Microsoft Unveils Maia 300 to Power AI Agent Inference
Microsoft unveiled its Maia 300 inference platform to handle multi-step AI agents, planning multi-gigawatt Azure deployments in 2027 to boost energy efficiency for complex workloads.
On October 12, Microsoft unveiled Maia 300, its next-generation artificial intelligence inference platform engineered to anchor the company's long-term cloud hardware infrastructure. Microsoft plans to launch multi-gigawatt Azure Maia deployments beginning in 2027, positioning Maia 300 as the primary inference engine across its entire suite of products and services, including Microsoft 365 Copilot, Teams built-in AI, GitHub coding tools, and external model serving through Microsoft Foundry.
The new hardware architecture is optimized for sustained agent workloads, extended reasoning tasks, and long context windows. Microsoft expects Maia 300 to achieve 40% more tokens per megawatt compared to prior efficiency baselines. In addition, when deployed in a full configuration of 96 specialized accelerators per rack, the platform is designed to produce up to seven times the rack-scale token throughput of the earlier Maia 200 hardware.
For developers, systems engineers, and enterprise practitioners, this design shift targets the immense computational demands of agentic workflows. Standard chatbot exchanges require only a single inference step, whereas autonomous AI agents repeatedly query a model to perform multi-step jobs. For instance, a coding agent tasked with repairing a bug might analyze several files, execute tests, interpret failures, and revise code across multiple sequential calls. Each step consumes tokens, increases latency, and requires additional datacenter memory to retain contextual information.
This continuous processing strain highlights power and memory availability as primary bottlenecks for expanding AI deployment. By dramatically expanding rack-scale throughput, Maia 300 aims to clear request backlogs when numerous users deploy Copilot Cowork or GitHub agents simultaneously. Ultimately, improving token generation per megawatt provides the operational efficiency required to make resource-heavy, multi-step AI services economically viable at scale.
This is our own summary of reporting by The Neuron
