Microsoft is making a major push into custom artificial intelligence hardware, with CEO Satya Nadella announcing plans to scale the company’s Maia accelerator family to multi-gigawatt deployments beginning in 2027. The strategy puts Microsoft’s own chips at the center of its AI inference infrastructure, which powers applications such as Microsoft 365 Copilot and Teams.
The announcement introduces Maia 300, a new AI accelerator that Microsoft says can deliver 40% more tokens per watt than leading GPU-based systems. Tokens are the units of text processed or generated by AI models, making tokens per watt an important measure of how much AI output hardware can produce for a given amount of electricity.
The move reflects growing pressure on major technology companies to control the cost and energy demands of AI. As usage expands, running models efficiently at scale can become as important as improving their underlying capabilities. Custom silicon gives Microsoft more control over the hardware, software and infrastructure used to deliver AI services.
Maia 300 Targets Greater AI Inference Efficiency
According to Nadella’s announcement, Maia 300 is designed to improve inference efficiency by 40% compared with leading GPU-based systems. Inference occurs when a trained AI model processes a request and generates an answer, code or other output. It is a recurring computing expense for services that handle millions of user interactions.
Microsoft’s efficiency claim could have significant implications for operating costs if it translates into comparable workloads and production environments. Higher output per watt can help data centers serve more requests without increasing electricity consumption at the same rate. However, actual savings will depend on model size, workload characteristics, hardware utilization and the wider data center system.
Microsoft has not positioned Maia as a replacement for every third-party accelerator. Nadella said the company would continue incorporating innovations from across the hardware ecosystem while building on its own silicon. That approach allows Microsoft to develop specialized infrastructure for its own services while retaining flexibility to use chips from other suppliers.
The company also said Maia 200 is already being deployed across the United States, Europe and Australia. According to Nadella, it has delivered up to 1.7 times more tokens per watt across experiences in Teams and Copilot. Those figures are Microsoft’s reported performance claims, and comparisons depend on the systems, models and workloads used for measurement.
The broader infrastructure strategy includes optimizing chips alongside networking, memory, cooling and software. Microsoft has previously described its Maia program as a cloud-focused effort to design and validate the complete system rather than treating the accelerator as an isolated component. That full-stack approach can influence how efficiently a model runs once it reaches production.
Microsoft Plans Multi-Gigawatt Deployments From 2027
The most significant part of the announcement is Microsoft’s plan to expand Maia deployments to multi-gigawatt scale starting next calendar year. Gigawatts measure power capacity, so the commitment signals an infrastructure rollout on a scale requiring substantial data center, electricity and supply-chain resources. Microsoft has not disclosed a final deployment total in the announcement.
The plan also highlights the economics of AI infrastructure. Companies are investing heavily in data centers and accelerators to meet demand for generative AI services, but operating costs can rise alongside usage. More efficient custom chips could help Microsoft manage that growth and reduce dependence on any single external hardware platform.
Microsoft’s own silicon effort puts it in competition with established AI chip suppliers, particularly in inference workloads. Yet the comparison is not simply about raw chip performance. Total cost of ownership, software compatibility, reliability, networking and the ability to deploy at scale will determine whether custom accelerators provide a sustained advantage.
For Microsoft customers, the potential benefit is more AI capacity and possibly lower operating costs over time. Greater efficiency could support more Copilot and Teams interactions, but it does not automatically mean lower prices for customers or better results from every AI model. Those outcomes will depend on how Microsoft translates infrastructure gains into product performance and pricing.
Maia 300 therefore represents a strategic investment in the computing layer behind Microsoft’s AI business. The promised 40% efficiency gain and planned multi-gigawatt rollout are significant targets, but their long-term impact will become clearer as the hardware is deployed and independently comparable production results emerge. For now, Microsoft is signaling that controlling the cost of inference is central to its next phase of AI expansion.
About The Author
