Microsoft’s results demonstrate cloud AI balancing act.

was a subtext of DeepSeek, and the drive for greater artificial intelligence (AI). The cloud business of the company reported revenue of $69.6bn in the quarter ending 31 December 2024.

Microsoft’s CEO Satya Nadella announced that Microsoft Cloud had exceeded $40bn in revenues for the first time. This is a 21% increase year-over-year. “Enterprises have begun to move beyond proof-of concept to enterprise-wide deployments in order to unlock the full ROI of AI,” said Nadella. “Our AI business has now exceeded an annual revenue run-rate of $13bn – up 175% from last year.”

Nadella acknowledged during the earnings call that the challenge for hyperscalers was to ensure that they have a fleet that is equipped with the most advanced AI-optimised equipment, while also keeping in mind the emergence and development of new AI models which do not require the latest hardware.

He said, “You don’t need to buy too much at once.” Moore’s Law will give you two times [performance improvement]every year. Your [software] optimization will give you 10x. You want to upgrade the fleet continuously, modernise it, age it, and have the right balance of demand-driven monetisation at the end of day.”

Microsoft is currently forced to purchase expensive AI acceleration hardware which is affecting its margins. Amy Hood, the chief financial officer at Microsoft, said that Microsoft Cloud margins were down year-overyear due to “the impact of scale our AI infrastructure”. She did note that this scaling does not necessarily allow it to grow its installed bases significantly. Microsoft, for example, expects a “moderate increase” in the renewal rate of its M365 Copilot and E5 subscriptions due to the size the existing installed base.

Nadella said that DeepSeek “has some real innovation” when asked about its impact on US AI dominance. DeepSeek is seen as a competitor to US AI dominance. One of them is the cost, both in terms of the hardware needed to run inference workloads as well as the cost to use the system, which is measured by tokens.

In response to a question about whether Microsoft is taking a step back in OpenAI funding because of the recent Stargate Project he said: “One thing to note in AI, you don’t launch the frontier model. If it is too expensive to provide, it is not worth it. It won’t generate demand. You need to optimise so that inferencing cost can come down, and they [the model]

can be consumed widely.”

Essentially, this translates to a lower token price which makes the AI model easier to use. The hardware required to run DeepSeek and other large language models is significantly reduced.

Nadella said: “When token prices drop, inference computing costs fall; this means people can consume even more.” “And more apps will be written.” It’s hard to imagine that we’re at the dawn of ’25 and you can run on a PC a model [previously] that required pretty massive cloud-based infrastructure.”

[HesaidDeepSeekdemonstratedAIwillbemoreubiquitousandrepresentedanopportunityforMicrosoft’sbusiness[HesaidDeepSeekdemonstratedAIwillbemoreubiquitousandrepresentedanopportunityforMicrosoft’sbusiness”ThisisgoodnewsforaPCplatformprovidersuchasus”saidNadella

Nadella’s prepared remarks also touched upon the emergence more efficient AI models. Although the prepared statement did not directly refer to DeepSeek, Nadella added that “as AI becomes more accessible and efficient, we will see an exponentially greater demand.”