Microsoft plans to deploy its custom Azure Maia AI chips at a multi-gigawatt scale beginning in 2027, according to executive Scott Guthrie. The proprietary silicon is designed to serve as the company’s primary inference platform across its product portfolio, supporting models from OpenAI, MAI, and the open-source community.
Microsoft’s current Maia 200 silicon is already serving Copilot workloads across several global regions, with GitHub Copilot and Azure AI Foundry integrations scheduled by the end of the year. The company claims Maia 200 achieves up to 1.7 times higher token efficiency per watt compared to leading GPUs in Microsoft Teams applications. CEO Satya Nadella characterized the initiative as Microsoft’s largest internal silicon investment to date.
Looking ahead, Microsoft announced the Maia 300 successor chip for 2027, targeting long-context processing, fast conversational latency, and agentic workflows. Microsoft projects a 40 percent improvement in tokens per megawatt over leading GPU systems, with a 96-chip rack expected to deliver up to seven times the token output of a Maia 200 rack.
Why it matters
Hyperscale shift to proprietary silicon will reduce long-term inference costs for enterprise AI applications.
Targeted chip optimizations for long-context and agentic tasks highlight the hardware requirements for next-gen workflows.
Increased multi-gigawatt internal deployments put direct margin pressure on third-party GPU hardware vendors.
Source: the-decoder.com



