The future of artificial intelligence was bound to be power-hungry. As AI systems like ChatGPT, Gemini, and Claude grew more expansive, so did their energy needs — skyrocketing demand for electricity from data centers and putting a strain on the U.S. power grid. Projections suggested that AI could account for up to 25% of U.S. electricity consumption by 2030, fueling a massive buildout of fossil fuel and nuclear plants to keep the data centers running.
Already, plans were laid out to build nuclear and coal plants to power AI’s surge. Then came DeepSeek, a Chinese AI company that seemingly rewrote the rules overnight. But this particular story is far from written.
Did DeepSeek just pull the plug on AI’s energy crisis?
DeepSeek’s most notable breakthrough was achieving the same results as its Western counterparts with a fraction of the energy. This sent shockwaves through financial markets. Utility stocks plunged. Chipmakers, especially Nvidia, lost hundreds of billions in value.
Granted, markets are known for their knee-jerk reactions. Nvidia gained back some of its losses, and other Big Tech companies like Meta seemed to brush off the news. But the energy question still loomed: have AI’s projected electricity demands been wildly overestimated?
Scaling AI seemed like a brute force, “bigger is better” type of problem. State-of-the-art models required more chips and more energy for training. Data centers, which house the servers powering AI models, already consume about 4.4% of U.S. electricity, and the expectation was for that figure to rise exponentially. Projections from Morgan Stanley, Wells Fargo, and other analysts varied widely, but many expected AI to add hundreds of terawatt-hours of electricity demand — enough to power millions of homes.
Utilities and energy companies rushed to build new power plants to accommodate the surge. Some of the largest tech firms, including Microsoft, Meta, and Google, scrambled to secure energy contracts. Some even struck deals to restart old nuclear plants.
Then came the DeepSeek bombshell. What if, suddenly, you needed far less power?
The company claimed that its model used 10 to 40 times less energy than its U.S. competitors. Even more stunning: DeepSeek said it trained its AI model for just $5.6 million, a tiny fraction of the hundreds of millions — sometimes billions — of dollars spent by OpenAI, Google, and Anthropic.
Granted, everyone sort of just believed DeepSeek, which is a little questionable. They could just be flat-out lying or exaggerating. But while they seemingly came out of nowhere, their models have been causing waves in China for the better part of a year.
Where Does the Innovation Come From?
In May 2024, DeepSeek released an AI model that seemed to offer strong performance for a very low price. While that model (DeepSeek-V2) wasn’t as good as the best in the field, it quickly triggered a price war in China. It prompted other major tech giants like ByteDance or Tencent to cut the price for their own AI models to be able to compete. Yet even as these other companies were losing money, DeepSeek was apparently profitable. They may be fudging the numbers, but whatever DeepSeek is doing, it seems to be working.
The secret seems to be that DeepSeek focused on efficiency rather than brute force. Instead of throwing thousands of power-hungry GPUs at training its AI, the company optimized its model architecture to run on older, less powerful Nvidia chips — the only ones legally available to Chinese companies under U.S. trade restrictions.
This efficiency-first approach blew apart assumptions about how AI needs to be trained. Instead of massive electricity consumption scaling endlessly, DeepSeek essentially suggested that you could train (and use) AI at a much lower energy cost.
Suddenly, the AI electricity boom wasn’t guaranteed.
Except this doesn’t necessarily mean that overall demand for AI will shrink. In fact, it might mean the opposite.
A classic paradox
This could be a classic example of the Jevons paradox — a phenomenon in which greater efficiency leads to greater overall consumption. The logic is simple: If AI becomes cheaper and more efficient, more people and businesses will use it. It’s like how building an extra lane on a highway doesn’t necessarily lead to less congestion. This is because more lanes mean more people will want to drive this route.
AI is also very scalable and can be used in numerous different ways. If you make it cheaper and more accessible, it could actually bring demand higher.
Simply put, if DeepSeek’s breakthrough (or anything else) makes AI accessible to more users, the total number of AI queries could skyrocket, offsetting the efficiency gains. The same thing happened with computer chips — once they became more efficient, computers proliferated, and overall energy use still went up. You just shift from a few big consumers to more smaller ones.
Despite the uncertainty, energy companies are still racing ahead with their AI-driven power plans.
Good or Bad News For the Planet?
Microsoft, Meta, and Oracle continue to build massive data centers, and they don’t seem to be slowing down. Meanwhile, Trump’s energy policies favor fossil fuels, with the administration wanting to make sure that gas and coal are central to AI power generation.
DeepSeek has introduced doubt. Utility companies that had been planning massive investments in new power plants may need to reassess their projections. Some speculate that there’s an AI bubble, similar to the dot-com bubble from 25 years ago. Just as telecom companies poured $500 billion into fiber-optic networks in the end of the 1990s, only to find that demand lagged behind expectations, today’s utilities, chipmakers, and data center developers are racing to expand power generation and infrastructure based on high AI energy forecasts —forecasts that DeepSeek’s efficiency breakthrough just called into question.
But it remains to be seen whether this will happen here too or we’ll see some form of Jevons paradox.
It’s still early days. AI will still require a lot of electricity, but we don’t know how much.
The AI-electricity story isn’t over — it’s just entering a new phase. DeepSeek may have disrupted the AI power market, but whether it’s the start of a revolution or just a speed bump remains to be seen for now.
