Microsoft has unveiled the Maia 300, its third-generation AI accelerator designed for inference tasks within the company's own infrastructure. The chip aims to improve energy efficiency, with the company targeting a 40 percent increase in tokens per megawatt compared to current GPU-based systems.
Production of the Maia 300 is scheduled to begin in 2027. Reports suggest Microsoft is currently in negotiations with TSMC to manufacture more than 300,000 units for the initial rollout, with long-term plans to scale production to over a million chips.
The previous generation, Maia 200, was released in January 2026 and produced in quantities of several tens of thousands. That model utilized a 3nm process, 216GB of HBM3e memory, and offered a 30 percent improvement in performance per dollar over earlier hardware.
Microsoft is exploring the possibility of offering Maia chips to external customers, with discussions reportedly involving companies like Anthropic. While no formal agreements have been signed, such a move would signal a shift toward competing directly in the AI hardware market.
