微软财报、CoreAI/MantleAI、其他说明

Microsoft Earnings, CoreAI/MantleAI, Additional Notes 《微软财报、CoreAI/MantleAI、其他说明– Ben Thompson的Stratechery》 Wednesday, November 12, 2025Listen to Podcast收听播客 Listen to this post:收听此⽂章： Good morning,早上好， Onthis week’s episode of Sharp China, Andrew and Bill discuss the implementation of the U.S.-China trade deal,and more setbacks for Nvidia. 在本周的《锐评中国》节⽬中，安德鲁和⽐尔讨论了中美贸易协议的实施情况，以及英伟达遭遇的更多挫折。 On to the Update: 接下来的更新： 微软财报 From theWall Street Journal: 来⾃《华尔街⽇报》： Microsoft is seeing more demand for its cloud computing and AI services than it can keep up with, a challengethat is supercharging the company’s income. Demand for the tech giant’s cloud services is so great thatMicrosoft will boost its AI capacity by more than 80% this year and double its total data-center footprint in thenext two years, Chief Executive Satya Nadella told investors late Wednesday. That outlook means Microsoftnow expects to spend more than previously projected on its AI infrastructure this fiscal year — and still beshort of the capacity it will need to meet demand. 微软的云计算和⼈⼯智能服务需求量之⼤，已超出其供应能⼒，这⼀挑战正极⼤地推动公司收⼊增⻓。⾸席执⾏官萨蒂亚·纳德拉周三晚间告诉投资者，对这家科技巨头云服务的需求如此之⼤，以⾄于微软今年将把其⼈⼯智能容量提⾼80%以上，并在未来两年内将其数据中⼼总⾯积扩⼤⼀倍。这⼀展望意味着微软现在预计本财年将在其⼈⼯智能基础设施上投⼊⽐之前预想的更多资⾦——但仍将⽆法满⾜需求。 Companies have been eager to host and train their artificial-intelligence models on Microsoft’s wide array ofservices. That keeps paying o!for Microsoft, which reported revenue of $77.7 billion for its first fiscal quarter. The figure exceeded Wall Street expectations. The company’s closely watched Azure cloud business grew byabout 40%, also topping expectations. Operating income increased 24% to $38 billion, more than consensusprojections. The company said its net income was $27.7 billion, or $3.72 per diluted share. 各公司⼀直渴望在微软⼴泛的服务上托管和训练他们的⼈⼯智能模型。这持续为微软带来回报，该公司报告其第⼀财季收⼊为777亿美元。这⼀数字超出了华尔街的预期。该公司备受关注的Azure云业务增⻓了约40%，也超出了预期。运营收⼊增⻓24%⾄380亿美元，超过了普遍预测。该公司表示其净收⼊为277亿美元，或每股摊薄收益3.72美元。 Despite beating Wall Street’s projections, Microsoft shares ticked lower by nearly 4% in after-hours trading, asthe company reported substantial spending on cloud and AI infrastructure. The company took a $3.1 billioncharge from its massive investment in Open AI. Some investors are worried there is overreliance on cloudcontracts from OpenAI, reflecting a broader industry concern. 尽管超出了华尔街的预期，但微软股价在盘后交易中下跌了近4%，因为该公司报告了在云和⼈⼯智能基础设施上的巨额⽀出。该公司因对OpenAI的巨额投资⽽产⽣了31亿美元的费⽤。⼀些投资者担⼼对OpenAI的云合同过度依赖，这反映了更⼴泛的⾏业担忧。 I’m a bit late on these earnings, but Microsoft’s stock is still down a bit, and CEO Satya Nadella seemed to anticipateinvestor concerns. His message onthe earnings callcame down to one word: “fungibility”, which he repeated insome variation eight times. From his prepared remarks: 我对这些财报的发布有点迟了，但微软的股价仍略有下跌，⾸席执⾏官萨蒂亚·纳德拉似乎预料到了投资者的担忧。他在财报电话会议上的讲话归结为⼀个词：“可互换性”，他以不同的形式重复了⼋次。以下是他的准备发⾔： We have the most expansive data center fleet for the AI era, and we are adding capacity at an unprecedentedscale. We will increase our total AI capacity by over 80% this year and roughly double our total data centerfootprint over the next 2 years, reflecting the demand signals we see. Just this quarter, we announced theworld’s most powerful AI data center, Fairwater in Wisconsin, which will go online next year and scale to 2gigawatts alone. And we have deployed the world’s first large-scale cluster of Nvidia GB300s. 我们拥有⼈⼯智能时代最庞⼤的数据中⼼集群，并且正在以前所未有的规模增加容量。今年我们将把⼈⼯智能总容量增加80%以上，并在未来两年内将数据中⼼总占地⾯积扩⼤⼀倍左右，这反映了我们所看到的市场需求信号。就在本季度，我们宣布了全球最强⼤的⼈⼯智能数据中⼼——位于威斯康星州的Fairwater，它将于明年投⼊运营，仅其⾃身就能扩展到2吉瓦。我们还部署了全球⾸个⼤规模的英伟达GB300集群。 We are building a fungible fleet that’s been continuously modernized and spans all stages of the AI life cycle,from pretraining to post training, to synthetic data generation and inference. And it also goes beyond GenAIworkloads to recommendation engines, databases and streaming. We’re optimizing this fleet across siliconsystems and software to maximize performance and e"ciency. 我们正在构建⼀个可互换的集群，它不断现代化，涵盖了⼈⼯智能⽣命周期的所有阶段，从预训练到后训练，再到合成数据⽣成和推理。它还超越了⽣成式⼈⼯智能⼯作负载，扩展到推荐引擎、数据库和流媒体。我们正在优化这个集群的芯⽚系统和软件，以最⼤限度地提⾼性能和效率。 It’s this combination of fungibility and continuous optimization that allows us to deliver the best ROI and TCOfor us and our customers. For example, during the quarter, we increased the token throughput forGPT-4.1 andGPT-5, two of the most widely used models, by over 30% per GPU. 正是这种可替代性和持续优化的结合，使我们能够为⾃⼰和客户提供最佳的投资回报率和总拥有成本。例如，在本季度，我们将GPT-4.1和GPT-5（两种最⼴泛使⽤的模型）的每个GPU的令牌吞吐量提⾼了30%以上。 This is, in many respects, the same argument thatMeta CEO Mark Zuckerberg was trying to make on Meta’searnings call: there are so many applications for GPUs that Microsoft would in fact be derelict in not buying more;moreover, the way to gain maximum leverage on its R&D is to have an ever larger fleet from which to wringcontinual e"ciency gains. 这在许多⽅⾯与Meta⾸席执⾏官⻢克·扎克伯格在Meta财报电话会议上试图提出的论点相同：GPU有如此多的应⽤，以⾄于微软实际上如果不购买更多GPU就是失职；此外，要最⼤限度地发挥其研发的杠杆作⽤，就必须拥有⼀个不断扩⼤的GPU集群，从中不断提⾼效率。 What is interesting is the pecking order when it comes to using those GPUs.