Microsoft bets on its own AI chips with next-gen Maia 300, CEO Nadella calls it a major shift

Microsoft is moving deeper into designing its own hardware to power artificial intelligence, indicating that it expects its next-generation Maia 300 system to improve the economics of AI computing.

In a market report published Tuesday, Microsoft CEO Satya Nadella described the Maia 300 effort as the company’s “biggest bet yet on our own silicon.” The framing points to a strategy that extends beyond software and cloud services into custom chip and system design for AI data centers.

According to the report, Microsoft expects the Maia 300 system to deliver 40% more AI-generated tokens per megawatt than “leading GPU-based systems.” “Tokens” are the basic units of text or output that large language models generate, so higher tokens per megawatt is a measure of how much AI output a system can produce for each unit of power.

The report also suggests that Microsoft views performance per watt as a central metric as demand for AI training and inference continues to strain power availability and increase operating costs. By targeting improved output for the same energy consumption, the company is effectively trying to make AI services more scalable as deployments expand.

Microsoft’s approach reflects a broader industry push toward specialized hardware for AI. Large language model workloads can be computationally intensive, and power efficiency can matter as much as raw speed when data centers compete for electricity and cooling capacity.

Even so, the post does not disclose additional specifics that investors and customers typically look for, such as measured latency, model quality outcomes, compatibility details with Microsoft cloud services, or the size of the performance gap under different workload types. It also does not provide a timeline for broad deployment or how quickly Maia 300 would replace or supplement existing GPU-based stacks.

In the absence of more granular benchmarks and rollout details in the cited report, the main takeaway is directional: Microsoft is treating its own silicon as a lever to improve unit economics for AI at scale. The 40% tokens-per-megawatt target, if validated across real-world deployments, would represent a meaningful advantage for companies running large volumes of inference.

What to watch next is whether Microsoft provides additional benchmark methodology, workload breakdowns, and deployment milestones for Maia 300. Further clarity on how the system integrates with Microsoft’s cloud infrastructure would also help determine how quickly customers can benefit from the efficiency gains.