Microsoft unveiled its third-generation Maia 300 AI chip on Monday and said the Maia family will become its primary inference system for products and services starting in 2027.
The company plans to scale Maia deployments to gigawatt levels. Reuters reported that the rollout could involve more than 1 million Maia 300 chips next year.
Microsoft has described the initiative as a way to reduce costs by running its AI services, including Copilot, on internally designed chips rather than processors purchased from Nvidia and AMD.
Microsoft presented Maia 300 at the OCP Global Summit in San Jose. It said the chip delivers 40% more tokens per megawatt than “leading GPU-based systems” but did not identify the systems used for comparison.
A full Maia 300 rack contains 96 chips and consumes about 167 kilowatts. Microsoft said it can produce up to seven times the token throughput of a Maia 200 rack.
Reuters noted that Nvidia’s GB300 racks contain 72 GPUs and consume about 142 kilowatts.
Each Maia 300 chip has eight stacks of HBM4 memory, 288 GB of memory capacity and 18 TB/s of memory bandwidth within a 950-watt power budget. Its compute chiplets are manufactured using TSMC’s N2P process.
Microsoft said Maia 300 is designed for inference, the operation of AI models after training. It targets long-context reasoning, conversational AI and agents that perform multistep tasks.
“It’s basically delivering more intelligence for every watt, every dollar, and every byte of memory by optimizing across the entire system,” Rani Borkar, president of Azure hardware systems and infrastructure, told Reuters.
Maia 200 is already operating in Microsoft data centers, supporting Microsoft Copilot and running OpenAI models, Microsoft’s MAI models and open-source models. GitHub Copilot and Microsoft Foundry are due to move onto Maia 200 by year-end.
Microsoft said Maia 200 delivers up to 1.7 times more tokens per watt than leading GPU infrastructure for AI features in Teams and Copilot.
In August, The Information reported that Microsoft was in talks with TSMC to manufacture more than 300,000 Maia 300 chips for 2027.
Amazon, Google parent Alphabet, Meta, OpenAI and Anthropic are also developing their own chips. Nvidia regularly receives more than half of its quarterly revenue from hyperscalers, according to Yahoo Finance.
Microsoft shares rose during Monday trading, while Nvidia and AMD shares declined.
Analysts cautioned that large cloud companies are unlikely to stop buying Nvidia chips entirely because customers need different processors for different workloads and Nvidia’s chips remain the most in demand.
Microsoft plans to provide more details at its Ignite conference in San Francisco, running November 17 to 20.
Featured image credit
