Nuclear power is failing, and AI can’t rescue it

Amory Lovins teaches engineering at Stanford, and is cofounder and chairman emeritus of RMI. 

An intensive influence campaign seeks to resurrect a “nuclear renaissance” from the industry’s slow-motion collapse documented in the independent annual World Nuclear Industry Status Report. Claims that past failures won’t recur have convinced many politicians that socializing nuclear investments rejected by private capital markets, weakening or bypassing rigorous safety regulation, suppressing market competition, and commanding military reactor and data-center projects as a national-security imperative will restore nuclear expansion and transform the economy.

This illusion neatly fits the industry’s business-model shift from selling products to harvesting subsidies.

A few awkward facts intrude. Even the most skilled firms and nations keep delivering big reactors with several times the promised cost and construction time. A swarm of startup firms that have never built a reactor are dubiously rebranding their inexperience as a winning advantage. New designs are said to be so safe they don’t need normal precautions (though not safe enough to waive nuclear energy’s unique exemption from accident liability). Political interference in nuclear licensing is eroding public confidence. Proposed smaller reactors cost more per kWh, produce more nuclear waste per kWh, and often need more-concentrated fuel directly usable for nuclear weapons.

And nuclear power faces the same fundamental challenges as fossil fuels: uncompetitive costs, runaway competitors, dwindling profits, and uncertain demand. Few if any vendors have made profits selling reactors — only fueling and fixing them. Nuclear electricity loses in open auctions, so only Congressional bailouts — $27 billion ($15 billion paid out) in 2005, $133 billion in 2021-22, tens of billions more in 2025 — saved most existing U.S. reactors from closure.

Now comes another vision: powering the glorious new world of artificial intelligence. This may be a trillion-dollar bubble, but it’s sellable until market realities intervene. The International Energy Agency expects data centers, mostly non-AI, to cause only a tenth of global electricity demand growth to 2030, doubling their share of usage — to just 3%. So AI won’t eat the grid. But IEA forecasts renewables will power data-center growth 10-20 times over, while Bloomberg NEF predicts over 100. Nuclear lost the race to power the grid, so new reactors have no business case or operational need.

Each year, nuclear adds as much net global capacity as renewables add every two days. Soaring renewables generate three times more global electricity than stagnant nuclear power, whose 9% world and 18% U.S. shares keep shrinking. In 2023-24, China added 197 times more solar and wind than nuclear capacity, at half the cost. In May, China added 93 GW of solar, or 3 GW per day.

Despite having turned nuclear power into a minor distraction, renewables are dismissed as “intermittent.” Again, facts intrude.

Military and industrial installations already prefer 100% renewables for their most critical applications, including Apple’s data centers in four states. Ten kinds of carbon-free resources can balance variable (but highly predictable) renewables, keeping the grid stable. Using a small subset, power systems with modest or no hydropower already sustain such annual renewable fractions of electricity use as Denmark 88+%, South Australia 74% (expecting 100% in two years), and Germany 54%.

And since a nuclear kWh costs several to many times more than a renewable or saved kWh — even more if nuclear load-follows to “complement” rather than curtail renewables — nuclear displaces less fossil fuel per dollar (or year), making climate change worse.

Nonetheless, nuclear power is being boosted by fierce lobbying and federal policy as essential for new AI data centers vital for prosperity and security. This case can’t withstand scrutiny. My essay “Artificial Intelligence Meets Natural Stupidity: Managing the Risks” shows:

• Data centers use about 4.5-5% of U.S. and 1.5% of world electricity, and lately caused only about 5% of world electricity demand growth. Of all data-center electricity, about one-fourth in the U.S. or one-ninth globally is for AI, the rest for traditional uses.
• Claims of soaring AI electricity use are projections, not realities, except in a few “hot spots” like two Virginia counties. In 2023, AI added roughly 0.04% to world and 0.1% to U.S. electricity use.
• Most proposed AI data centers are speculative and unlikely to get built; many built won’t thrive. Major power-supply investments risk getting stranded.
• Demand for AI services is enormously uncertain. So is their business case: AI’s proven value in narrowly specialized technical applications looks too small to repay its immense investments. Many general users don’t need or want to pay for AI.
• Big Tech firms rarely sign specific nuclear power purchase agreements. Much of the hype is about vague statements of interest in buying electricity timely at an attractive price, or modest, symbolic investments. Big Tech rightly prefers renewables as faster, surer and cheaper.
• The efficiency of turning electricity into AI services roughly quadruples each year, so a new data center must roughly quadruple its sales of AI services each year for decades to keep using and paying for the same amount of electricity — a tall order.
• This spring, innovators showed how operating AI data centers slightly more flexibly without compromising service can power at least the next decade of U.S. AI growth with no new generators, stranding more electricity and gas investments.
• The coal industry’s 1999 campaign to create panic that the Internet would falter without huge power expansions misled investors, worsening a 2000-02 bloodbath when hundreds of new power plants weren’t needed. Today’s trends, pushing an AI case for unsellable and too-late nuclear and gas projects, rhyme with that disaster.

The latest risk to the AI/nuclear case came into focus in Sparks, Nevada, in June, when Redwood Energy (a new activity of dominant battery-recycler Redwood Materials) revealed North America’s largest microgrid.

Twenty MW-DC of photovoltaics are laid flat on level ground. Water-recovering Roomba-like crawlers clean them nightly. About 800 battery packs from retired or crashed cars — the world’s largest use of second-life batteries — are wrapped in white plastic and set on cinderblocks, safely separated. They’re good for another few years, then hot-swappable. Novel power electronics and software meld those diverse batteries into 63 MWh of storage with 2-48-hour nominal duration. (Redwood Energy is already engineering similar microgrids an order of magnitude larger, enough to run most existing data centers.)

The resulting 100%-solar microgrid produces 10 MW-AC of ultrareliable 24/7/365 power that runs modular Crusoe data centers onsite, eliminating transmission costs, losses and approvals. This all-solar power is more reliable than grid power, cheaper than the utility’s 8¢/kWh retail price, and all built in four months.

Thus, we needn’t guess or debate whether a particular data center will get built and flourish. Instead, we can commit to build its onsite solar power plant, perhaps by competitive procurement, only when the data center’s 1.5 to 2.5 years of construction is mostly done. Needing no grid connection, the solar microgrid needs few if any approvals — just cheap land. It’s inherently safe, silent, automatic, virtually water- and maintenance-free, based on common commodities, zero-emission, portable and durably profitable.

Can your reactor do that? If not, why build it?