Microsoft's Maia 200 AI Chip Powers Microsoft 365 Copilot
Microsoft's Maia 200 AI chip is now powering Microsoft 365 Copilot.
Why it matters: Legal teams evaluating AI tools should note this shift to proprietary chips, which could lower costs and improve performance of AI-driven platforms.
- Maia 200 has over 140 billion transistors, built on TSMC's 3nm process for efficiency.
- It delivers over 10 petaFLOPS FP4 (4-bit) and 5 petaFLOPS FP8 (8-bit) AI compute within 750W power.
- Microsoft says Maia 200 provides 30% better performance per dollar than previous hardware.
- A more advanced Maia 300 chip, optimized for long-context AI, is planned for 2027.
Microsoft has launched its own AI accelerator, the Maia 200 chip, which now powers Microsoft 365 Copilot as well as Teams and Cowork across multiple regions including the US, Europe, and Australia.
Built on Taiwan Semiconductor Manufacturing Company’s advanced 3nm process, Maia 200 contains more than 140 billion transistors, enabling high efficiency in AI computations. The chip includes 216GB of HBM3e memory—a high-speed memory type—providing 7 TB/s bandwidth, and features 272MB of fast on-chip SRAM.
In terms of AI performance, Maia 200 delivers over 10 petaFLOPS using FP4 (4-bit floating point) precision and 5 petaFLOPS using FP8 (8-bit floating point) within a 750-watt power budget, according to Tom's Hardware. FP4 and FP8 are lower-precision formats optimized for faster AI calculations with lower energy use, a key consideration for large-scale cloud AI.
Scott Guthrie, Microsoft's Executive Vice President of Cloud + AI, describes Maia 200 as their "most efficient inference system" so far, citing a 30% improvement in performance per dollar compared to earlier hardware. This efficiency helps Microsoft reduce costs and gain better control over AI service delivery.
Despite Maia 200’s deployment, Microsoft continues to source AI accelerators from Nvidia and AMD, maintaining hardware diversity, as noted by TechRadar.
Looking ahead, Microsoft plans to introduce the Maia 300 chip in 2027, which is designed to handle long-context reasoning and reduce latency for conversational AI, according to a Microsoft Blog post. This indicates ongoing investment in bespoke AI hardware innovation.
For legal tech professionals and in-house counsel, these developments highlight an evolving AI infrastructure that could influence AI adoption strategies, platform costs, and vendor choice in legal applications.
By the numbers:
- 140 billion+ transistors — inside Maia 200 chip built on TSMC 3nm process
- 10 petaFLOPS FP4 & 5 petaFLOPS FP8 — AI compute performance within 750 watts
- 30% better performance per dollar — claimed efficiency improvement over previous hardware
Yes, but: Microsoft still relies on Nvidia and AMD chips alongside Maia 200, so hardware diversification continues.
What's next: Microsoft plans to release Maia 300 in 2027, targeting long-context AI and reduced latency for conversational applications.