Microsoft has introduced a new in-house chip called the Maia 300, designed to power the artificial intelligence services running across the company's infrastructure. Built as the third iteration in Microsoft's line of custom AI processors, the Maia 300 represents the company's continued push to develop its own silicon for artificial intelligence workloads.
Rollouts at multi-gigawatt scale are planned for the following year, with projections pointing to upwards of a million chips handling workloads spanning software development to complex business operations. Proprietary AI models from Microsoft, along with those from OpenAI routed through the Azure cloud platform, were among the workloads handled by the preceding version.
The Maia 300's architecture, according to Microsoft, treats the processor, memory, and surrounding server infrastructure as a unified system rather than separate components — a configuration the company says drives greater computational efficiency. The architecture breaks away from a single large processor, instead distributing work across an array of compact, task-specific chips referred to as chiplets.
Fabricated on TSMC's 2-nanometre process, the processor chiplets handle core computation, while a separate set of components — built to the 3nm node — takes responsibility for shuttling data between different parts of the system. By assigning distinct manufacturing technologies to each functional component, the processor avoids the uniformity of a single production method applied across its entire architecture.
According to figures put forward by Microsoft, the Maia 300 draws roughly 40 percent more usable computation from each megawatt than comparable GPU-driven hardware. No comparison systems were named by the company, which means the figure stands as an unverified assertion on Microsoft's part rather than a finding backed by outside testing.
Each rack destined to house the processors will hold 96 of the new chips once installation is complete, with power draw for each rack projected at around 167 kilowatts. Nvidia's GB300 AI server racks, the company's newest generation, pack 72 GPUs into each unit and draw roughly 142 kilowatts of power, based on available documentation.
Varying rack layouts indicate that energy draw and throughput are factors to be weighed against each other, rather than assessed simply by the number of chips a system holds. Growing appetite for AI-driven computation has pushed technology firms toward building specialised chips engineered around their own processing needs.
Built to handle Microsoft's own services and cloud infrastructure, the Maia 300 is also engineered to extract more computational output from every unit of energy consumed. The scale of the planned expansion has implications for the company's capital commitments, with Microsoft's capital expenditure projected to surpass $175 billion across the current calendar year.
Chipmakers across the industry have been directing resources into proprietary silicon, aiming to reduce their dependence on Nvidia at a time when that company's graphics processing units hold a commanding position across much of the artificial intelligence computing landscape. Deploying the Maia 300 stands to extend Microsoft's foothold across its own AI infrastructure.
