Chinese AI startup DeepSeek stunned the world with the release of its R1 model, which appears to perform nearly as well as leading models from Google and OpenAI, despite the company’s claim that it used a relatively modest number of GPUs to train it.
DeepSeek’s relative efficiency has experts and investors questioning whether AI really needs the massive hardware outlays everyone had been predicting. And that could change data center demand — and the energy needed to power them.
The company claims it ran 2,048 Nvidia H800 GPUs for two months to train a slightly older model, a fraction of the compute that OpenAI is rumored to use.
Few companies are as exposed as Nvidia, the share price of which was down 16% at the time of publishing. Perhaps even more vulnerable are the startups and power producers that are betting big on new nuclear and natural gas capacity.
Nuclear power, in particular, has been on the cusp of a renaissance for years, driven by advances in fuel and reactor designs that promise to make a new generation of power plants safer and cheaper to build and operate. Until now, there was little reason to blaze ahead. Nuclear is still expensive relative to wind, solar, and natural gas. Plus, next-generation nuclear has yet to be tested at commercial scale.
The surge in power demand from AI changed the equation. With data centers predicted to consume as much as 12% of all electricity in the U.S. — more than triple their share in 2023 — and forecasts of underpowered AI data centers by 2027, tech companies have been racing to secure new supplies, and throwing billions of dollars at the problem. Google has pledged to buy 500 megawatts of capacity from nuclear startup Kairos, Amazon led a $500 million investment in another nuclear startup, X-Energy, and Microsoft is working with Constellation Energy on a $1.6 billion renovation of a reactor at Three Mile Island.
But what if the problem has been overblown?
There is no hard and fast rule suggesting that the only way to improve AI performance is to use more compute. For a while, that tactic worked well, but more recently, more compute hasn’t yielded the same results. AI researchers have been casting about for solutions, and it’s possible that DeepSeek found one for its R1 model.
Not everyone is convinced, of course.
“While DeepSeek’s achievement could be groundbreaking, we question the notion that its feats were done without the use of advanced GPUs,” Citigroup analyst Atif Malik wrote.
Still, history suggests that even if DeepSeek is hiding something, someone else will probably find a way to make AI cheaper and more efficient. After all, it’s easier and potentially faster to task some PhDs with developing better models than it is to build new power plants.
The current wave of new reactors aren’t scheduled to come online until 2030, and new natural gas power plants won’t be available until the end of the decade at the soonest. In that context, tech companies’ power investments appear to be hedges in case their software bets don’t pan out.
If they do, expect tech companies to scale back their power ambitions. When given the choice between spending billions on physical assets or software, tech companies almost always chose the latter.
Where will that leave nuclear startups and energy companies? It depends. Some might be able to produce power at a low enough cost that it won’t matter if AI’s power needs ebb. The world is electrifying, and even before the AI bubble started inflating, demand for electricity was expected to grow.
But absent demand from AI, those cost pressures are probably going to increase. Wind, solar, and batteries are cheap and getting cheaper, and they’re inherently modular and mass-produced. Developers can roll out new renewable plants in phases, delivering electricity (and revenue) before the entire project is complete while offering some control over their future in the face of uncertain demand. The same can’t be said of a nuclear reactor or a gas turbine. Tech companies know this, which is why they’ve been quietly investing in renewables to power their data centers.
Few people predicted the current AI boom, and it’s unlikely that anyone knows how the next five years will play out. As a result, the safer bets in energy will probably flow to proven technologies that can be rapidly deployed and scaled according to a rapidly evolving market. Today, renewables fit that bill.
