微软发布 Maia 300 定制 AI 推理芯片，每兆瓦 token 数提升 40%
Microsoft Unveils Maia 300 With 40% More Tokens Per Megawatt
微软发布下一代自研 AI 芯片 Maia 300，面向推理负载，预计 2027 年开始量产部署，官方称其每兆瓦 token 数比主流 GPU 系统高 40%。该芯片采用多芯粒架构，计算芯粒基于台积电 N2P 工艺，配备 8 组 HBM4 显存，峰值内存带宽 18 TB/s，TDP 为 950 瓦；在 96 加速器的配置下，机架级 token 吞吐量最高可达 Maia 200 的 7 倍。微软表示 Maia 200 已在 Copilot 等生产负载中运行，并部署于美国、欧洲和澳大利亚，GitHub Copilot 与 Microsoft Foundry 预计在 2026 年底前采用 Maia 200。
正文
当前显示原文，中文全文译文尚未就绪。
Microsoft has launched Maia 300, its next-generation custom AI chip designed to power inference workloads, with production deployments set to begin in 2027. The company said the chip is expected to deliver 40% more tokens per megawatt than leading GPU-based systems.
The company said that Maia 200 is already running production workloads, including Microsoft Copilot. Microsoft plans to make Maia its primary inference system for its products and services.
“We will scale our Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for Microsoft products and services,” Microsoft wrote in a blog post.
Microsoft said Maia 200 is being deployed across the US, Europe and Australia. The chip is running OpenAI models, Microsoft AI (MAI) models and open-source model families. GitHub Copilot and Microsoft Foundry are expected to adopt Maia 200 by the end of 2026.
Maia 300 is designed for inference workloads involving long-context reasoning, conversational AI and multi-step AI agents. Microsoft said the system would deliver up to seven times the rack-scale token throughput of Maia 200 in a configuration with 96 accelerators.
The company said Maia 200 delivers up to 1.7 times more tokens per watt for built-in AI capabilities in Microsoft Teams, Copilot and Cowork experiences compared with leading GPU infrastructure.
Maia 300 will use eight HBM4 memory stacks, providing peak memory bandwidth of 18 terabytes per second within a 950-watt thermal design power (TDP). The chip also uses a multi-chiplet architecture, with compute chiplets built on TSMC’s N2P process node.
Microsoft said the system is built to reduce data movement across compute, memory and networking resources, which it identifies as a bottleneck in large-scale AI inference.
The company said Maia 300 is designed to support larger context windows and higher throughput through low-precision formats, including enhanced MXFP4, which reduces memory and bandwidth requirements.
The Maia programme is part of Microsoft’s effort to develop its own silicon and reduce the cost of serving AI workloads at scale. By developing chips alongside its software and models, the company aims to optimise infrastructure for its own services and Azure customers.
Microsoft said Maia 300’s design was informed by its experience running models and services on Maia 200. The next-generation system is expected to support GitHub's coding tools, Microsoft Foundry's model-serving workloads, model development by the MAI team and Dragon Copilot's healthcare workflows.