Microsoft on October 12 disclosed its third-generation custom AI accelerator, Maia 300, positioning the chip as the "primary AI inference system" for the company's products and services starting in 20271.
Maia 300 is built on Taiwan Semiconductor Manufacturing Co.'s N2P process node2. Microsoft expects the chip to deliver 40 percent more tokens per megawatt than leading GPU-based systems, according to Scott Guthrie, executive vice president of Microsoft's cloud group. In a full 96-accelerator configuration, Maia 300 delivers up to 7x the rack-scale token throughput of Maia 200, its predecessor.
Microsoft plans to scale its Azure Maia family to "multi-gigawatt deployments starting in 2027," Guthrie said, a move that would make Maia the default inference platform across Microsoft's cloud and product stack. The chip has been designed to handle reasoning models and agentic workloads, not just conventional inference.
ANALYSIS The efficiency claim, 40 percent more tokens per megawatt versus GPU-based alternatives, frames the value proposition in operational-cost terms rather than raw performance, consistent with inference workloads where power and cooling budgets constrain scale.
The efficiency gains are intended to let Microsoft rely primarily on its in-house silicon for inference rather than chips from Nvidia Corp. and AMD. Production rollout to live AI workloads begins in 2027.
ANALYSIS Designating Maia 300 as the "primary" inference system does not eliminate GPU procurement, but it shifts the default: third-party accelerators move from baseline to supplement for Microsoft's inference fleet. The practical effect is a structural reduction in per-token dependence on external silicon suppliers.