In-depth analysis of Microsoft's AI grand strategy: not investing in computing power competition, not betting on LLM, aiming to be the global operating system for AI agents.

Microsoft announced its Q3 2025 financial report, with revenue and profit exceeding market expectations. This quarter, revenue reached $77.67 billion, a year-over-year increase of 18%, with earnings per share at $3.72. The key driver was the cloud division, with Azure's revenue growth rate at 40%. However, in order to strengthen its AI and cloud capabilities, CapEx soared to $34.9 billion, setting a historical high. Additionally, due to the impact of investments related to OpenAI, non-operating gains and losses decreased by $3.7 billion. Despite strong fundamentals, the acceleration of capital expenditures remains the main concern for investors.

(OpenAI has completed its capital restructuring and established PBC! The latest valuation is 500 billion USD, with Microsoft holding a 27% stake)

In 2025, when the entire AI industry is crazily expanding computing power, Microsoft is going against the tide. The company quietly halted the construction of some data centers, which once sparked external doubts: Is Microsoft slowing down amid the global AI infrastructure frenzy? However, Microsoft CEO Satya Nadella revealed a completely different strategic mindset in a recent in-depth interview and conference call: Microsoft is not slowing down, but is clearer than anyone else: the competition for the next generation of AI will not rely on a single model, nor will it be about betting all chips on a generation of GPUs.

Microsoft did not tie down OpenAI, but instead developed a horizontal and vertical ecosystem.

The outside world generally believes that Microsoft has invested billions of dollars in OpenAI, so it would naturally tie its technology direction closely with the GPT series. However, Nadella's statement is quite different. He candidly pointed out that large language model companies actually face a structural risk:

“If you are a modeling company, you are likely to encounter the 'winner's curse': the innovations you work hard to create will immediately become commoditized once they are copied.” His meaning is clear: no one knows which model structure will come out on top. Worse yet: open-source models and enterprise fine-tuning can catch up to cutting-edge models in a short period of time. In other words, the model capabilities you invest $50 billion in training could be matched instantly by some open-source model fine-tuned with private data.

Scaffolding ( is the AI moat, integrating Microsoft infrastructure, models, and agents.

Therefore, Microsoft will not tie its future solely to GPT, but will also utilize cutting-edge models from OpenAI, supporting open source and other vendors like Meta, Anthropic. Nadella believes that the models themselves will gradually become commoditized. The true moat is not in the models, but in the scaffolding. Therefore, while developing its own MAI models, there are also products like Copilot and Azure, cultivating its own ecosystem. Having data and context engineering is the true moat for Microsoft.

Microsoft is not unable to build, but is unwilling to build a giant data center for a generation of GPUs.

In 2025, many companies are frantically building GB200 data centers. However, Microsoft's strategy is completely different: they are halting the construction of some data centers and instead renting computing power from external neocloud and mining companies. As for the reason behind this, Nadella said: I don't want to build a gigawatt-level data center that can only be used for a specific generation of GPU or a specific model architecture.

He explained that the design and requirements of GB200 are different from those of GB300, and by the time of Vera Rubin Ultra, the power consumption and cooling needs will be completely different. Microsoft's strategy is to develop infrastructure that can grow over time, rather than allowing funds to be trapped in infrastructure that seems impressive at first glance but becomes a sunk cost in a few months.

More than half of the cost of building AI data centers is spent on procuring GPUs.

)Barclays downgrades Oracle ORCL rating, close to junk bonds! CapEx surges, cash flow may be cut off next year(

According to reports, the construction cost of an AI data center can reach up to 50 to 60 billion USD per GW, which is three times that of traditional data centers, with over half of the cost coming from the purchase of GPU computing hardware such as NVIDIA. Since the beginning of 2025, global technology companies' estimated CapEx for the coming years has nearly doubled. Among them, Oracle ) has a debt-to-equity ratio of 500%, and Barclays Bank estimates that under the condition of stable CapEx, it will run out of cash as early as November next year. In contrast, Microsoft's debt-to-equity ratio is only 30%, indicating a relatively healthy financial situation.

Industry insiders reveal that the real lifespan of AI data center CPUs is only 1 to 3 years.

Industry insiders with a Google background revealed that the lifespan of CPUs used in AI data centers is only 1 to 3 years.

Michael Burry, the main character of the big short, also stated that many AI companies claim that the AI usage period is not actually that long; they rely on extending the usage period to beautify the depreciation amounts in their financial reports each year. Burry estimates that between 2026 and 2028, the total underestimated depreciation amount of super-large cloud service providers could reach as high as $176 billion. Based on this calculation, he predicts: “By 2028, the earnings of Oracle ( will be overestimated by 26.9%, while Meta will be overestimated by 20.8%.”

) Big Short protagonist Michael Burry criticizes AI giants again: underestimating depreciation and inflating earnings is modern fraud (

Microsoft is unwilling to be bound by capital expenditures and is purchasing computing power from mining companies.

Nadella emphasized fungibility ) interchangeability (, and the condition for Microsoft to make a significant investment is the ability to adapt to various large language models, complete multi-stage training, data generation, inference, and also support multiple generations of GPUs, which gives meaning to the investment. This is why Microsoft prefers to rent external computing power instead of being tied down by a single chip generation. It also explains why many cloud computing providers like IREN have recently become partners of Microsoft rather than competitors.

) IREN signed a $9.7 billion AI cloud deal with Microsoft, and the stock price rose over 7% (

Microsoft's business model has shifted from to C to to Agent.

In the past, Microsoft's business model was to sell software services to consumers, but now their goal is to sell infrastructure to AI agents )Business to Agent(. Microsoft does not aim to win the model war but rather to be Microsoft in the era of AI agents. Models will become more numerous, updated, and stronger. Hardware will become denser and consume more power with each generation. Data centers will continuously be redesigned to meet new power demands. But one thing will not change: AI agents require a world-class, reliable, auditable, and cross-generational compatible infrastructure to operate.

What Microsoft needs to do is this. This is also the real message that Satya Nadella wants to convey in this interview: the models will change, the chips will change, but the “operating environment of AI agents” is the battlefield that remains constant.

This article deeply analyzes Microsoft's AI grand strategy: not investing in computing power competitions, not betting on LLMs, aiming to be the global operating system for AI agents, first appearing in Chain News ABMedia.