Deepseek’s efficient AI training has caused much discussion in the AI community and caused volatility in AI related stocks. However, we should not be surprised at advances like those made in developing Deepseek. The various technologies used for computing, networking, memory and storage that enable today’s AI training have a long history of innovations leading to greater efficiency and lower power consumption.
These advances will continue in both hardware and software and enable data centers to do more with less. They will also make AI training more accessible to more organizations, enable doing more with current data centers and driving digital storage and memory growth to support more AI training.
Driving the growth projections for data centers are estimates that future data centers doing heavy AI tasks could require multiple giga-watt, GW, power consumption. This can be compared to the estimated 5.8GW of power consumed by San Francisco, CA. In other words, single data centers are projected to require as much power as a large city. This is causing data centers to look at generating their own power, using renewable and non-renewable power sources, including modular nuclear reactors.
What if we could make future data centers more efficient in AI training and inference and thus slow the anticipated data center power consumption growth? More efficient AI training approaches like those used by Deepseek could give make AI training more accessible and allow more training with less energy consumption.
DeepSeek achieved efficient training with significantly less resources compared to other AI models by utilizing a “Mixture of Experts” architecture, where specialized sub-models handle different tasks, effectively distributing computational load and only activating relevant parts of the model for each input, thus reducing the need for massive amounts of computing power and data. This approach, combined with techniques like smart memory compression and training only the most crucial parameters, allowed them to achieve high performance with less hardware, l0wer training time and power consumption.
More efficient AI training will enable new models to be made with less investment and thus enable more AI training by more organizations. Even if data for training is compressed, more models mean more storage and memory will be needed to contain the data needed for training. Digital storage demand for AI will continue to grow, enabled by more efficient AI training. In my opinion, there are likely even more efficiencies possible in AI training and that additional developments in AI training methodologies and algorithms, beyond those used by Deepseek, that could help us constrain future energy requirements for AI.
This is important to enable more efficient data centers and to make more effective investments to implement AI and will be needed to provide better AI returns on investments. If we don’t develop and implement these current and future advances, the projected growth in data center power consumption will threaten sustainability efforts and could be an economic barrier to AI development. Let’s look at data center power consumption projections, including projections for data storage power consumption.
A recent report from the US Department of Energy, produced by the Lawrence Berkeley National Laboratory examined historical trends and projections for data center power consumption in the United States from 2014 through 2028, see below. Up until about 2018 the total percentage of generated energy consumed by data centers had been fairly flat and less than 2%. Growing trends for cloud computing and in particular various types of AI drove power consumption to 4.4% by 2023. Projections going forward to 2028 were projected to grow to 6.7-12.0%. This growth could put serious pressure on our electrical grid.
During the period leading up to 2018, although computing and other data center activities increased, greater efficiencies achieved through architectural and software changes such as virtual machines and containers as well as the rise of special purpose processing and new scaling and networking technologies were able to constrain the total data center energy consumption.
AI and other growing computing applications require more and more digital storage and memory to hold the data being processing. Storage and memory use power and the figure below from the DOE report, shows estimated data center digital storage energy consumption from 2014 and projected to 2028.
The chart, informed by data from IDC, shows higher growth since 2018 with projections of about a 2X increased power consumption out to 2028, with a greater percentage of this growth in power consumption from NAND flash-based SSDs. This is likely due somewhat to increasing growth in SSDs for data center applications, particularly for primary storage because of their higher performance, but most of this growth is probably due to more intense writing and reading of SSDs to support AI and similar workflows, writing and reading in SSDs uses more energy than when the SSDs are not being accessed.
HDDs, increasingly used for secondary storage, for data retention, where the data isn’t being immediately being processed, have been become increasingly more power efficient even as the total storage capacity of these devices have increased. As a consequence, SSDs could account for almost half of data center storage energy consumption by 2028.
However, the projected growth of power consumption for storage and memory in these projections, is much less than that required for GPU processing for AI models. New storage and memory technologies, such as pooling of memory and storage and memory as well as storage allocation using software management will likely create more efficient storage and memory use for AI applications and thus also help to make more efficient AI modeling.
Deepseek and similar more efficient AI training approaches could reduce data center power requirements, make AI modelling more accessible and increase data storage and memory demand. Even more efficiencies are possible and this could help make data centers more sustainable.
