Microsoft has unveiled its Azure Maia 300 AI chip as part of its broader effort to build more of its own computing infrastructure and reduce its dependence on external semiconductor suppliers such as Nvidia and AMD.
Announced at the OCP Global Summit 2026 in San Jose, California, Maia 300 is the latest addition to Microsoft’s in-house AI accelerator programme. The company plans to expand deployments of its Maia chip family to multi-gigawatt scale beginning in 2027, with the chips expected to become a primary system for running AI models across Microsoft products and services.
The move comes as technology companies invest heavily in specialised hardware to meet growing demand for AI assistants, cloud computing and enterprise AI applications.
Microsoft Maia 300 Targets Better AI Performance
Microsoft says Maia 300 is designed to deliver more AI processing capacity while improving energy efficiency.
According to the company’s figures, Maia 300 can generate 40% more tokens per megawatt than leading GPU-based systems. Microsoft also claims that a rack containing 96 Maia 300 accelerators can deliver up to seven times the rack-scale token throughput of its previous-generation Maia 200 system.
Tokens are the basic units of information processed by AI models. Depending on the model and data format, they can represent parts of words or other pieces of information. Token throughput measures how much model output a system can generate over a given period.
The number of tokens produced for each unit of electricity is an important measure for AI infrastructure providers. Better efficiency can help cloud companies serve more requests without increasing power consumption at the same rate.
However, actual performance depends on the model, workload, hardware configuration and operating conditions. Microsoft’s figures are company-reported comparisons rather than independently verified results.
Why Microsoft Wants Its Own AI Chips
Microsoft operates large cloud and AI services that require substantial computing resources. Using specialised processors designed for its workloads gives the company another way to optimise performance, power consumption and infrastructure costs.
Nvidia and AMD remain important suppliers in the AI computing market. Developing its own accelerators allows Microsoft to diversify its hardware options and gain greater control over how its systems are designed for specific applications.
Custom AI chips can also be developed alongside software and AI models. This allows hardware engineers to optimise memory, data movement and computing resources for the workloads the company expects to run.
Microsoft’s strategy does not necessarily mean it will stop using third-party processors. Instead, its Maia programme provides an additional hardware platform that can complement chips from external suppliers.
Maia 200 Already Powers Microsoft’s AI Services
Before Maia 300, Microsoft introduced Maia 200, which is already being used for production AI workloads.
According to Microsoft, Maia 200 supports Microsoft 365 Copilot, OpenAI models, the company’s own MAI models and selected open-source AI models.
The company also reports that Maia 200 delivers up to 1.7 times more tokens per watt for built-in AI capabilities in Microsoft Teams, Copilot and Cowork experiences compared with leading GPU-based infrastructure.
These systems demonstrate Microsoft’s approach to designing hardware around the AI services it operates at scale.
The experience gained from running real-world workloads on Maia 200 is also helping the company develop Maia 300. Microsoft aims to improve how its hardware handles large models, long-context requests and AI applications that require multiple steps to complete a task.
How Maia 300 Could Change AI Infrastructure
Modern AI applications increasingly require more than generating a short answer to a question. They may need to process lengthy documents, analyse large amounts of information, generate code or perform several connected actions.
These workloads place demands on processors, memory bandwidth, networking and electricity supplies.
Microsoft has designed Maia 300 to address these requirements through improvements to its computing architecture and the way data moves between system components.
If the chip delivers the performance and efficiency Microsoft expects, it could help the company run more AI requests using its available data-centre capacity.
For cloud providers, improving processing efficiency is important because operating AI services involves substantial ongoing expenses, including electricity, cooling, networking and hardware maintenance.
Higher throughput per rack could also help Microsoft expand AI capacity while managing the physical space and power required by its data centres.
What This Means for Nvidia and AMD
Microsoft’s investment in in-house AI chips reflects a broader effort among major technology companies to develop specialised processors for their own cloud platforms and AI services.
Nvidia has established a strong position in AI computing through its GPUs and software ecosystem, while AMD also supplies processors for data centres and AI workloads.
The expansion of Microsoft’s Maia programme introduces another option for running AI models within its cloud infrastructure. As Microsoft deploys more of its own accelerators, it may have greater flexibility in deciding which workloads run on custom chips and which use external hardware.
However, competing with established AI chip suppliers involves more than achieving strong performance figures. Hardware availability, software compatibility, developer support, reliability and the ability to manufacture chips at scale also matter.
Microsoft’s Maia chips are therefore best understood as part of a broader hardware diversification strategy, rather than evidence that the company will no longer need Nvidia or AMD.
What Comes Next?
Microsoft plans to begin multi-gigawatt deployments of its Maia chip family in 2027. Maia 300 is expected to play a central role as the company expands its AI infrastructure to support growing demand across its cloud and software businesses.
The next important test will be how the new accelerators perform in large-scale production environments and whether their efficiency advantages translate into lower operating costs.
Microsoft will also need to continue developing the software and infrastructure required to support a wide range of AI models on its custom hardware.
Conclusion
Microsoft’s Maia 300 represents another step in its effort to build a more specialised and flexible AI infrastructure. With reported improvements in token throughput and energy efficiency, the chip could help support the company’s growing AI services while reducing its reliance on any single hardware supplier. Its long-term impact will depend on real-world performance, deployment scale and how effectively Microsoft integrates Maia into its cloud ecosystem.