A 300-megawatt nuclear reactor sounds like a vast source of electricity. Run at its full electrical rating for every hour of a 365-day year, however, it would produce 2.628 million megawatt-hours. That is 2.628 terawatt-hours, enough to cover only a small fraction of the electricity now associated with US data centres.

Put 200 terawatt-hours in the numerator and the result is 76.1 reactors. Round the input, allow for real maintenance outages and the scale becomes roughly 80 reactors of that size. The comparison is mathematically straightforward, but understanding it requires keeping power, energy, reactor size and forecast uncertainty separate.

The 2.6 TWh figure is a unit conversion

A megawatt is a rate of producing or consuming electricity at a particular moment. A megawatt-hour is the energy delivered when one megawatt continues for one hour. A terawatt-hour is one million megawatt-hours.

The calculation is therefore 300 MW multiplied by 8,760 hours, which equals 2,628,000 MWh, or 2.628 TWh. There is no assumption about reactor physics hidden inside that conversion. “Running flat-out” simply means holding the reactor’s 300 MW electrical output without interruption for the full year.

Real reactors do stop for refuelling and maintenance. The US Energy Information Administration’s capacity-factor data show that nuclear generators averaged 91.0 percent in 2025, after 90.8 percent in 2024. At 91 percent, a 300 MW reactor would deliver about 2.39 TWh in a year rather than 2.63 TWh.

Why “around 80 reactors” is reasonable

Using the ideal upper limit, 200 TWh divided by 2.628 TWh gives 76.1 reactors. Since a fraction of a reactor is not useful in this comparison, that means 77 units running continuously. Using the recent 91 percent US capacity factor instead gives about 83.6, or 84 units. “Around 80” sits between those two ways of expressing the same scale.

This is an energy-equivalence calculation, not a construction plan. Annual energy does not reveal the highest instantaneous demand, the location of that demand or whether transmission can carry electricity from a generator to a cluster of server buildings. Data centres often seek continuous power, but they are distributed unevenly and can create acute local grid constraints even when a national annual total looks manageable.

Expressed as an average load, 200 TWh spread across a year is about 22.8 GW. That average is another way to reach the same answer: divide 22.8 GW by 0.3 GW and the result is about 76 reactors before outages. It also shows that the comparison concerns a sustained national load, not a brief surge on a hot afternoon.

The size assumption matters too. The Department of Energy describes small modular reactors as units with up to 300 MW of capacity. A 300 MW reactor is therefore small beside many conventional commercial units. A 1,000 MW reactor operating at 91 percent would generate close to 8 TWh annually, making the equivalent for 200 TWh about 25 large units rather than 80 small ones.

The firm measured number is 176 TWh

The strongest recent national baseline comes from Lawrence Berkeley National Laboratory. Its 2024 United States Data Center Energy Usage Report estimated that data centres consumed 176 TWh in 2023, about 4.4 percent of all US electricity. Usage had climbed from about 58 TWh in 2014, with most of the acceleration occurring after 2017.

That makes 200 TWh a rounded statement about the industry’s present scale, not a claim that every US data centre has been separately metered and summed for 2026. The latest measured benchmark was already much closer to 200 TWh than to 100 TWh, and the modelled trajectory rose sharply from there. If the calculation used 176 TWh instead, the result would be about 67 ideal 300 MW reactors or 74 reactors at a 91 percent capacity factor.

National estimates are needed because the category covers facilities of very different sizes and ownership structures. Hyperscale campuses attract attention, but smaller enterprise server rooms also consume electricity. Researchers assemble equipment shipments, operating assumptions and infrastructure overhead into a bottom-up model, which is why the result is an estimate rather than a utility meter reading.

The same laboratory’s original scenarios put 2028 consumption between 325 and 580 TWh. Its 2025 update, published in June 2026, extended the horizon to 2030 and estimated a range of 521 to 843 TWh. The reference case was 649 TWh, or 11.8 percent of projected US electricity consumption.

AI changes the slope, not the definition

Data centres are not synonymous with artificial intelligence. The national estimate includes conventional servers, storage systems, networking equipment, power conversion and cooling. Search engines, video streaming, financial systems, cloud software and ordinary corporate computing all contribute to the total.

AI matters because accelerator-rich servers are power dense and because demand for them has expanded rapidly. Berkeley Lab estimated that electricity used by GPU-accelerated AI servers grew from less than 2 TWh in 2017 to more than 40 TWh in 2023. Its projections vary widely because future chip shipments, utilisation rates, equipment life and cooling choices remain uncertain.

ScienceBlog has previously covered research that mapped the environmental footprint of AI data centres and examined ways to reduce it. Efficiency can slow the increase in electricity per computation, but total consumption can still rise if the volume of computing grows faster than those savings.

The forecasts stop at their modelling horizons

The Berkeley Lab scenarios provide ranges through 2028, while the laboratory’s later update extends its modelling horizon to 2030. Those end dates are not predictions that data-centre electricity consumption will crest and decline.

If the 649 TWh reference case were supplied entirely by hypothetical 300 MW reactors, it would equal about 247 units running at 100 percent or roughly 272 units at a 91 percent capacity factor. The low and high scenarios would produce still wider equivalent counts. That exercise shows why a small change in the assumed demand produces dozens of reactors’ worth of difference.

It also explains why efficiency assumptions matter. A few percentage points saved across hundreds of terawatt-hours can equal the annual output of several power stations. Conversely, higher utilisation of installed AI hardware can add a large continuous load without any change in the number printed on a chip’s nameplate rating.

The Department of Energy’s data-centre resource hub presents the forecast as a planning challenge involving new generation, grid upgrades, efficiency and demand flexibility. Those elements cannot be replaced by a single national energy ratio.

Nuclear can contribute without being the whole answer

Nuclear power offers steady output and the US already operates a fleet with about 100 GW of capacity. The EIA’s reactor-level generation and capacity data show why output must be judged over time rather than from nameplate capacity alone. Restarts, uprates and new reactor designs could add supply, but each option has its own construction schedule, financing requirements and grid connection.

Data-centre operators are also pursuing renewables, gas generation, storage, geothermal power and contracts with existing nuclear plants. Some facilities may shift workloads in time or operate backup resources during grid emergencies. The practical electricity mix will be a portfolio shaped by geography, cost, reliability and regulation.

The headline’s comparison is useful because it translates an abstract national load into generators people can picture. Its boundaries are equally important: 200 TWh is a rounded current-scale estimate, 2.6 TWh assumes perfect operation, and 80 applies specifically to 300 MW reactors. Within those limits, the arithmetic is correct.