One of the significant discussions in the energy industry over the past year has centered on the energy consumption of artificial intelligence and its potential repercussions. The consensus has been that the impact on suppliers may be positive, while the implications for the green transition may not be as favorable.
However, the recent advancements from DeepSeek, a Chinese group that launched a much more efficient large language model, have shifted this narrative. The focus has now turned to how much lower energy demand forecasts may need to be adjusted and what this means for utility companies in the sector.
This issue predominantly concerns the United States. Due to the immediate demand for quick responses, data centers that support AI functions are better off being close to their end users, whereas facilities used to train models can be established where electricity costs are lower. As a result, Europe, with its higher energy prices, has remained less involved in this aspect and has been somewhat insulated from the developments in the US energy market.
It seems likely that power forecasts in the US will need to be revised downward. Training DeepSeek’s model consumed less than ten percent of the computing power required for Meta’s Llama. Although data centers will still require cooling, this efficiency represents a significant change in the landscape.
Once operational, DeepSeek may prove to be more resource-efficient than existing models, such as OpenAI’s ChatGPT, due to its ability to deactivate unused components. While reduced costs might lead to increased utilization, the potential risks exist on the downside.
Translating these efficiencies into reliable long-term energy forecasts poses challenges, especially since US electricity projections have been rather varied and often questionable. The multitude of different forecasts complicates the understanding of underlying assumptions.
Moreover, if cheaper and more efficient AI does increase overall usage, this revelation does not necessarily assure confidence among investors in the US power industry. If DeepSeek truly alters the AI landscape, the resulting unpredictability concerning future energy demand may lead hyperscale companies to hesitate on entering long-term energy contracts.
A slowdown in energy demand could negatively impact energy stock performance, but it might support the energy transition. An abrupt rise in demand would likely have led to more gas-fired power plants being constructed, whereas a slower increase provides a better opportunity for renewables, battery storage, and nuclear energy to contribute. The developments surrounding DeepSeek illustrate that while AI advancements can help the broader world, they may also challenge the investment landscape.
photo credit: www.ft.com
