Nuclear vs. the alternatives: what should power AI?
A data center needs a huge amount of power every minute of every day. Nuclear, natural gas, solar with batteries, wind, geothermal and the grid all compete for the job. Here's how they stack up on cost, reliability, speed and pollution, and what a nuclear engineer would pick.
Plain-English explainer by Joe Miller, a nuclear engineer. Not investment advice. Figures are the latest public numbers we could find and change fast; sources are at the bottom of each page.
What a data center needs from its power
- Power every minute of the year. An AI data center runs around the clock. A few seconds without power can crash a training run worth millions of dollars.
- Big blocks of power. A new AI campus wants 100 megawatts to more than 1,000 megawatts, as much as a city.
- Speed. Companies want power in one to three years, not ten. The wait to connect to the grid in many places is now several years.
- Low cost and low carbon, because power is the biggest running cost, and most of these companies promised to cut emissions.
No single source wins on all four. That's why the answer is almost always a mix.
Which sources run around the clock?
What new power costs
Those numbers are for new plants. Power from existing nuclear plants, which were paid off long ago, is far cheaper; that's why tech companies are racing to sign deals with plants that are already running (see why Big Tech is buying nuclear). Lazard also found that battery storage costs went up in 2026 because of tariffs on imported batteries, and that new gas plants cost more because turbines are scarce.
Side by side
| Source | Runs 24/7? | Time to new power | Carbon | The catch |
|---|---|---|---|---|
| Existing nuclear (restart or buy) | Yes, about 93% of the time | 0 to 4 years | None | Only a few plants left to restart; buying from running plants takes power away from other customers |
| New nuclear (big or small reactors) | Yes | About 2030 at the earliest; big plants about 10 years | None | Most expensive new power; first-of-a-kind projects run late |
| Natural gas, combined cycle | Yes, when fueled | 3 to 5 years; turbines sold out into 2029-2030 | About half of coal | Fuel prices swing; turbine shortage; pollution near the plant |
| Gas turbines on site ("behind the meter") | Yes | Months, if you can find turbines | Higher than combined cycle | Costly to run; air permits and neighbors (see Colossus) |
| Solar plus batteries | Only with enough batteries | 1 to 2 years | None while running | Needs lots of land and 4+ hours of batteries; nights and cloudy weeks |
| Wind | No, about 34% of the time | 2 to 4 years | None while running | Wind comes and goes; transmission lines needed |
| Geothermal | Yes, about 70% or more | 3 to 5 years | Very low | Only a few hundred megawatts signed so far; best sites in the West |
| Coal | Yes, when run | No new plants being built | Highest | Old plants kept open longer for data centers; most pollution |
| The grid (a mix of everything) | Yes | Several years in many regions | Depends on region | Waiting lines to connect; costs can spread to homeowners |
Times and ranges are typical U.S. figures and vary a lot by project.
Each option, in plain English
Nuclear
The best fit on paper: it runs about 93% of the time, makes no carbon, and a 1,000-megawatt plant fits on about a square mile. The problem is time and cost. Restarting a closed plant takes two to four years, and only a handful are left. New reactors, including small modular reactors, are aimed at about 2030 at the earliest and cost the most of any new power. Big Tech has still signed close to 10 gigawatts of nuclear deals, because nothing else gives carbon-free power around the clock.
Natural gas
Today's workhorse. Gas is the largest source of U.S. electricity, it can run on demand, and new combined-cycle plants are relatively cheap. About 30% of planned data center capacity is set to use power built on site, mostly gas, according to a Cleanview analysis reported by E&E News. The catch: turbines are sold out. GE Vernova, Siemens Energy and Mitsubishi have a combined backlog of about 220 gigawatts, with delivery slots stretching into 2029 and 2030, and Wood Mackenzie expects turbine prices to be nearly triple their 2019 level by the end of 2027. Gas also burns fuel whose price swings, and it emits carbon and smog-forming pollution.
Solar plus batteries
The cheapest new energy and the fastest to build. Solar went to almost 9% of U.S. electricity in 2025. The catch: it only produces about a quarter of the time. To run a data center all night, you need a large battery bank, and to ride through a cloudy week you still need something else. Land is the other limit: about 5 to 10 acres per megawatt of panels.
Wind
Cheap when it blows, often strongest at night, which pairs well with solar. The catch: it runs about a third of the time, and the best wind is often far from where data centers are being built.
Geothermal
The dark horse. Heat from deep underground runs around the clock with almost no emissions. New drilling methods borrowed from the oil and gas industry are opening up more sites. Google has signed with Fervo and Ormat, and Meta with Sage Geosystems and XGS Energy, each deal in the 115 to 150 megawatt range. The catch: it's still small, and the best sites are in the West.
Coal and the grid
No one is building new coal plants, but some utilities are keeping old ones open longer to meet data center demand. Most data centers still plug into the grid, which is a mix of everything above. The problem is the line to connect: several years in many regions, and the cost of new grid capacity can land on everyone's bill (see your electric bill).
Why gas is winning right now
If you need a gigawatt in two years, gas is usually the only thing you can get, if you can find turbines. Developers are using truck-mounted generators and turbines taken from aircraft and ships to get power faster. Elon Musk's Colossus in Memphis ran dozens of gas turbines on site and ended up in court over air permits (see Elon Musk, Colossus and data centers in space). Expect most new data center power in the 2020s to be gas, with nuclear, geothermal and solar plus batteries growing behind it.
My verdict as a nuclear engineer
Short term, through about 2030: gas and the grid do the heavy lifting, with solar and batteries cutting fuel use during the day. Nuclear restarts like Crane (2027) and Duane Arnold (2029) add a few gigawatts of clean, steady power.
Long term, the 2030s: if small reactors come in close to their targets, nuclear is the best match for what a data center needs: steady, compact, carbon-free and immune to fuel price swings for decades. Geothermal is the one to watch as a second clean, round-the-clock option. Solar and wind will keep growing because they're cheap, but they need storage or backup to run a data center.
What it means for you: the mix your local data center picks decides your air quality, your water and, through the grid, part of your electric bill. Ask your utility commission and local officials which power a new data center will use, and who pays for it.
Members: $15 a month
Half the topics are free. Members get the other half, plus the monthly Data Center Critical newsletter with the month's AI, chip and power news. Cancel anytime.
Sources
Checked November 16, 2026.
- Utility Dive: Renewables remain cheapest, but LCOE is rising, Lazard says (July 16, 2026)
- Lazard: Levelized Cost of Energy+ (2026)
- Turbomachinery International: The gas turbine bottleneck (Sept. 11, 2026)
- E&E News: Data centers increasingly turn to off-grid gas (Feb. 5, 2026)
- U.S. Energy Information Administration: Electricity generation, capacity and sales
- Solar Power World: Renewables produced more than 25% of U.S. electricity in 2025 (Feb. 25, 2026)
- Enki AI: Geothermal deals for data centers (2025-2026)
- TechCrunch: Google invests in Fervo's $462M round (Dec. 10, 2025)
- ESS News: Battery storage costs up 27% since 2020, says Lazard (July 13, 2026)
- Forbes: The AI boom is making nuclear power bankable again (July 26, 2026)