Microsoft has introduced Maia 300, the third generation of its in-house artificial-intelligence accelerator, and says production deployments will begin in 2027. The company plans to scale the Azure Maia family to multi-gigawatt installations and make it the primary inference system for Microsoft products and services. The announcement places custom silicon at the centre of Microsoft’s effort to control the cost, energy use and performance of running large AI models.
Maia 300 is designed for inference rather than model training. Inference is the stage in which a trained AI model answers questions, generates text or code, and carries out multi-step tasks. Microsoft says the new system is optimised for long-context reasoning, memory-intensive workloads, low-latency conversations and agent-based processes that repeatedly use tools and make decisions.
Microsoft’s efficiency claims
According to Microsoft, Maia 300 can deliver 40% more tokens per megawatt than leading GPU-based systems. The company did not identify the competing systems used for that comparison, Reuters noted, so the figure should be treated as a vendor claim rather than an independently benchmarked result. Microsoft also says a full rack with 96 accelerators can provide up to seven times the rack-scale token throughput of Maia 200.
Each Maia 300 accelerator includes 288 GB of HBM4 high-bandwidth memory and 18 terabytes per second of peak memory bandwidth within a 950-watt thermal design envelope. A complete rack contains 96 accelerators, has 27.6 TB of HBM capacity and operates within a 167-kilowatt power envelope. Microsoft says liquid cooling removes more than 80% of the rack’s thermal load.
The processor uses a multi-chiplet architecture. Compute chiplets are built on TSMC’s N2P process, while memory-controller and input-output functions are placed on separate dies chosen for bandwidth, connectivity and cost. Microsoft argues that this structure allows more silicon area to be dedicated to computation and reduces the energy spent moving data.
From one rack to a superpod
Maia 300’s network can connect up to 128 racks into a superpod containing 12,288 accelerators. Microsoft’s design uses an Ethernet-based transport and a two-tier network so workloads can communicate across accelerators and racks through a common programming model. The company says the architecture is intended to support large inference jobs without treating networking as a separate layer.
The announcement does not mean Maia 300 is available to Azure customers today. Microsoft says production deployment starts in 2027. Maia 200 is already running some internal workloads, including Microsoft Copilot and Microsoft’s own AI models, with GitHub Copilot and Microsoft Foundry expected to follow. Experience from those services informed the Maia 300 design.
Reuters reported that Microsoft could deploy more than one million Maia 300 chips next year. The company described the broader rollout as multi-gigawatt, but deployment numbers and schedules remain plans that could change with manufacturing capacity, data-centre construction and component supply. Microsoft has said it expects capital expenditure above $175 billion in the calendar year, underlining the scale of its AI infrastructure programme.
Pressure on the GPU market
Custom accelerators give cloud companies another way to manage infrastructure costs and reduce reliance on general-purpose AI GPUs supplied by outside vendors. Microsoft is not presenting Maia as the only hardware in Azure; its cloud platform also uses chips from Nvidia and AMD. The strategic shift is about owning more of the stack, from silicon and networking to models and software.
The Maia 300 launch therefore matters beyond a single processor specification. It shows Microsoft moving toward a vertically integrated inference platform for services used across Copilot, developer tools and enterprise AI. The key test will come in 2027, when planned production deployments can be measured under real workloads against the efficiency and throughput figures announced by the company.
