“With cost destruction comes proliferation,” Susan Su, a climate investor at the venture capital firm Toba Capital, told me. “Plus the fact that it’s open source, I think, is a really, really big deal. It puts the power to expand and to deploy and to proliferate into billions of hands.”
If you’ve seen lots of chitchat about Jevons paradox of late, that’s basically what this line of thinking boils down to. After Microsoft’s CEO Satya Nadella responded to DeepSeek mania by posting the Wikipedia page for this 19th century economic theory on X, many (myself included) got a quick crash course on its origins. The idea is that as technical efficiencies of the Victorian era made burning coal cheaper, demand for — and thus consumption of — coal actually increased.
While this is a distinct possibility in the AI space, it’s by no means a guarantee. “This is very much, I think, an open question,“ energy expert Nat Bullard told me, with regards to whether DeepSeek-type models will spur a reduction or increase in energy demand. “I sort of lean in both directions at once.” Formerly the chief content officer at BloombergNEF and current co-founder of the AI startup Halcyon, a search and information platform for energy professionals, Bullard is personally excited for the greater efficiencies and optionality that new AI models can bring to his business.
But he warns that just because DeepSeek was cheap to train — the company claims it cost about $5.5 million, while domestic models cost hundreds of millions or even billions — doesn’t mean that it’s cheap or energy-efficient to operate. “Training more efficiently does not necessarily mean that you can run it that much more efficiently,” Bullard told me. When a large language model answers a question or provides any type of output, it’s said to be making an “inference.” And as Bullard explains, “That may mean, as we move into an era of more and more inference and not just training, then the [energy] impacts could be rather muted.”
DeepSeek-R1, the name for the model that caused the investor freakout, is also a newer type of LLM that uses more energy in general. Up until literally a few days ago, when OpenAI released o3-mini for free, most casual users were probably interacting with so-called “pretrained” AI models. Fed on gobs of internet text, these LLMs spit out answers based primarily on prediction and pattern recognition. DeepSeek released a model like this, called V3, in September. But last year, more advanced “reasoning” models, which can “think,” in some sense, started blowing up. These models — which include o3-mini, the latest version of Anthropic’s Claude, and the now infamous DeepSeek-R1 — have the ability to try out different strategies to arrive at the correct answer, recognize their mistakes, and improve their outputs, allowing for significant advancements in areas such as math and coding.
But all that artificial reasoning eats up a lot of energy. As Sasha Luccioni, the AI and climate lead at Hugging Face, which makes an open-source platform for AI projects, wrote on LinkedIn, “To set things clear about DeepSeek + sustainability: (it seems that) training is much shorter/cheaper/more efficient than traditional LLMs, *but* inference is longer/more expensive/less efficient because of the chain of thought aspect.” Chain of thought refers to the reasoning process these newer models undertake. Luccioni wrote that she’s currently working to evaluate the energy efficiency of both the DeepSeek V3 and R1 models.
Another factor that could influence energy demand is how fast domestic companies respond to the DeepSeek breakthrough with their own new and improved models. Amy Francetic, co-founder at Buoyant Ventures, doesn’t think we’ll have to wait long. “One effect of DeepSeek is that it will highly motivate all of the large LLMs in the U.S. to go faster,” she told me. And because a lot of the big players are fundamentally constrained by energy availability, she’s crossing her fingers that this means they’ll work smarter, not harder. “Hopefully it causes them to find these similar efficiencies rather than just, you know, pouring more gasoline into a less fuel-efficient vehicle.”
In her recent Substack post, Su described three possible futures when it comes to AI’s role in the clean energy transition. The ideal is that AI demand scales slowly enough that nuclear and renewables scale with it. The least hopeful is that immediate, exponential growth in AI demand leads to a similar expansion of fossil fuels, locking in new dirty infrastructure for decades. “I think that's already been happening,” Su told me. And then there’s the techno-optimist scenario, linked to figures like Sam Altman, which Su doesn’t put much stock in — that AI “drives the energy revolution” by helping to create new energy technologies and efficiencies that more than offset the attendant increase in energy demand.
Which scenario predominates could also depend upon whether greater efficiencies, combined with the adoption of AI by smaller, more shallow-pocketed companies, leads to a change in the scale of data centers. “There’s going to be a lot more people using AI. So maybe that means we don’t need these huge, gigawatt data centers. Maybe we need a lot more smaller, megawatt-size data centers,” Laura Katzman, a principal at Buoyant Ventures, told me. Katzman has conducted research for the firm on data center decarbonization.
Smaller data centers with a subsequently smaller energy footprint could pair well with renewable-powered microgrids, which are less practical and economically feasible for hyperscalers. That could be a big win for solar and wind plus battery storage, Katzman explained, but a boondoggle for companies such as Microsoft, which has famously committed to re-opening Pennsylvania’s Three Mile Island nuclear plant to power its data centers. “Because of DeepSeek, the expected price of compute probably doesn’t justify now turning back on some of these nuclear plants, or these other high-cost energy sources,” Katzman told me.
Lastly, it remains to be seen what nascent applications cheaper models will open up. “If somebody, say, in the Philippines or Vietnam has an interest in applying this to their own decarbonization challenge, what would they come up with?” Bullard pondered. “I don’t yet know what people would do with greater capability and lower costs and a different set of problems to solve for. And that’s really exciting to me.”
But even if the AI pessimists are right, and these newer models don’t make AI ubiquitously useful for applications from new drug discovery to easier regulatory filing, Su told me that in a certain sense, it doesn't matter much. “If there was a possibility that somebody had this type of power, and you could have it too, would you sit on the couch? Or would you arms race them? I think that is going to drive energy demand, irrespective of end utility.”
As Su told me, “I do not think there’s actually a saturation point for this.”
The model for 7 p.m. on June 12. Darker colors indicate higher PM2.5 levels, the particles associated with wildfire smoke. The numbered circles refer to the number of regional wildfires.
The model for 7 a.m. on June 13. 
The model for 1 p.m. on June 13. 
