微软揭开了其 Azure Maia 定制 AI 芯片下一代产品的面纱，称其将开启计算基础设施的新阶段，届时所有组件将紧密集成，以满足代理推理的需求。这家雷德蒙巨头表示，今年早些时候宣布的 Azure Maia 200 已经开始服务生产工作负载，而其继任产品已规划于 2027 年推出。Azure Maia 300 被描述为“一个完整的智能系统”，专为当下新兴的 AI 工作负载类型而设计。
Microsoft has lifted the lid on the next generation of its Azure Maia custom AI chips, saying it will usher in a new phase of its compute infrastructure where everything is closely integrated to meet the demands of agentic inference. The Redmond giant says its Azure Maia 200, announced earlier this year, is now serving production workloads and that a successor is already planned for 2027. Azure Maia 300 is described as "a complete intelligent system" designed for the new types of AI workloads emerging today.
微软声称，完整的 96 加速器配置将提供高达 Maia 200 机架规模令牌吞吐量的七倍，同时相比领先的基于 GPU 的系统，每兆瓦提供的令牌数多出 40%。根据微软云与 AI 执行副总裁斯科特·古思里的一篇博客文章，其下一代芯片采用以推理为先的架构进行优化，以适应长上下文、内存密集型和解码密集型的 AI 工作负载。
Microsoft claims a full 96-accelerator configuration will deliver up to seven times Maia 200's rack-scale token throughput, while offering 40 percent more tokens per megawatt than leading GPU-based systems. Its next-gen chip is created with an inference-first architecture optimized for long-context, memory-intensive and decode-heavy AI workloads, according to a blog post by Microsoft's EVP for Cloud and AI, Scott Guthrie.
该公司设想将其部署在规模可达 96 个加速器的机架级系统中，而每个 Maia 300 将配备八个 HBM4 内存堆栈，提供高达 18 TB/s 的峰值内存带宽，以更好地支持大模型和长上下文工作负载的执行。这听起来类似于英伟达此前为其 GPU 加速器开发的那类机架级 AI 设备，以及 AMD 最近通过其 Helion 和 Helios 基础设施所推出的方案，所有这些方案均旨在充分发挥各自专属的硅片和软件栈的优势。
The firm envisions it being deployed in rack-scale systems comprising up to 96 accelerators, while each Maia 300 will feature eight HBM4 memory stacks providing up to 18 TB/s of peak memory bandwidth, the better to support execution of large models and long-context workloads. This sounds similar to the kind of rack-scale AI equipment that has already been developed by Nvidia for its GPU accelerators, and more recently by AMD with its Helion and Helios infrastructure, all designed to make the most of their own specific silicon and software stacks.
微软表示，Maia 300是一个“完整的系统”：其软件、硬件（包括服务器芯片）、网络基础设施以及数据中心都是经过共同设计的，旨在作为一个统一的平台来满足大规模云计算场景下的推理需求。该公司强调，这种系统级的集成对于那些需要多步骤执行、复杂数据处理以及工具集成的应用至关重要；这种集成方式会显著改变推理任务的计算需求和算术复杂度。
Microsoft says that Maia 300 is "built as a complete system, with software, silicon, physical rack infrastructure, networking, and datacenters co-designed to operate as a single platform to meet the demands of cloud-scale agentic inference." It insists that this system-level approach is necessary for agentic workloads, which involve multi-step execution, long-context decoding and tool use, and "change the computational profile and arithmetic intensity of inference."
这种日益重视内部整合的趋势，与微软在2024年提出的“AI Access Principles”（即“AI使用原则”）形成了鲜明对比——该原则提倡在微软的技术栈中建立合作伙伴关系。微软甚至向欧盟委员会的竞争监管机构投诉称，谷歌因其这种内部整合策略而具有竞争优势。微软当时表示：“只有谷歌一家公司在从芯片到移动应用商店的整个AI产业链上都实现了垂直整合，从而具备了强大的竞争力和独立性；其他所有公司都必须依赖合作伙伴关系来推动创新和竞争。”
The growing emphasis on in-house integration contrasts with Microsoft's 2024 "AI Access Principles," which stressed partnerships across its technology stack. The firm even complained to the European Commission's competition authorities that Google had a competitive advantage because of its approach. "Only one company – Google," said Microsoft at the time, "is vertically integrated in a manner that provides it with strength and independence at every AI layer from chips to a thriving mobile app store.
如今，微软已经拥有了自己的AI芯片、适用于大规模部署的服务器架构，以及自主研发的AI模型，这使它对自己的技术栈拥有更大的控制权。该公司计划从2027年开始将Azure Maia平台扩展到“多吉瓦级别的应用场景”。我们向微软寻求了评论，但微软方面表示将在下个月的Microsoft Ignite 2026活动中（该活动将在旧金山举行）公布更多关于其战略的细节。
Everybody else must rely on partnerships to innovate and compete." Microsoft now has its own AI chips, rack-scale designs, and internally developed MAI models, giving it more control over its stack. It plans to scale Azure Maia to "multi-gigawatt deployments" starting in 2027. We asked Microsoft for comment. Guthrie said that the firm will share more details on its strategy at next month's Microsoft Ignite 2026 event in San Francisco. ®