Microsoft plans Maia 300 as main Azure inference chip in 2027

Microsoft unveiled Maia 300 for 2027 multi-gigawatt Azure inference, claiming better token efficiency than leading GPUs.

Microsoft announced Maia 300, its third-generation custom AI accelerator, and said Maia will become the primary inference system for Azure and Microsoft products at multi-gigawatt scale starting in 2027, a plan The Decoder attributes to executive Scott Guthrie. Both reports say Microsoft claims 40 percent more tokens per megawatt than leading GPU systems and up to seven times a Maia 200 rack's token throughput in a 96-accelerator system. The Verge describes a TSMC N2P chiplet design that separates compute, memory controllers, and I/O for long-context reasoning and agentic workloads, while The Decoder adds eight HBM4 stacks totaling 288 GB at 18 TB/s and a 950-watt rating, with shipping in 2027. The sources disagree on Maia 200's current production footprint: The Verge says it already runs Copilot, GitHub Copilot, and Microsoft Foundry, whereas The Decoder says Copilot is live in the United States, Europe, and Australia and that GitHub Copilot and Foundry follow by year-end. The Decoder also says Satya Nadella called the silicon Microsoft's biggest bet, that the chips serve OpenAI, MAI, and open-source models, and that Maia 200 delivers up to 1.7 times more tokens per watt than leading GPUs in Teams.

• Microsoft announced Maia 300, its third-generation custom AI accelerator, as the planned primary inference system for Azure and Microsoft products at multi-gigawatt scale starting in 2027.
• Microsoft claims Maia 300 delivers 40 percent more tokens per megawatt than leading GPU systems and up to seven times a Maia 200 rack's token throughput in a 96-accelerator configuration.
• The Verge reports a TSMC N2P chiplet design that separates compute, memory controllers, and I/O, targeting long-context reasoning and agentic workloads.
• The Decoder reports Maia 300 ships in 2027 with eight HBM4 stacks totaling 288 GB at 18 TB/s and a 950-watt rating.
• Sources disagree on Maia 200: The Verge says it already runs Copilot, GitHub Copilot, and Microsoft Foundry; The Decoder says Copilot is live in the US, Europe, and Australia, with GitHub Copilot and Foundry to follow by year-end.
• The Decoder attributes the multi-gigawatt 2027 plan to Scott Guthrie, quotes Satya Nadella calling the silicon Microsoft's biggest bet, and says the chips serve OpenAI, MAI, and open-source models.
• The Decoder says Maia 200 pushes up to 1.7 times more tokens per watt than leading GPUs in Teams.