In This Article
Microsoft has introduced Maia 300, the next generation of its in-house AI accelerator, as it prepares to scale its Maia systems to multi-gigawatt deployments beginning in 2027. The company says the system is built to serve growing inference demand, with a focus on memory efficiency, data movement and performance across large AI workloads.
Key takeaways
Microsoft’s announcement highlights how AI infrastructure is shifting from individual chips to integrated systems designed around real-world workloads.
- Maia 300 production deployments are scheduled to begin in 2027, alongside Microsoft’s plan to scale Maia across multi-gigawatt deployments.
- Microsoft says Maia 300 can deliver 40% more tokens per megawatt than leading GPU-based systems.
- A 96-accelerator rack is rated for up to seven times Maia 200’s rack-scale token throughput.
- Maia 200 is already serving Microsoft production workloads, including Copilot.
An inference-focused system
Microsoft designed Maia 300 for long-context reasoning, responsive conversational AI and multi-step agentic tasks. Rather than optimizing only for raw compute, the company says the system targets data movement and memory use—important factors in serving large models efficiently.
The accelerator uses eight HBM4 stacks for up to 18 terabytes per second of memory bandwidth within a 950-watt thermal design power. Its chiplet design separates compute, memory-controller and I/O functions; compute chiplets use TSMC’s N2P process. Support for low-precision formats, including enhanced MXFP4, is intended to reduce memory needs and support larger context windows.
Scaling from chip to rack
Microsoft’s full Maia 300 rack configuration combines 96 accelerators in a 167-kilowatt power envelope, with 27.6 terabytes of HBM capacity and 1.7 petabytes per second of aggregate HBM bandwidth. The company says its Ethernet-based AI transport connects accelerators and racks, extending its software-defined dataflow approach across the system.
Microsoft forecasts 40% more tokens per megawatt than leading GPU-based systems and up to seven times the rack-scale token throughput of Maia 200. These are company-reported comparisons; real-world results will depend on workloads and deployment conditions.
Maia 200 is already in production
The announcement also marks Maia 200’s transition into active use. Microsoft says the accelerator is running Copilot and supporting development of its MAI models, with GitHub Copilot and Microsoft Foundry workloads expected to follow by year-end. Maia 200 supports OpenAI, Microsoft and open-source models, and is being deployed in the United States, Europe and Australia.
Microsoft reports up to 1.7 times more tokens per watt than leading GPU infrastructure for built-in AI features in Teams, Copilot and Cowork. Experience with Maia 200, the company says, informed Maia 300’s design.
What enterprise leaders should watch
More efficient AI infrastructure can expand the capacity available for enterprise applications, but compute alone does not ensure useful business outcomes. Organizations also need well-governed, accessible data and reporting systems to put AI insights to work. For Microsoft-focused enterprises, Spargent’s Microsoft Fabric and Power BI consulting addresses that data foundation, helping teams improve analytics delivery and reduce reporting bottlenecks as AI capabilities grow.
- Microsoft to expand Azure Maia system to multi-gigawatt deployments, The Official Microsoft Blog.
- Microsoft unveils Maia 300, designed for multi-gigawatt deployments, Techzine Global.
