ACCORDING to Microsoft, Maia 200 was unveiled on January 26 as the company’s second-generation AI processor, built on TSMC’s 3nm process and pitched as its most formidable inference system to date. The chip houses over 140 billion transistors on a single die and delivers more than 10 petaFLOPS in FP4 precision and more than 5 petaFLOPS in FP8 precision, while its thermal design power is kept below 750W and it ships with 216GB of HBM3e memory capable of 7 TB/s.
Microsoft notes Maia 200 offers a 30% improvement in performance-per-dollar over its predecessor and aims to challenge rivals in the inference market rather than targeting NVIDIA’s training GPUs. Initial deployment has begun in Microsoft’s Iowa data centres to support the internal Superintelligence Team and to back services like Copilot and heavyweight models such as OpenAI’s GPT-5.2, with the broader strategy described as about autonomy and reducing hardware costs for Azure’s scale.
The article also states that Microsoft has shifted away from NVIDIA’s InfiniBand in favour of Ethernet for interconnectivity, signalling an intention to fracture NVIDIA’s ecological monopoly.