Microsoft is preparing to bring its self-developed Azure Maia AI chip from a supporting role in the lab to the main stage. Executive Scott Guthrie revealed in an article that the company plans to deploy the Maia chip at a scale of multiple gigawatts starting in 2027, making it the primary inference system across Microsoft's product lines.

The current Maia200 has already been running Microsoft Copilot, GitHub Copilot, and Foundry in the US, Europe, and Australia. The CEO Satya Nadella referred to this strategy as "our biggest bet on our own silicon chips." This chip is versatile—it can simultaneously support inference for OpenAI, Microsoft's own MAI, and open-source models. Microsoft also released data indicating that in Teams scenarios, its token throughput per watt is up to 1.7 times higher than leading GPUs, meaning more intelligence with less power.

The next-generation Maia300, which will be shipped in 2027, targets tasks that require high computing power, such as long input, fast chat, and intelligent agent operations. Microsoft claims that its token output per megawatt is 40% higher than leading GPU systems. A rack containing 96 chips can produce up to seven times more tokens than a Maia200 rack. From a hardware specification perspective, each Maia300 integrates eight HBM4 memory stacks, totaling 288GB with a bandwidth of 18TB/s, and has a rated power consumption of 950 watts—equivalent to the power consumption of an entire rack in the past, now concentrated in a single chip. More detailed technical reports have already been released, and additional details will be shared at the Ignite 2026 conference in November.

When a cloud giant deploys its own chip at a gigawatt scale and explicitly states it will serve as the main inference system, it clearly aims not only to reduce dependence on external GPUs but also to create a self-sufficient chain linking models, chips, and cloud services, resulting in a bill that others cannot calculate or block.