Microsoft positions Maia 300 as its primary AI inference system

Microsoft describes Maia 300 as an inference-first accelerator built for long-context, memory-intensive and decode-heavy workloads.[2] The company says it is designed for long-context reasoning, latency-sensitive conversations and sustained, multi-step agentic tasks.[2] Microsoft says Maia 300 will become the primary AI inference system for its products and services, while it plans to rely mainly on its own silicon to handle incoming requests.[1][3]

Microsoft expects Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems.[2] It also says a full rack with 96 accelerators can provide up to seven times Maia 200’s rack-scale token throughput.[2] The rack combines those accelerators within a 167-kilowatt power envelope and provides 27.6 terabytes of HBM capacity.[2] Microsoft says the rack’s scale-up domain has 1.7 petabytes per second of aggregate HBM bandwidth.[2] These are company-reported specifications and expectations; the evidence does not provide an independent comparison or pricing.

Microsoft’s announcement was dated October 12, 2026, and the company says Maia 300 will begin production deployments in 2027.[2] It also plans to scale its Azure Maia family to multi-gigawatt deployments starting that year.[2] Microsoft says Maia 300 uses eight HBM4 stacks, providing up to 18 terabytes per second of memory bandwidth within a 950-watt thermal design power.[2] Its chiplet design separates compute, memory-controller and I/O functions, which Microsoft says is intended to improve efficiency, scalability and cost structure.[2]

GitHub Copilot and Microsoft Foundry were expected to follow by the end of the year.[2] CNBC reported that Microsoft’s accelerators cannot be rented by the hour to other companies, and that Maia chips had not been tested outside Microsoft.[3] The evidence therefore describes an expansion of Microsoft’s own silicon use, not an end to its GPU use.