Microsoft announced the Maia 300, the third generation of its custom AI chips for Azure, focused on inference. The announcement appeared on the company's official blog on October 12, 2026, signed by Scott Guthrie, Executive Vice President of Cloud + AI.
According to Microsoft, the Maia 300 is expected to deliver 40% more tokens per megawatt compared to leading GPU-based systems. In a full 96-accelerator configuration, the rack offers up to 7 times the token throughput of Maia 200. Large-scale production begins in 2027, with multi-gigawatt deployment plans that will make Maia the primary inference system for Microsoft products and services.
What changes in the hardware
The chip uses a multi-chiplet architecture: compute chiplets are built on TSMC's N2P process, while other functions use cheaper nodes. Each accelerator carries eight HBM4 stacks, totaling 288 GB of memory and 18 TB/s of bandwidth. TDP is 950 watts. A rack of 96 accelerators operates at about 167 kW and delivers 27.6 TB of aggregate HBM capacity.
The company highlights support for low-precision formats such as enhanced MXFP4, software-defined dataflow architecture, and optimizations for long-context workloads, low-latency conversations, and multi-step agentic tasks. Maia 200 is already in production and serving workloads such as Microsoft Copilot.
Why it matters
The move reduces Microsoft's reliance on NVIDIA and AMD GPUs for large-scale inference. With the explosive growth of AI agents and reasoning models, energy efficiency has become the main bottleneck. Maia 300 is positioned as a direct response to that pressure, focusing on useful tokens per watt rather than raw compute peaks.
The company plans to share more details at Microsoft Ignite 2026. Technical details are available in the architecture paper and on the dedicated site.
Sources
Transparency: This content was created, edited, or reviewed with the assistance of artificial intelligence. Information was cross-checked with public posts on X and sources available on the internet. Consult the original sources for full context.
By GeekikiBot
