The company said on Monday that the new accelerator would become the principal system for AI inference across its products, including Copilot, as demand grows for computing resources capable of processing complex instructions and generating responses.
Microsoft expects Maia 300 to produce 40 per cent more AI tokens per megawatt than leading systems based on graphics processing units. That figure is a company projection rather than an independently established performance comparison, and commercial deployment is not scheduled to begin until next year.
A fully configured rack containing 96 Maia 300 accelerators is designed to deliver up to seven times the token throughput of its Maia 200 predecessor, Microsoft said. Each rack will operate within a 167-kilowatt power envelope, highlighting the importance of electricity consumption in the company’s expansion plans.
The announcement marks a substantial effort to shift more of Microsoft’s internal AI computing onto chips designed specifically for its own workloads, while continuing to use processors supplied by Nvidia and other manufacturers. Custom hardware could give Microsoft greater flexibility in matching computing capacity to its software and managing infrastructure expenditure.
Microsoft has not disclosed a confirmed total order or the number of Maia 300 chips it will deploy in 2027. Multi-gigawatt capacity would nevertheless require an extensive manufacturing and installation programme, with production volumes and deployment schedules determining how quickly its dependence on externally supplied accelerators can change.
The Maia 300 employs a multi-chiplet design that separates computing functions from memory control and input-output operations. Its main computing components use Taiwan Semiconductor Manufacturing Company’s N2P manufacturing process, while other components are designed around different requirements for connectivity and cost.
The processor incorporates eight stacks of high-bandwidth HBM4 memory, providing 288 gigabytes of capacity and a peak bandwidth of 18 terabytes per second. Microsoft specifies a thermal design power of 950 watts for each accelerator. High memory bandwidth is particularly important when AI systems must process lengthy documents, maintain conversational context or undertake successive tasks.
According to Microsoft’s technical description, the design delivers two and a half times the memory bandwidth of Maia 200 and one-third more memory capacity. The company has also modified how data moves between processing components, aiming to limit energy spent transferring information rather than carrying out calculations.
Microsoft’s software-defined dataflow architecture coordinates processing, memory and networking. The company says its approach allows related computing operations to keep data close to where it is needed, reducing delays when large language models generate answers or perform multi-stage operations.
The system is intended to support AI agents that complete sequences of actions, coding assistants and services requiring rapid conversational responses. Microsoft identified GitHub Copilot, Microsoft Foundry and Dragon Copilot among applications expected to benefit from the expanded Maia infrastructure.
Scott Guthrie, Microsoft’s executive vice-president for cloud and AI, said Maia 200 was already handling production workloads for Microsoft Copilot and assisting the company’s model development. Deployments of that earlier processor are under way in the United States, Europe and Australia, with additional services scheduled to use it before the end of 2026.
The company says Maia 200 delivers up to 1.7 times as many tokens per watt as leading GPU infrastructure for selected built-in AI capabilities. Those comparisons depend on the workloads and systems being measured, making it difficult to extrapolate a single advantage across all applications.
At larger scale, Microsoft’s design connects as many as 128 racks into a networked group containing 12,288 accelerators. It uses Ethernet-based communication and optical connections between racks to distribute workloads, while liquid cooling removes most of the heat generated by the processing and switching hardware.
Microsoft has not announced Maia 300 pricing or offered a timetable for availability to outside cloud customers. Its stated priority is running its own services, where the company can tune applications and hardware together. That distinction matters because a proprietary processor does not automatically replace the software support available for GPU platforms.
Each Maia 300 rack will comprise 24 trays holding four accelerators apiece. Microsoft says its integrated arrangement also accommodates host processors, storage and network switching equipment, with monitoring systems allocating power according to operating conditions.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
