Microsoft to Scale Maia AI Chips to Gigawatts by 2027

Microsoft plans to deploy its proprietary Azure Maia AI processors at a multi-gigawatt scale beginning in 2027. Executive Scott Guthrie revealed that the custom hardware will serve as the company's primary inference architecture across its enterprise and consumer products. CEO Satya Nadella framed the effort as "our biggest bet yet on our own silicon." Currently, the Maia 200 chip already powers Microsoft Copilot in the United States, Europe, and Australia, with integration into GitHub Copilot and Foundry set to follow by the end of the year.

The hardware is built to handle workloads from OpenAI, MAI, and diverse open-source models. According to Microsoft, the chip delivers up to 1.7 times more tokens per watt than leading GPUs when running internal workloads such as Microsoft Teams.

The roadmap centers on its successor, the Maia 300, scheduled to ship in 2027. Tailored for long context inputs, low-latency chat, and complex agentic tasks, each Maia 300 accelerator integrates eight HBM4 memory stacks for a total of 288 GB of memory and 18 TB/s of bandwidth, rated at 950 watts. Microsoft claims the chip achieves 40 percent more tokens per megawatt than top GPU systems. Furthermore, a 96-chip rack of Maia 300 units is designed to yield up to seven times the token throughput of a standard Maia 200 rack.

For developers and cloud practitioners, Microsoft's massive custom silicon commitment indicates a long-term transition toward specialized infrastructure for AI inference. By tuning hardware specifically for massive context windows and agent orchestration, Azure aims to lower operational costs while improving latency across its AI portfolio. Further architectural details were outlined in a recent technical report, with additional disclosures planned for November at Ignite 2026.