CEO acknowledges efficiency of DeepSeek and need to assess model optimisation rather than just buying more AI servers
DeepSeek and the drive to greater artificial intelligence (AI) efficiency was the subtext underlying Microsoft’s latest quarterly filing. The company’s cloud business reported revenue of $69.6bn for the quarter that ended 31 December 2024.
Microsoft CEO Satya Nadella said Microsoft Cloud surpassed $40bn in revenue for the first time, up 21% year-over-year. “Enterprises are beginning to move from proof-of-concepts to enterprise-wide deployments to unlock the full ROI [return on investment] of AI,” he said. “And our AI business has now surpassed an annual revenue run rate of $13bn, up 175% year-over-year.”
During the earnings call, Nadella acknowledged that the challenge for hyperscalers is the need to ensure they have a fleet of the most modern AI-optimised hardware balanced with the emergence of new, more efficient AI models that do not need the latest hardware.
“You don’t want to buy too much of anything at one time,” he said. “Because of Moore’s Law, every year is going to give you 2x [performance improvement]. Your [software] optimisation is going to give you 10x. You want to continuously upgrade the fleet, modernise the fleet, age the fleet, and at the end of the day, have the right ratio of modernisation and demand-driven monetisation.”
For now, Microsoft is having to buy expensive AI acceleration hardware, which is impacting its margins. Chief financial officer Amy Hood said that Microsoft Cloud margins are down year-over-year driven by “the impact of scaling our AI infrastructure”. However, she noted that this scaling is not necessarily enabling it to grow its installed base significantly. For instance, Microsoft only expects “moderate growth” in the annual renewal rate of its E5 and M365 Copilot subscriptions, due to the size of the existing installed base.
When asked about the impact of DeepSeek, which is seen as a rival to US AI dominance, Nadella’s view was that it “has some real innovations”. One of these is in terms of cost – both the cost of the hardware required to train and run inference workloads and the cost to use it, which is measured in tokens.
Responding to a question on whether Microsoft is taking a backseat in OpenAI funding due to the recent Stargate Project, he said: “One of the key things to note in AI is you don’t just launch the frontier model. If it’s too expensive to serve, it’s no good. It won’t generate any demand. You’ve got to have that optimisation so that inferencing costs come down and they [the model] can be consumed broadly.”
This equates to a lower price per token, which makes the AI model more affordable to use. With more efficient large language models like DeepSeek, the hardware required to run the model is also significantly lower.
“When token prices fall, inference computing prices fall; that means people can consume more,” said Nadella. “And there will be more apps written. It’s unimaginable to think that here we are in the beginning of ’25, where on the PC, you can run a model that [previously] required pretty massive cloud infrastructure.”
He said DeepSeek demonstrated that AI will be much more ubiquitous, and represented an opportunity for the Microsoft business. “For a hyperscaler like us, a PC platform provider like us, this is all good news as far as I’m concerned,” said Nadella.
His prepared remarks also touched on the emergence of more efficient AI models, and although the prepared statement does not reference DeepSeek directly, Nadella added: “As AI becomes more efficient and accessible, we will see exponentially more demand.”
