微软发布自研AI芯片Maia 300 拟2027年投产推动大规模降本Microsoft Unveils Custom Maia 300 AI Chip Targeting 2027 Production

微软周一公布其第三代自研人工智能加速器Maia 300，计划在2027年正式投产。这款新芯片将主要承担微软AI服务日常接收的大量推理请求，从而逐步扩大自主硬件在数据中心运算中的实际占比。在行业对先进计算处理器需求持续高企的环境下，微软正试图通过加大自研芯片的使用范围，降低对长期主导市场的外部图形处理器的单独依赖。

Microsoft introduced its third-generation custom artificial intelligence accelerator, the Maia 300, on Monday, scheduling production to commence in 2027. The upcoming silicon is designed to primarily field incoming user and enterprise requests across the company's expanding suite of AI-powered services. With market demand for high-end silicon remaining strong, the software giant is ramping up deployment of homegrown processors to handle workloads that historically relied almost exclusively on Nvidia graphics processing units across cloud infrastructure.

针对未来部署规模，微软执行副总裁斯科特·格思里在官方博客中确认，公司正瞄准数吉瓦级别的Maia系列芯片基础设施建设。不过，微软总经理阿利斯泰尔·斯皮尔斯并未对该部署规划给出固定时间表。他解释称，由于人工智能行业演变极为迅速，微软需要保留业务选择权，以便在恰当的时间根据所运行的模型匹配最适合的芯片算力。

Detailing the planned footprint, Executive Vice President Scott Guthrie noted in a company blog post that Microsoft is targeting multi-gigawatt deployments across the broader Maia accelerator family. However, General Manager Alistair Speirs declined to specify a fixed timeline for reaching that capacity. Speirs stated that the company is deliberately reserving operational flexibility in a rapidly shifting industry, allowing infrastructure teams to deploy the appropriate silicon for specific models at the right time.

在硬件物理架构设计上，微软预计每个数据中心机架将容纳96颗Maia 300芯片，其安装密度达到前代产品Maia 200的两倍。得益于这种更高的集群密度，该芯片每个机架处理人工智能Token的数量预计将提升高达7倍。这一吞吐量增长使服务器能够在单位机房空间内承载更密集的计算任务。

In physical server architecture, Microsoft expects to pack 96 Maia 300 chips into each data center rack, doubling the installation density achieved by the Maia 200 generation. Due to this denser hardware configuration, the company projects that the new accelerator will process up to seven times more artificial intelligence tokens per rack, substantially expanding server compute throughput within existing footprint constraints.

除了机架吞吐表现，能耗控制也是该芯片设计的关键重点。微软测算显示，与图形处理器相比，Maia 300在每兆瓦电力供应下能够多产生40%的Token。微软总经理阿利斯泰尔·斯皮尔斯表示，数据中心需要能够带来更优性能并更好利用现有电力的新架构，因为电力消耗会直接转化为数据中心的实际运营成本。

Beyond spatial density, power efficiency forms a central focus of the new silicon design. Microsoft calculations show that the Maia 300 delivers 40% more tokens per megawatt when compared directly with graphics processing units. Microsoft General Manager Alistair Speirs explained that facilities require new system architectures to extract superior performance from constrained power envelopes, pointing out that power consumption directly drives ongoing operational costs across large-scale data center facilities.

微软在自研算力领域的探索始于2023年推出的首款人工智能芯片Maia 100，随后在今年1月公布了Maia 200。首席执行官萨提亚·纳德拉曾在7月向分析师表示，Maia 200相比微软数据中心现有最新一代硬件实现了30%的单位成本性能提升，并且已经开始承担OpenAI相关模型的运算任务。此外，微软在通用计算领域也早已采用自研中央处理器，与英特尔和超威半导体的硬件协同运行。

Microsoft began building its custom artificial intelligence silicon pipeline with the release of the Maia 100 in 2023, followed by the announcement of the Maia 200 this past January. Chief Executive Officer Satya Nadella told analysts in July that the Maia 200 delivered 30% better performance per dollar than the latest generation hardware operating in the company's server fleet, noting that it was actively processing OpenAI workloads. For general-purpose infrastructure workloads, Microsoft has likewise deployed homegrown central processing units alongside processors sourced from Intel and Advanced Micro Devices.

自主研发硬件直接服务于微软控制成本与强化产品体验的目标。首席财务官艾米·胡德上个月在高盛举办的会议上指出，使用自研芯片帮助公司在十年间实现了利润率改善。在终端应用层面，微软技术院士吉姆·克莱温表示，随着Maia 300投入使用，微软365 Copilot办公助手的相关功能运行速度预计将得到加快。

The internal silicon push aligns directly with company financial goals and core product performance. Chief Financial Officer Amy Hood said at a Goldman Sachs conference last month that deploying proprietary chips has contributed to margin improvement over the course of a decade. On the software front, Jim Kleewein, a technical fellow at Microsoft, confirmed that specific capabilities within the Microsoft 365 Copilot productivity suite should run faster when powered by the upcoming Maia 300 hardware.

在整体硬件采购规划中，投行Evercore ISI分析师在发给客户的报告中预估，到2027年第一季度，自研Maia系列将占微软人工智能芯片总开支的约10%，该机构同时维持对微软股票的买入评级。尽管自研算力正在逐步铺开，外部芯片制造商仍将在微软数据中心版图中占据重要地位。英伟达首席执行官黄仁勋上周出席了微软发布搭载英伟达芯片的Surface笔记本电脑活动，而微软总经理阿利斯泰尔·斯皮尔斯也证实，在复杂的人工智能模型训练场景下，微软将继续与英伟达和超威半导体保持紧密合作。

Assessing the financial impact on hardware procurement, Evercore ISI analysts estimated in a note distributed to clients on Wednesday that Maia silicon will account for approximately 10% of Microsoft's total AI chip spending by the first quarter of 2027, with the investment firm maintaining a buy recommendation on Microsoft shares. Even as homegrown processors expand, merchant semiconductor vendors continue to play a foundational role in the software giant's infrastructure. Nvidia Chief Executive Officer Jensen Huang appeared at an event last week announcing a Surface laptop equipped with an Nvidia processor, and General Manager Alistair Speirs confirmed that Microsoft continues to partner with both Nvidia and Advanced Micro Devices for computationally intensive model training.

与其他主要云服务巨头的商业模式不同，微软的自研人工智能加速器目前并未单独向外部企业提供按小时付费的算力租赁服务，这与谷歌的张量处理器以及亚马逊的Inferentia和Trainium芯片形成鲜明对比。咨询机构Moor Insights and Strategy首席执行官帕特里克·穆尔黑德指出，除微软内部团队之外，外部目前尚无其他机构实际测试过这些自研加速器。这种架构部署体现出微软现阶段将自研硬件优先服务于内部云平台与自主产品矩阵的运营导向。

Regarding market availability, Microsoft's proprietary AI accelerators diverge from offerings deployed by rival hyperscalers. Unlike Google's tensor processing units or Amazon's Inferentia and Trainium chips, Microsoft does not rent its Maia hardware by the hour to external corporate customers. Patrick Moorhead, chief executive officer of advisory firm Moor Insights and Strategy, observed that nobody outside of Microsoft has tested the processors yet. The company's deployment model keeps the custom hardware focused on powering its internal cloud infrastructure and native product ecosystem rather than operating as standalone merchant compute capacity.