Current outcome
Microsoft announced its third-generation Maia 300 AI chip, positioned as the primary inference system for its own products and services, with plans to scale to multi-gigawatt deployments from 2027 to reduce reliance on Nvidia GPUs; model training continues on Nvidia and AMD. New details: over 1 million Maia 300 units may be deployed next year; compute dies use TSMC 2nm process, each with 288GB HBM4 memory and 18TB/s bandwidth; each rack holds 96 chips at about 167kW, with inference throughput up to 7x Maia 200 and ~40% more tokens per MW than leading GPU systems; previous-generation Maia 200 is already running in Microsoft Copilot.
Progress timeline
3 material updates- #01
Microsoft’s new AI chip makes it less reliant on Nvidia’s GPUs
Microsoft announced its third-generation Maia 300 AI chip, positioned as the primary inference system for its own products and services with improved cost efficiency and support for reasoning and agentic workloads, reducing its reliance on Nvidia GPUs.
Source evidence: The Verge AI
- #02
微软将加大自研AI芯片使用，Maia 300预计2027年投产
New details: Maia 300 expected to enter production in 2027; multi-gigawatt deployment on Maia series; each data center rack can hold 96 Maia 300 chips, double the previous Maia 200; up to 7x AI token throughput per rack vs GPU and ~40% more tokens per MW; Evercore ISI estimates Maia chips will account for ~10% of Microsoft AI chip spending by Q1 2027; model training still uses Nvidia and AMD chips; not yet available for rental by other enterprises.
State after update: Microsoft announced its third-generation Maia 300 AI chip, positioned as the primary inference system for its own products and services with improved cost efficiency and support for reasoning and agentic workloads, reducing its reliance on Nvidia GPUs; production targeted for 2027 and multi-gigawatt internal deployment planned, with model training continuing on Nvidia and AMD.
Source evidence: theblockbeats
- #03
微软要让自研芯片成为AI推理主力，Maia 300明年或部署超百万颗
New details: Microsoft may deploy over 1 million Maia 300 units next year; Maia 300 compute dies use TSMC 2nm process, each with 288GB HBM4 memory and 18TB/s bandwidth; each rack draws about 167kW; previous-generation Maia 200 is already running in Microsoft Copilot, supporting OpenAI, Microsoft and open-source models.
State after update: Microsoft announced its third-generation Maia 300 AI chip, positioned as the primary inference system for its own products and services, with plans to scale to multi-gigawatt deployments from 2027 to reduce reliance on Nvidia GPUs; model training continues on Nvidia and AMD. New details: over 1 million Maia 300 units may be deployed next year; compute dies use TSMC 2nm process, each with 288GB HBM4 memory and 18TB/s bandwidth; each rack holds 96 chips at about 167kW, with inference throughput up to 7x Maia 200 and ~40% more tokens per MW than leading GPU systems; previous-generation Maia 200 is already running in Microsoft Copilot.
Source evidence: theblockbeats