nuclear-news

Nuclear power is a parasite on AI’s credibility

The IEA expects renewables to add 10–20-fold more electricity supply than data centers raise demand. Renewables and storage are already 93% of US and 95% of world electric capacity expected to be added this year. These cheapest and fastest options can come online sooner than a data center, and they already reliably power critical loads like data centers.

Global nuclear power in a good year adds only as much net capacity as renewables add every two days

 July 20, 2025 by beyondnuclearinternational, https://beyondnuclearinternational.org/2025/07/20/nuclear-power-is-a-parasite-on-ais-credibility/

The following is the press release announcing a new essay by Stanford University’s Amory Lovins, Artificial Intelligence Meets Natural Stupidity: Managing the Risks.

Claims of soaring electricity needs to power ravenous new AI data centers underpin the energy emergency declared for national security. Yet new research synthesized by a prominent energy expert, Amory Lovins, explains how hidden order-of-magnitude uncertainties in AI’s energy needs are risking major speculative losses and energy-market distortions—and he highlights timely remedies.

In fact, US electricity use fell in 2023, and in 2024, it rose only 2%—less than in three other years of the past ten. Forecasts of future electricity use have lately risen, especially in a few hotspots that promote and subsidize new data centers.

Yet that’s far from a broad trend, and most of the forecast growth is for other or reshored industries, electric vehicles, and electrifying buildings and factories. Data centers used only about 4.5% of US electricity in 2024. Of global electricity growth, the International Energy Agency (IEA) says only 5% in 2024 was for new data centers, rising to 5–10% of growth in 2025–30. Both nationally and globally, most data centers aren’t even made or run for AI; they’re for traditional functions like search engines, e-mail, and e-commerce.

Big Tech firms are indeed investing at least a trillion dollars in new AI data centers. Hundreds are planned, some as power-hungry as a small city. However, only a small fraction of those proposed are likely to be built, and not all those built are certain to thrive. Overforecasting seems endemic, severe, and underrecognized. It’s caused by peculiarities of the current data-center marketplace. But underlying those are many fundamental unknowables—even about the dominant model’s basic validity.

These make future demand for AI services extremely uncertain. Industry leaders and analysts warn of a potential financial bubble. Moreover, the electricity needed to produce a given amount of AI service is durably falling by about fourfold every year. That’s faster than purchases of AI services (costly to produce, but now often given free as bait) seem set to keep growing, yielding the revenue to buy the electricity. 

Assuming explosive growth in power for AI thus looks like a double bubble that can cause toil and trouble for utility investors and for other electricity customers, as Utility Dive reported yesterday. Two precedents counsel caution: widely believed 1999 coal-industry claims that information technology would use half of US electricity by 2020 proved about 2400% too high, and in 2010–18, the world’s data centers did 550% more computing with 6% more electricity. Lovins offers important new precautions and solutions. 

Even ambitious forecasts of AI’s electricity needs could be met by any of three proven methods: running data centers more flexibly on rare occasions without materially compromising service; freeing up supplies that are now largely wasted by other customers; and siting new data centers and clean energy (solar, wind, storage) together near little-used existing gas plants. 

“Both time- and location-based adaptability offer promising pathways to transform data centers from electricity liabilities into grid and regional assets,” according to Stanford’s Amory Lovins. “By aligning compute demand with cleaner energy availability—whether by time-shifting workloads or siting them in regions with surplus renewables—data centers can support grid resilience, reduce carbon intensity and other impacts, save infrastructure, and cut cost, if not distorted by short-term economic incentives such as local tax breaks.

The IEA expects renewables to add 10–20-fold more electricity supply than data centers raise demand. Renewables and storage are already 93% of US and 95% of world electric capacity expected to be added this year. These cheapest and fastest options can come online sooner than a data center, and they already reliably power critical loads like data centers.

Fossil and nuclear plants, both favored by federal policy, would be far slower and costlier: turbines for new gas plants are sold out to at least 2031, and global nuclear power in a good year adds only as much net capacity as renewables add every two days.

“Renewables’ high speed and low cost have run off with the world power market. For anyone who reads the data and respects the market, it’s game over. Nuclear energy is a parasite on AI’s credibility. Pairing them makes them both less investable,” added Lovins. 

Buying slower, costlier power by misunderstanding AI and grids risks higher retail rates and painful investor losses—as occurred when hundreds of unneeded power plants were built in a similar panic a quarter-century ago (part of the dot-com bubble’s $5-trillion losses).

Utility regulators should protect the public from these speculative risks by requiring data-center developers to post a bond or insurance policy guaranteeing full payment for their future power needs. Then the parties seeking profits will bear the independently priced risks that their projects create. Responsible AI use may also reduce the risk that AI-enabling more and cheaper oil and gas production may emit more carbon than AI saves.

Physicist Amory Lovins is a globally recognized expert on energy productivity, renewable energy, and sustainable design based at Stanford University. The full article is available for download.

July 21, 2025 - Posted by Christina Macpherson | ENERGY