Newshot

Microsoft's In-House Maia 200 Chip Enters Production: 1.7x Higher Token-per-Watt Performance

AI summary

At Ignite 2026 on November 17, Microsoft announced that its second-generation in-house AI accelerator, Maia 200, has entered large-scale production and begun running Microsoft's largest frontier AI models in its data centers. Compared with "leading GPU infrastructure," the accelerator delivers up to 1.7x higher token throughput per watt in built-in AI features such as Microsoft Teams, Copilot, and Cowork. Maia 200 is a custom chip designed for AI inference, built on TSMC's 3nm process with FP4 and FP8 tensor cores, 216GB of HBM3e memory, and 272MB of on-chip SRAM; a single chip provides over 10 petaFLOPS at FP4 precision and scales via Ethernet to clusters of up to 6,144 accelerators. Microsoft said the chip has been deployed in the US, Europe, and Australia serving production workloads including Microsoft Copilot, and will gradually support GitHub Copilot and Microsoft Foundry; the company also plans to expand Azure Maia systems to multi-gigawatt deployments in 2027, making them Microsoft's primary inference infrastructure for its products and services.

Translation(model translation, not the source's own text)

Microsoft's in-house chip Maia 200 enters production: 1.7x improvement in token performance per watt

IT Home, November 18: At the Ignite 2026 event held yesterday (November 17), Microsoft announced that Maia 200, its second-generation in-house AI accelerator, has officially entered mass production and has begun running Microsoft's largest-scale frontier AI models in data centers.

Compared with "leading GPU infrastructure," the accelerator delivers up to 1.7x higher token throughput per watt in built-in AI features such as Microsoft Teams, Copilot, and Cowork.

IT Home note: Maia 200 is Microsoft's custom chip designed for AI inference, manufactured using TSMC's 3nm process, featuring FP4 and FP8 tensor cores, 216GB HBM3e memory, and 272MB on-chip SRAM; a single chip can deliver over 10 petaFLOPS at FP4 precision and scale over Ethernet to clusters of up to 6,144 accelerators.

Microsoft says the chip has been deployed in the United States, Europe, and Australia, serving production workloads such as Microsoft Copilot, and will gradually support GitHub Copilot and Microsoft Foundry. The company also plans to expand Azure Maia systems to multi-gigawatt-scale deployment in 2027, making them the primary inference infrastructure for Microsoft products and services.

Related articles

  • Microsoft's decision model Microsoft-Decision-1.1 debuts, outperforming GPT-6 Luna and Gemini
  • Microsoft and Apple outages in succession; Google urgently responds to Gmail outage rumors: "Service is running normally"
  • Multilingual error rate drops by 50%: Microsoft upgrades AI transcription engine for doctor-patient conversations, now covering 1/3 of physicians in the U.S.
  • Microsoft's first speech-to-speech model: MAI Realtime 1 debuts, starting at $0.01 per minute
  • Microsoft's most cost-effective cybersecurity AI model MAI Cyber 2 Flash debuts: performance comparable to GPT-6 Astra, with per-task cost about 2% of it
  • Microsoft announces PowerShell 7 will be provided by default with Windows; Win11 26H2 is the last version with 5.1 pre-enabled

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by IT之家
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗