Power up. From supercharging employee productivity and streamlining processes to inspiring innovation, Microsoft’s AI is designed to help you build the next big thing. No matter where you're starting, push what's possible and build your way with Azure's industry-leading AI. Check it out.
AI tools produce substantial carbon emissions, which is one of the reasons why DeepSeek’s purportedly more-efficient-than-others AI model made such a splash in the industry. And in a recent study, French sustainability software company Greenly put DeepSeek’s claims to the test.
According to Greenly, training DeepSeek models takes less time and uses fewer Nvidia chips. When training DeepSeek’s V3 model and Meta’s Llama 3.1 on the same scenario, DeepSeek used 2.78 million graphics processing unit (GPU) hours and Meta’s model used 30.8 million. As training is usually the most carbon-intensive step in operating an AI model, DeepSeek’s faster training time ups its efficiency. Additionally, DeepSeek uses 2,000 Nvidia chips, while Meta’s model uses over 16,000 and ChatGPT uses more than 25,000—and DeepSeek’s chips are less “energy intensive” than those used by ChatGPT.
“The company had to develop these innovations out of necessity—due to US sanctions restricting access to Nvidia’s most advanced AI chips,” Greenly’s study said. “This limitation forced DeepSeek to design models that maximize efficiency rather than relying on large-scale computing power.”
That design model includes DeepSeek’s mixture-of-experts design, which makes the tool delegate user tasks to sub-models, “only activating the necessary computing power for a given request.”
DeepSeek’s relationship (or potentially lack thereof) to data centers helps make it more sustainable, too, Greenly’s study said. Because DeepSeek is an open-weight model, or publicly available, Greenly notes that it can be run on physical devices, rather than solely in cloud computing or through data centers. And by reducing the need for data centers, DeepSeek could in turn reduce the energy used by facilities, which are poised to use twice as much energy in five years as they do now.
All that said, “such gains could easily be short-lived,” the study stated, due to Jevons paradox, or the idea that the more efficient something is, the more it’ll be used—thereby creating more emissions.
“The widespread adoption of generative AI tools mean that any efficiency gains could be rapidly offset by exponential growth in usage,” the study said.