Sign In or Create an Account.

By continuing, you agree to the Terms of Service and acknowledge our Privacy Policy

Energy

What DeepSeek Means for AI Energy Demand

It’s complicated!

The Deepseek logo on wires.
Heatmap Illustration/Getty Images

It took the market about a week to catch up to the fact that the Chinese artificial intelligence firm DeepSeek had released an open-source AI model that rivaled those from prominent U.S. companies such as OpenAI and Anthropic — and that, most importantly, it had managed to do so much more cheaply and efficiently than its domestic competitors. The news cratered not only tech stocks such as Nvidia, but energy stocks, as well, leading to assumptions that investors thought more-energy efficient AI would reduce energy demand in the sector overall.

But will it really? While some in climate world assumed the same and celebrated the seemingly good news, many venture capitalists, AI proponents, and analysts quickly arrived at essentially the opposite conclusion — that cheaper AI will only lead to greater demand for AI. The resulting unfettered proliferation of the technology across a wide array of industries could thus negate the energy efficiency gains, ultimately leading to a substantial net increase in data center power demand overall.

“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.”

Green

You’re out of free articles.

Subscribe today to experience Heatmap’s expert analysis 
of climate change, clean energy, and sustainability.
To continue reading
Create a free account or sign in to unlock more free articles.
or
Please enter an email address
By continuing, you agree to the Terms of Service and acknowledge our Privacy Policy
Politics

AM Briefing: Climate United Sues the EPA

On the Greenhouse Gas Reduction Fund, Canada’s new prime minister, and CERAWeek

The Battle Over Climate Grants Is Heating Up
Heatmap Illustration/Getty Images

Current conditions: Firefighters successfully controlled brush fires in Long Island that prompted New York Gov. Kathy Hochul to declare a state of emergency • Brisbane, Australia, recorded its wettest day in more than 50 years • Forecasters are keeping an eye on a storm system developing across the central U.S. that could pack a serious punch this week.

THE TOP FIVE

1. Nonprofit sues EPA, Citibank, over missing climate funds

The nonprofit Climate United filed a lawsuit over the weekend against the Environmental Protection Agency and Citibank for withholding $7 billion in climate funds awarded as part of the Biden administration’s Inflation Reduction Act. The move escalates a dispute over some $20 billion in grants from the IRA’s Greenhouse Gas Reduction Fund, which was designed to help mobilize private capital toward clean energy and climate solutions. President Trump’s EPA Administrator Lee Zeldin has been on a mission to claw back the funds, claiming their distribution was rushed and mismanaged. In its lawsuit, Climate United says it has been unable to access the $7 billion it was awarded, and that the EPA and Citibank have given no explanation for this. It wants a judge to order that the money be released. “We’re not trying to make a political statement here,” Beth Bafford, chief executive of Climate United, toldThe New York Times. “This is about math for homeowners, for truck drivers, for public schools — we know that accessing clean energy saves them money that they can use on far more important things.” The Trump administration has reportedly demanded that the eight organizations tapped to receive the money turn over records to the FBI and appear in federal court later this month.

Keep reading...Show less
Yellow
Climate Tech

AI Weather Forecasters Still Need NOAA

While they’re getting more accurate all the time, they still rely on data from traditional models — and possibly always will.

A robot forecaster.
Heatmap Illustration/Getty Images

The National Oceanic and Atmospheric Administration has had a bruising few weeks. Deep staffing cuts at the hands of Elon Musk’s efficiency crusaders have led to concerns regarding the potential closure of facilities critical to data-gathering and weather-forecasting operations. Meteorologists have warned that this could put lives at risk, while industries that rely on trustworthy, publicly available weather data — from insurance to fishing, shipping, and agriculture — are bracing for impact. While reliable numbers are difficult to come by, the agency appears to have lost on the order of 7% to 10% of its workforce, or more than 1,000 employees. NOAA’s former deputy director, Andrew Rosenberg, wrote that Musk plans to lay off 50% of the agency, while slashing its budget by 30%.

Will that actually happen? Who the heck knows. But what we can look at are the small cracks that are already emerging, and who could step in to fill that void.

Keep reading...Show less
Yellow
Climate

AM Briefing: U.S. Abandons a Key Climate Financing Coalition

On energy transition funds, disappearing butterflies, and Tesla’s stock slump

America’s Shrinking Climate Financing Footprint
Heatmap Illustration/Getty Images

Current conditions: Australians have been told to prepare for the worst ahead of Cyclone Alfred, and 100,000 people are already without power • Argentina’s Buenos Aires province has been hit by deadly flooding • Critical fire conditions will persist across much of west Texas through Saturday.

THE TOP FIVE

1. Trump administration questions aid programs about their climate ambitions

Many foreign aid programs have reportedly received a questionnaire from the Trump administration that they must complete as part of a review, presumably to help the government decide whether or not the groups should receive any more federal funds. One of the questions on the list, according toThe New York Times, is: “Can you confirm this is not a climate or ‘environmental justice’ project or include such elements?” Another asks if the project will “directly impact efforts to strengthen U.S. supply chains or secure rare earth minerals?” President Trump issued an executive order freezing foreign aid on his first day back in office. The Supreme Court subsequently ruled that aid must be released. The Times notes that “many of the projects under scrutiny have already fired their staff and closed their doors, because they have received no federal funds since the review process ostensibly began. … Within some organizations, there are no staff members left to complete the survey.”

Keep reading...Show less
Yellow