project is still stuck in "selection and testing," you may need to pick up the pace in the second half of 2026.
- Focus on the security framework of intelligent agents. Domestic policies have begun to regulate the security boundaries of intelligent agents, and when purchasing, the supplier's security governance capabilities should be included in the evaluation criteria—not "whether the function exists," but "how to handle the fallout if something goes wrong."
- Re-examining the AI budget structure.Token costs are becoming ongoing operational expenses (the three major telecom operators are already selling token packages), a cost model completely different from traditional software licensing fees. It is recommended to set up a separate ledger.
The $37 billion figure is impressive, but for most Chinese companies, what may be more instructive is not how much Microsoft earned, but the customers behind it—why they are willing to sign multi-year contracts, and what kind of business value these contracts ultimately deliver.
References: Microsoft FY2026 Q3 Financial Report / Xinhua News Agency report on May 26 / "Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents"
