According to Woofun AI, Microsoft (MSFT.US) has officially established the Maia accelerator family as its core inference infrastructure, and JPMorgan (JPM.US) promptly issued a research report maintaining an overweight rating on Microsoft with a target price of $625.
The core logic behind this investment decision is that Microsoft, through large-scale deployment of its self-developed chips, is fundamentally restructuring the cost-efficiency model of its AI services rather than relying solely on external computing procurement.
The market's focus has shifted from mere revenue growth to structural improvements in capital expenditure efficiency and cloud business gross margins.
The substantive breakthrough in technology iteration occurred on October 12, when Microsoft announced that the Maia accelerator family will expand to multi-gigawatt deployment scale starting in 2027, officially becoming the primary inference system for its products and services.
The third-generation product, Maia 300, is scheduled to enter production in 2027, delivering 40% more tokens per megawatt than leading GPU systems, and achieving 7 times the token throughput of the previous generation Maia 200 in a fully configured rack-level setup with 96 accelerators.
Notably, Maia 200 already delivers 1.7 times more tokens per watt than traditional GPU infrastructure in terms of built-in AI capabilities.
This performance leap is not coincidental but stems from deep innovation in software-defined dataflow and computing architecture.
At the hardware level, the chiplet architecture significantly boosts high-bandwidth memory (HBM) bandwidth per watt, and combined with data-center-scale scale-up networking, achieves exponential growth in computing density.
Currently, Maia 200 fully supports Copilot's production workloads, with GitHub Copilot and Microsoft Foundry expected to follow by the end of 2026.
Microsoft has accumulated extensive experience supporting OpenAI models, MAI models, and leading open-source model families, and JPMorgan believes that Maia's transition from a supporting role to the primary inference system marks a key turning point in Microsoft's AI infrastructure strategy.
Data compiled by Woofun AI shows that this architectural innovation directly drives down unit inference costs, laying the technical foundation for subsequent large-scale deployment.
Scale effects represent another core variable driving cost optimization.
Reuters reported that Microsoft plans to deploy more than 1 million Maia 300 chips by 2027 to support internal AI services.
Meanwhile, Microsoft's infrastructure expansion plans are equally aggressive, with global data center capacity expected to more than triple by 2032, and JPMorgan forecasts that capacity will grow to over 30 gigawatts by 2030.
This million-chip-level deployment, synchronized with data center capacity expansion, aims to optimize the overall cost structure through economies of scale.
On financial performance, JPMorgan estimates Microsoft's capital expenditure for fiscal year 2027 at approximately $210 billion, and approximately $175 billion for calendar year 2026.
Despite the enormous absolute capital expenditure, Maia deployment focuses primarily on internal use cases and does not pursue external computing revenue, meaning investors need to focus on marginal changes in capital expenditure intensity and cloud software cost of sales.
The deployment of a million chips corresponds to multi-gigawatt capacity, and Maia 300's 40% improvement in tokens per megawatt means significantly increased inference throughput under the same power constraints, thereby lowering unit computing costs at the macro level.
From a long-term financial impact perspective, the full implementation of Maia as the primary inference system will directly improve the profitability quality of Microsoft Cloud.
Currently, Microsoft Cloud gross margin sits in the mid-to-high range of 60% to 70%, and JPMorgan believes this metric is poised to rise further as Maia deployment deepens.
Since the deployment focus is on internal use cases rather than external computing revenue, the returns on capital expenditure will be more reflected in reduced cloud software cost of sales and improved operational efficiency.
The decline in inference costs and improvement in cloud gross margins constitute the twin drivers of Microsoft's long-term competitiveness in AI infrastructure.
JPMorgan's decision to maintain its overweight rating is based on confidence in this cost structure optimization path.
In the future, with the full production of Maia 300 in 2027 and continued expansion of data center capacity, Microsoft's cost advantage in the AI era will become even more pronounced, which is not only the result of technological victory but also a reflection of strategic resolve and execution capability.
Comments