微软将在多年依赖英伟达图形处理单元之后，加大使用其自研人工智能芯片的力度，因为这家软件巨头正试图强化其人工智能技术栈。
Microsoft will ramp up use of its homegrown artificial intelligence chips after years of reliance on Nvidia's graphics processing units, as the software giant attempts to bolster its AI stack.
微软表示，从周一发布的第三代Maia人工智能加速器开始，它将主要依赖自研硅片来处理传入的请求。
Starting with its third-generation Maia AI accelerator, unveiled on Monday, Microsoft said it will mainly rely on its own silicon to handle incoming requests.
微软在人工智能领域的最大优势在于其Azure云，该云为客户运行一系列模型。其Microsoft 365 Copilot和GitHub Copilot为软件工程师和其他办公室工作人员提供人工智能驱动的服务。此外，该公司还开发内部模型，尽管这些模型尚未在由OpenAI、Anthropic、Google主导的市场中掀起大波澜。芯片领域的新兴机遇源于对先进处理器的持续需求以及市场对选择的渴望。
Microsoft's greatest strength in AI has been its Azure cloud, which runs a series of models for clients. Its Microsoft 365 Copilot and GitHub Copilot offer AI-powered services to software engineers and other office workers. And the company develops in-house models, though they've yet to make a big splash in a market dominated by OpenAI Anthropic Google The emerging opportunity in chips stems from unrelenting demand for advanced processors and a desire in the market for choice.
英伟达的GPU一直是人工智能工作负载和模型的标准，它们填充了微软现有的数据中心，同时也用于训练来自OpenAI和Anthropic的模型。
Nvidia's GPUs have been the standard for AI workloads and models, and they fill Microsoft's existing data centers, while also serving to train models from OpenAI and Anthropic.
但随着需求持续增长，模型领导者们已经接受了来自包括亚马逊、超微半导体、Cerebras、谷歌以及最近几个月的微软在内的公司的多种硅片。
But as demand keeps growing, the model leaders have embraced a variety of silicon from companies including Amazon Advanced Micro Devices Cerebras Google and, in recent months, Microsoft.
对于通用用途，微软已经在使用其自身的中央处理单元以及来自英特尔的硅片。微软财务总裁艾米·胡德上个月在高盛大会上表示，使用其自身芯片有助于在十年期间实现“利润率提升”。微软于2023年推出了其首款人工智能芯片Maia 100，并在今年1月推出了Maia 200。微软首席执行官萨提亚·纳德拉在7月告诉分析师，Maia 200“每美元性能比我们机队中最新一代硬件高出30%”，并表示它正在处理OpenAI的模型。
For general-purpose use, Microsoft already taps its own central processing units as well as silicon from Intel Finance chief Amy Hood said at a Goldman Sachs conference last month that use of its own chips has helped with "margin improvement over the course of a decade." Microsoft introduced its inaugural AI chip, the Maia 100, in 2023 It touted the Maia 200 this past January and CEO Satya Nadella told analysts in July that Maia 200 "delivers 30% better performance per dollar than the latest generation hardware in our fleet" and said it was handling OpenAI models.
Maia 300 的生产将于 2027 年开始。根据执行副总裁斯科特·古斯里（Scott Guthrie）的一篇博客文章，微软正瞄准在 Maia 系列芯片上进行多吉瓦级别的部署。但总经理阿利斯泰尔·斯皮尔斯（Alistair Speirs）拒绝提供时间表。
Production of the Maia 300 starts in 2027. Microsoft is targeting multi-gigawatt deployments of chips in the Maia family, according to a blog post from Executive Vice President Scott Guthrie. But Alistair Speirs, a general manager, declined to provide a timeline.
“我们在那里保留了一点选择余地，因为这个行业变化非常快，我们希望在合适的时间部署适合我们当时所使用模型的合适硅片，”斯皮尔斯说。
"We reserve a little bit of optionality there, because this is an industry that's changing very fast, and we want to deploy the right silicon for the right models that we're using at the right time as well," Speirs said.
微软预计将在每个数据中心机架中装入 96 颗 Maia 300 芯片，这一数量是 Maia 200 的两倍。新版本每机架可处理的 AI token 数量最高可达 GPU 的七倍，且每兆瓦的 token 数量比 GPU 高出 40%。
Microsoft expects to pack 96 Maia 300 chips into each data center rack, double the number for the Maia 200. The new version can run through up to seven times more AI tokens per rack, with 40% more tokens per megawatt when compared with GPUs.
“我们需要新的架构，以实现基本上更好的性能和更好地利用我们现有的电力，”斯皮尔斯说。“你可以把电力视为数据中心等设施的直接运营成本。”公司技术研究员吉姆·克莱韦因（Jim Kleewein）表示，Microsoft 365 Copilot 的部分功能将因 Maia 300 而提速。
"We need new architectures that will give us essentially better performance, better use of the power that we have," Speirs said. "You can think of power translating pretty directly into operational cost for something like a data center as well." Parts of Microsoft 365 Copilot should speed up with Maia 300, said Jim Kleewein, a technical fellow at the company.
在周三分发给客户的一份备忘录中，Evercore ISI 分析师估计，到 2027 年第一季度，Maia 将占微软在 AI 芯片上的支出的约 10%。分析师建议买入微软股票。
In a note distributed to clients on Wednesday, Evercore ISI analysts estimated that Maia will represent about 10% of Microsoft's spending on AI chips by the first quarter of 2027. The analysts recommend buying Microsoft shares.
英伟达仍将是该方程式中的重要部分。上周，英伟达首席执行官黄仁勋（Jensen Huang）在微软宣布一款搭载英伟达处理器的 Surface 笔记本电脑的活动上发表了讲话。斯皮尔斯还说，在模型训练方面，微软与英伟达和 AMD 合作。
Nvidia will still be a big part of the equation. CEO Jensen Huang spoke at an event last week where Microsoft announced a Surface laptop containing an Nvidia processor. And when it comes to training models, Microsoft partners with Nvidia and AMD, Speirs said.
与谷歌的张量处理单元或亚马逊的 Inferentia 和 Trainium 不同，微软的加速器无法按小时出租给其他公司。
Unlike Google's tensor processing units or Amazon's Inferentia and Trainium, Microsoft's accelerators can't be rented by the hour to other companies.
“到目前为止，尚没有公司在微软之外对它们进行过测试，”咨询公司 Moor Insights and Strategy 的首席执行官帕特里克·穆尔黑德（Patrick Moorhead）说。梅丽莎·奥托：英伟达是 AI 贸易的风向标。
"Nobody has tested them outside of Microsoft yet," said Patrick Moorhead, CEO of advisory firm Moor Insights and Strategy. Melissa Otto: Nvidia is the bellwether of the AI trade