Deepseek AI
NurPhoto via Getty Images
Deepseek’s innovative AI training methods have sparked discussions in the AI community and caused fluctuations in AI-related stocks. The progress made in developing Deepseek should not come as a surprise, as the technologies used for computing, networking, memory, and storage have a history of leading to more efficient AI training with lower power consumption.
Advancements in hardware and software will continue to enhance data centers, enabling them to achieve more with less resources. These improvements will make AI training more accessible to a wider range of organizations, increase the efficiency of current data centers, and drive growth in digital storage and memory to support increased AI training.
Projections indicate that future data centers focused on heavy AI tasks may require massive power consumption in the range of multiple gigawatts, comparable to the power consumed by large cities. This is prompting data centers to explore options like generating their own power using renewable and non-renewable sources, including modular nuclear reactors.
Efforts to make data centers more efficient in AI training could potentially slow down the expected growth in power consumption. By utilizing technologies like Deepseek’s “Mixture of Experts” architecture, AI training can be achieved with fewer resources, shorter training times, and lower power consumption.
Efficient AI training will allow for the development of new models with reduced investments, enabling more organizations to engage in AI training. As the demand for digital storage grows with compressed data for training, there will be a need for increased storage and memory capacity. The continued growth in digital storage is supported by more efficient AI training methods, with possibilities for further enhancements beyond what Deepseek has achieved.
It is crucial to develop and implement more efficient AI training methods to ensure better returns on AI investments and sustainable data center operations. Failure to adopt these advancements could lead to a substantial increase in data center power consumption, posing challenges to energy sustainability and hindering AI development.
Historical and projected US data center energy consumption growth
DOE and Lawrence Livermore Lab
Data from a recent report by the US Department of Energy shows a significant increase in data center power consumption from 2014 to 2028, driven by the growing trends in cloud computing and AI. This growth could strain the existing electrical grid infrastructure.
Efforts to improve efficiency through architectural and software changes have helped in conserving energy in data centers up to a certain point. However, the increasing demands of AI and other computing applications require more digital storage and memory, leading to higher power consumption as shown in projections from the DOE report.
Estimated data center storage energy consumption history and trends from 2014 to 2028
DOE and Lawrence Livermore Lab
Growth in data storage power consumption is expected to continue, particularly with the rise of NAND flash-based SSDs, driven by the increasing use of SSDs for primary storage in AI workflows. Efforts to develop new storage and memory technologies could lead to more efficient storage solutions for AI applications and contribute to sustainable data center operations.
Efficient AI training approaches like Deepseek’s have the potential to reduce data center power requirements, increase accessibility to AI modeling, and drive demand for data storage and memory. Further advancements in AI training methodologies and algorithms could lead to even greater efficiencies, making data centers more sustainable in the long run.
