Microsoft unveils Maia 300, plans multi-gigawatt inference buildout
What happened
On October 12, Microsoft announced Maia 300, its third in-house AI chip, planned for production from 2027 as the main inference system for its products and services, scaling Maia-family inference deployments to multiple gigawatts. Maia 200 is already deployed in the US, Europe and Australia, with up to 1.7x more tokens per watt in Teams and Copilot. Earlier reports said Maia 300 is expected to deliver 40% more tokens per megawatt than leading GPU systems; later reports said 40% more tokens per watt than mainstream GPU systems. In a full 96-accelerator configuration, rack-level throughput is up to 7x Maia 200. A 22:45 report added architecture details: 288GB HBM4 per accelerator, 18TB/s bandwidth, cross-die NUMA, enhanced MXFP4 and sub-3-bit storage formats, and a two-tier network connecting up to 12,288 accelerators. Multi-gigawatt deployment remains a plan.
Written by AI from the coverage · updated 52 minutes ago
Coverage
Follow the reports to see the story from different sides.
- Hacker News front pageDeep dive into Azure Maia 300 architecture
Microsoft 详解 Azure Maia 300 架构，通过软件、芯片、机架、网络与数据中心协同设计优化云规模智能体推理。每个加速器配备 288 GB HBM4 内存和 18 TB/s 带宽，采用跨芯粒 NUMA 内存组织、增强型 MXFP4 与 sub-3-bit 存储格式，双层网络可连接最多 12,288 个加速器。
- AI HOT · Products微软发布 Maia 300，计划明年起将 Maia 推理部署扩展至多吉瓦级
微软发布自研芯片 Maia 300，称其每瓦 token 产出比主流 GPU 系统高 40%。微软计划明年起将 Maia 芯片家族扩展到多吉瓦级部署，作为自家产品和服务的首选推理系统。Maia 200 已在美国、欧洲和澳大利亚部署，在 Teams 和 Copilot 场景中每瓦 token 产出提升最高 1.7x。
- The Verge · AIMicrosoft’s new AI chip makes it less reliant on Nvidia’s GPUs
微软宣布第三代自研 AI 芯片 Maia 300，计划从 2027 年开始投入生产部署，作为旗下产品与服务的主要 AI 推理系统。微软预计，相比领先的 GPU 系统，其每兆瓦 token 产出提高 40 percent；在完整的 96-accelerator 配置下，机架级 token 吞吐量最高达到 Maia 200 的 7x。
Heat over time
Heat now 18·Comparable peak 19（Oct 12 23:00）·Comparable change over 24 hours –
The trend only compares accounts observed without gaps, so its range may be smaller than the current heat. Hover or tap the chart for each hour; the left and right arrow keys step through it.