微软宣布自研 Maia 芯片将成为旗下产品主要推理系统
微软要用自家 Maia 芯片替代 NVIDIA 和 AMD 做推理了，Maia 300 明年上，每兆瓦 token 多 40%，还能看它怎么拼芯粒设计。
Satya Nadella 宣布微软将把自研 Maia 芯片作为旗下产品的主要推理系统，规模扩展到多吉瓦级别，以降低对 NVIDIA 硬件的依赖。第三代加速器 Maia 300 计划 2027 年启用，微软称其每兆瓦 token 产出比领先 GPU 系统高 40%，整机架 96 颗加速器、167kW 功耗、27.6TB HBM，token 吞吐量达 Maia 200 机架的 7 倍。Maia 300 采用多芯粒设计，每颗芯片搭配 8 组 HBM4、峰值带宽 18.0 TB/s、TDP 950W，并通过增强 MXFP4 低精度格式压缩 KV cache 以支持长上下文。据报道微软正与 TSMC 洽谈 2027 年交付超过 30 万颗 Maia 300，目前仍在谈判中。现任产品 Maia 200 已服务于 Microsoft Copilot 等生产流量。
Satya Nadella officially announced that Microsoft will make its in-house Maia chips the primary inference system for its products, scaling to multiple gigawatts.
This will make it less reliant on NVIDIA hardware, and Maia 300 is meant to replace AMD chips too, not only Nvidia. The key claim is 40% more tokens per megawatt than leading GPU-based systems
A full rack packs 96 accelerators into 167kW with 27.6TB of HBM, and Microsoft claims it delivers up to 7x the token throughput of a Maia 200 rack.
The shift will start in 2027 with Maia 300, a third-generation accelerator built for long-context reasoning, low-latency chat and multi-step agent tasks.
Its predecessor, Maia 200, already serves production traffic including Microsoft Copilot and Microsoft claims Maia 200 produces up to 1.7x more tokens per watt than leading GPU infrastructure.
Maia 300 goes after the cost of moving data, which Microsoft says now limits large inference systems more than raw arithmetic does.
Maia 300 uses a “multi-chiplet” design, A traditional chip is one large slab of silicon that does everything: the math (compute), managing the memory (memory controller) and talking to other chips (I/O). Maia 300 instead splits those jobs into several smaller pieces of silicon, called chiplets, packaged together so they act as one chip. That leaves more room on the compute chiplets for the parts that actually run the AI math, and each chiplet can be made with whatever manufacturing process suits its job best.
Each chip pairs 8 HBM4 stacks delivering 18.0 TB/s of peak bandwidth with a 950W TDP, and uses on-die SRAM as a scratchpad to relieve that memory.
Support for low-precision formats such as an enhanced MXFP4 shrinks the KV cache, the stored attention state that grows with every token, so longer contexts fit within the same power budget.
Supply is the harder constraint, since Microsoft has reportedly been negotiating with TSMC for more than 300,000 Maia 300 units for 2027 delivery, a figure still under negotiation rather than a signed purchase order.
