Deepseek AI will increase data storage and make AI more affordable

Deepseek’s effective AI training has caused a great deal of debate in the AI ​​community and has caused volatility in reserves related to AI. However, we must not be surprised by advances such as those in Deepseek development. The various technologies used for calculating, networking, memory and storage that allow today’s AI training to have a long history of innovations that lead to greater efficiency and lower energy consumption.

These advances will continue on both material and software and allow data centers to do more with less. They will also make AI training more accessible to more organizations, allow to do more with current data centers and lead digital storage and memory development to support more AI training.

Driving development projections for data centers estimates that future data centers that do heavy AI duties could require multiple giga-watt, GW, energy consumption. This can be compared to the estimated 5.8GW power consumed by San Francisco, Ca. In other words, individual data centers are expected to require as much power as a big city. This causes data centers to consider creating their own power, using sources of renewable and non -renewable energy, including articulated nuclear reactors.

What if we could make future data centers more effective in training and including AI and thus slow down the expected increase in consumption of data center? The most effective AI training approaches, such as those used by Deepseek, could give AI training more affordable and allow more training with less energy consumption.

Deepseek has achieved effective training with significantly fewer resources compared to other AI models using a “mixture of experts” architecture, where specialized sub-models handle different tasks, effectively distributing computing load and only activating the relevant parts of the model for each entrance, thereby reducing the model. the need for a huge amount of computing power and data. This approach, combined with techniques such as smart memory compression and the training of only the most critical parameters, allowed them to achieve high performance with less material, L0wer training time and energy consumption.

The most effective AI training will allow new models to be made with less investments and thus allow more AI training than more organizations. Even if the data for training is compressed, more models mean more storage and memory will be needed to limit the data needed for training. Demand for AI digital storage will continue to grow, activated by more effective AI training. In my opinion, it is likely to be even greater performance in AI training and that additional developments in AI training and training algorithms, beyond those used by Deepseek, which could help us limit future energy Requirements for AI.

This is important to make more efficient efficient data centers and make more efficient investments to implement AI and will need to provide better AI yields in investment. If we do not develop and implement these current and future developments, the intended increase in the energy consumption of the data center will threaten sustainability efforts and could be an economic obstacle to the development of AI. Let us consider the power consumption forecasts of the data center, including provisions for data storage consumption.

A Recent report by the US Department of Energyproduced by the Lawrence Berkeley National Laboratory examined the historical trends and projections for the energy consumption of the Data Center in the United States from 2014 to 2028, cf. Below. By about 2018, the total percentage of energy produced by the data centers was quite level and less than 2%. Increasing trends for cloud computing and particularly various AI types led to energy consumption to 4.4% by 2023. Views in 2028 are expected to increase to 6.7-12.0%. This development could exert severe pressure on our electricity grid.

During the period until 2018, although the computing and other activities of the Data Center have increased, the greatest efficiency achieved through architectural and software changes, such as virtual machines and containers, as well as increasing the processing of special purpose and new technologies Edge and networking was able to limit the total energy consumption of the data center.

AI and other developing IT applications require more and more digital storage and memory to keep the processing data. The storage and use of memory use and the following figure from the DOE report presents the consumption of energy consumption of digital storage of an estimated center from 2014 and is displayed by 2028.

The diagram, updated by IDC data, shows higher growth than 2018 with projections of approximately 2x increased energy consumption by 2028, with a higher percentage of this increase in Flash Nand -based SSD energy consumption. This is likely to be somewhat due to the increase in the development of SSDs for data center applications, especially for primary storage due to their higher performance, but most of this development is probably due to more intense writing and reading SSDs to support the AI and similar work flows, writing and reading in SSDs uses more energy than when SSDs do not have access.

HDDs, increasingly used for secondary storage, data maintaining data, where data has not been immediately processed, have become increasingly effective than power, even when the overall storage capacity of these devices has increased. Consequently, SSDs could represent almost half of the data consumption of the data center by 2028.

However, the projected development of energy consumption for storage and memory in these views is much smaller than that required to process GPU for AI models. New storage and memory technologies, such as memory and storage and memory concentration, as well as storage distribution, using software management, will probably create more efficient storage and memory use for AI applications and thus also help more effective AI modeling.

Deepseek and similar effective AI training approaches could reduce the data center requirements of the data center, make AI modeling more accessible and increase data storage and memory demand. Even greater efficiency is possible and this could help the data centers more viable.