A cheaper AI system that reportedly uses a small fraction of the resources consumed by competing U.S. models upended assumptions about future electricity demand late last month when it was announced by Chinese startup DeepSeek.
The surprise news of DeepSeek’s R1 model drove down expectations or share values for electricity generation, small modular nuclear reactor, uranium, gas, and tech companies, among others, temporarily knocking nearly 17% off the stock price of Nvidia, one of the main tech companies at the centre of the AI craze.
“AI’s energy needs have led companies such as OpenAI, Alphabet Inc., and Microsoft Corporation to seek new sources of power, such as shuttered nuclear plants. It has also complicated their ambitious climate goals,” Bloomberg reported at the time. “DeepSeek’s model appears to be more efficient and can achieve the same results for a fraction of the energy use, which may mean AI will have a smaller climate impact than thought.”
Just days after Donald Trump unveiled his US$500-billion Stargate AI initiative—reportedly enraging “first pal” and wannabe co-president Elon Musk by leaving him off the podium for the big reveal—DeepSeek “unveiled a large language model that can compete with U.S. giants but at potentially a fraction of the cost,” The Associated Press reported. “DeepSeek had already hit the top of the chart for free apps on Apple’s App Store by Monday morning [January 27], and analysts said such a feat would be particularly impressive given how the U.S. government has restricted Chinese access to top AI chips.”
The pushback came a scant four days later, with the New York Times reporting that the DeepSeek chatbot’s answers “reflect China’s view of the world” on topics that included late United States president Jimmy Carter’s position on Taiwan, China’s repression of Uyghurs in Xinjiang, the country’s handling of the COVID-19 pandemic, and Russia’s war in Ukraine. Some of that critical analysis came from NewsGuard, a team of online misinformation specialists that “found a similar propensity for disinformation and conspiratorial ideas in ChatGPT after it became public in 2022,” the Times said.
DeepSeek also mirrors other chatbots’ tendency to “hallucinate” with search responses that are “inaccurate, irrelevant, or nonsensical”, and follows the Chinese government’s censorship rules with some of the topics it avoids, the news story stated.
Scrambled Demand Forecasts
But those qualms didn’t stop DeepSeek from scrambling predictions of a massive surge in future energy demand. Ontario, for one, has been planning for a 75% increase in power generation by 2050.
“For months, energy analysts and electric grid operators have been projecting a massive rise in the amount of power [the United States] will need to support the energy-guzzling artificial intelligence industry without causing widespread blackouts,” Politico reported. “That has led to proposals to build long-range power lines and power plants of all kinds. It drove up the stock of energy companies promising to fuel AI data centres. And it kicked off arguments about who will pay for such massive investments and whether [the United States] can afford to support AI’s energy habits without releasing more of the carbon pollution driving climate change.”
Those assumptions crashed and burned “seemingly overnight” with the arrival of a ChatGPT-like AI that is said to run far more efficiently and cheaply. The news prompted OpenAI CEO Sam Altman to declare DeepSeek’s model “impressive, particularly around what they’re able to deliver for the price,” vowing on social media that his company “will obviously deliver much better models, and also it’s legit invigorating to have a new competitor!”
A New Set of Assumptions
But in an in-depth thread on Bluesky, award-winning AI journalist Karen Hao traced the deeper implications of DeepSeek’s announcement, far beyond a top-line narrative focused on tech competition between China and the U.S. “The biggest lesson to be drawn from DeepSeek is the huge cracks it illustrates with the current dominant paradigm of AI development,” she wrote. With an approach that requires 1/50 the resources of the current market leaders, “DeepSeek has demonstrated that scaling up AI models relentlessly, a paradigm OpenAI introduced and champions, is not the only, and far from the best, way to develop AI.”
Until DeepSeek burst their bubble, OpenAI and their “peer scaling labs” had sold the idea that constantly scaling up their systems was the best route to artificial general intelligence (AGI), the point where an AI can match or exceed human intelligence. But “this has always been more of an argument based in business than in science,” Hao said.
“There is empirical evidence that scaling AI models can lead to better performance. For businesses, such an approach lends itself to predictable quarterly planning cycles and offers a clear path for beating competition: Amass more chips,” she explained.
But “there are myriad huge negative externalities of taking this approach—not least of which is that you need to keep building massive data centres, which require the consumption of extraordinary amounts of resources”—distorting power supplies, consuming drinking water, extending the lives of gas and coal plants, worsening air quality, accelerating the carbon pollution that worsens the climate crisis, and with the more recent hype around Stargate, “ceding more and more control over critical energy and water infrastructure to Silicon Valley.”
But with DeepSeek, all of that may have changed.
“Scientifically, there’s no law of physics that says AI advancements must come from scaling rather than approaches using the same or fewer resources. Scaling is just an incredibly easy-to-follow formula,” Hao wrote. “OpenAI has been burning through staggering sums of cash to keep up its scaling paradigm and has yet to figure out how to balance its chequebooks—and it turns out it didn’t need to spend so much cash.”
So “it doesn’t matter if you’re a company in the U.S., China, or elsewhere. DeepSeek should be a cue to pivot hard toward investing in far more efficient methods of AI development. Even if you care nothing about community and climate impacts, it’s just better business.”
We’ve Been Here Before
While Hao had by far the most detailed assessment of DeepSeek’s advantage and OpenAI’s assumptions, energy modellers reacted swiftly to the news. The Financial Times headlined that DeepSeek had exposed the “guesswork” in the massive push for more electricity generation to meet AI demand.
“This abrupt reaction highlights that the market currently does not yet have adequate tools and information to assess the outlook for AI-driven electricity demand,” said Thomas Spencer, a lead energy and AI analyst at the International Energy Agency.
Investment analysts at Citi said robust growth projections for companies that supply power to data centres had been based on the sky-high growth projections, and “more computationally efficient AI could bring these trends into question.”
The recent history also shows that “one simply cannot rely on electricity demand forecasts that come from the electricity generation industry and their entourage,” Corporate Knights Director of Research Ralph Torrie wrote on LinkedIn. “Hopefully the hype around AI will get sorted out before the wave of malinvestment it might otherwise have triggered. Will we ever learn?”
“Many researchers had already believed the AI electricity demand forecasts were inflated,” agreed Joe Romm, a former acting assistant secretary of efficiency and renewables in the U.S. Department of Energy and long-ago editor of the popular Climate Progress newsletter. More specifically, “most of the hype around a nuclear renaissance and small modular reactors (SMRs) has been built around these gargantuan (over)estimates of power demand growth for AI,” even though “SMRs are particularly ill suited for data centres.”
