Microsoft Maia 300 inference accelerator

Microsoft CEO Satya Nadella announced the Maia chip family will become the primary inference system for Microsoft products at multi-gigawatt scale, with deployments starting in 2027. The Maia 300, a third-generation accelerator, is designed for long-context reasoning, low-latency chat, and multi-agent tasks. Microsoft claims it delivers 40% more tokens per megawatt than leading GPU-based systems and 7x the token throughput of a Maia 200 rack. The chip uses a multi-chiplet design that separates compute, memory, and I/O onto specialized silicon. Maia 200 already serves Microsoft Copilot and Teams, delivering up to 1.7x more tokens per watt than GPU infrastructure according to Microsoft. Microsoft is reportedly negotiating with TSMC for over 300,000 Maia 300 units for 2027 delivery, with no purchase order signed. Korean media reporting cited by a recognized source projects Maia 300 shipments at over 1 million units, each carrying eight HBM4 stacks totaling 288 GB. Microsoft, Google, Amazon, and Meta are all reported to be accelerating in-house AI chip development, with SK Hynix already supplying HBM3E for Maia 200 and Samsung seeking a larger share of the custom chip market.