Inference efficiency increased by 40%: Microsoft (MSFT.US) million chip deployment and remodeling costs

According to Woofun AI, Microsoft (MSFT.US) officially established the Maia accelerator family as the core inference infrastructure, and JPMorgan Chase (JPM.US) immediately published a research report to maintain Microsoft's increase rating and set the target price at $625. The core logic of this investment decision is that Microsoft is fundamentally reconstructing the cost-benefit model of its AI services through large-scale deployment of self-developed chips rather than simply relying on external computing power procurement. The focus of the market has moved from simple revenue growth to structural improvements in capital expenditure efficiency and gross margin of cloud business.

A substantial breakthrough in technology iteration occurred on October 12, when Microsoft announced that the Maia accelerator family will expand to multi-gigawatt deployment from 2027, officially becoming the main inference system for its products and services. The third-generation Maia 300 is scheduled to be put into production in 2027. It has 40% more tokens per megawatt than leading GPU systems, and the token throughput is 7 times that of the previous generation Maia 200 in a rack-scale configuration fully equipped with 96 accelerators.

Notably, the Maia 200 already has 1.7 times more tokens per watt than traditional GPU infrastructure in terms of built-in AI capabilities. This jump in performance is not an accident, but is due to deep innovation in software-defined data flow and computing architectures. At the hardware level, the small chip architecture significantly increased the high-bandwidth memory (HBM) bandwidth per watt, and achieved exponential growth in computing power density in line with the vertical expansion network of data centers.

Currently, the Maia 200 fully supports Copilot's production workloads, and GitHub Copilot and Microsoft Foundry are expected to follow suit by the end of 2026. Microsoft has accumulated rich experience supporting the OpenAI model, the MAI model, and leading open source model families. Komo believes that Maia's transformation from a supporting role to a main inference system marks a key shift in Microsoft's AI infrastructure strategy.

According to data compiled by Woofun AI, this architectural innovation directly contributed to the reduction in unit inference costs and laid the technical foundation for subsequent large-scale deployment.

The effect of scale is another core variable driving cost optimization. Reuters reports that Microsoft plans to deploy more than 1 million Maia 300 chips in 2027 to support internal AI services. Meanwhile, Microsoft's infrastructure expansion plans are just as aggressive. It is expected to more than double the global data center capacity by 2032, and Komo predicts that capacity will increase to more than 30 gigawatts by 2030.

This million-chip deployment and simultaneous expansion of data center capacity are aimed at optimizing the overall cost structure through economies of scale. In terms of financial performance, Komo expects Microsoft's capital expenditure to be around US$210 billion for the 2027 fiscal year and approximately US$175 billion for the 2026 calendar year.

Although the absolute value of capital expenditure is huge, the Maia deployment mainly focuses on internal use cases and does not pursue external computing revenue, which means investors need to focus on marginal changes in capital expenditure intensity and cloud software sales costs. The deployment of one million chips corresponds to multiple gigawatts of capacity, and the Maia 300 increased the number of tokens per megawatt by 40%, which means that under the same power constraints, inference throughput has increased significantly, thereby reducing unit computing costs at the macro level.

Judging from the long-term financial impact, the full implementation of Maia as the main reasoning system will directly improve the profit quality of Microsoft Cloud. Currently, Microsoft Cloud's gross margin is in the middle to high-end range of 60% to 70%. Xiaomo believes that with the deepening of Maia's deployment, this indicator is expected to rise further. Since deployment focuses on internal use cases rather than external computing revenue, the return on capital expenses will be reflected more in reduced cloud software sales costs and increased operational efficiency. The reduction in inference costs and the improvement in cloud gross margin have formed a two-wheel drive for the long-term competitiveness of Microsoft's AI infrastructure. Xiaomo maintains an increase in holdings rating based on confidence in this cost structure optimization path. In the future, with the full commissioning of the Maia 300 in 2027 and the continuous expansion of data center capacity, Microsoft's cost advantage in the AI era will become more prominent. This is not only the result of technological success, but also a reflection of strategic strength and execution capabilities.