The original announcement from DeepSeek had observers worrying that perhaps the AI gold rush - and the potential demands on the grid were overblown. Then news came out that DeepSeek may have "distilled" information (stolen information) from OpenAI in developing its large language model (LLM). But it still points to the possibility that we can build LLMs more quickly and efficiently, without necessarily relying on a single brute force approach of throwing more money, chips, and power at the issue. However, IF DeepSeek's claims are half-true, they may increase demand for AI solutions because they cost less. The result would be more "inference," the downstream end-uses associated with everything from autonomous driving to advanced queries. So the jury is out as far as how much energy use will ultimately be triggered by AI.
Meanwhile, we should start looking for issues related to interconnection inflation - utilities employing capacity numbers that overstate what will eventually get build. In the supply interconnection queue, LBLN's 2024 analysis showed that historically 19% of what was in the queue went to commerciality (among other things, a lot of place-holding going on). One has to imagine the data companies are doing the same thing, especially if utility interconnection requirements are not rigorous (involving control of land, for example).
It's still the Wild West, so hype rules today. Tomorrow may tell a different story...
Peter Kelly-Detwiler
Top Contributor