One 300 megawatt (MW) reactor can carry a campus with a 300 MW connection, except while the reactor is down. That outage, more than the megawatts, decides how reactors and campuses pair. The megawatts still need care, because a reactor and a campus count them differently.
On 10 November, Canada's Ontario Power Generation (OPG) reported on its first small modular reactor (SMR) at Darlington. It is no longer expected to meet its October 2030 target or its cost estimate. The day before, the US Nuclear Regulatory Commission (NRC) permitted construction of four Xe-100 reactors in Texas.
In October the International Atomic Energy Agency (IAEA) published a report on nuclear energy and artificial intelligence (AI). It calls the idea of "a 500 MW reactor to a 500 MW data centre" a misconception. Here we size each reactor class against AI factories/hyperscale data centres, from 20 MW to gigawatt scale. We covered small reactor costs in September, before OPG's update.
A 200 MW single-tenant hyperscale campus needs a 200 to 240 MW connection, depending on what the 200 counts

A reactor's core makes heat, in megawatts thermal (MWt). It sends out electricity in megawatts electric (MWe). GE Vernova Hitachi's BWRX-300 design turns 870 MWt into 300 MWe net. X-energy's Xe-100 design makes 80 MWe from 200 MWt.
A campus figure may be its nameplate, the most it expects to draw from its grid connection. Or it may be its information technology (IT) capacity, what the servers alone can draw. The IAEA report gives single-tenant hyperscale campuses a power usage effectiveness (PUE) of 1.1 to 1.2. Multitenant campuses have historically run at 1.3 to 1.5.
A PUE of 1.1 to 1.2 adds 10 to 20% to the IT load, for cooling and building services. So a 200 MW hyperscale campus needs a 200 MW connection if 200 is its nameplate. It needs 220 to 240 MW if 200 is its IT capacity. The gap is up to 40 MW, half an Xe-100.
Such US campuses peak at about 80% of nameplate, on Electric Power Research Institute (EPRI) data the IAEA cites. So the campus peaks near 160 MW on the first reading, and 176 to 192 MW on the second.
NAVER and NVIDIA plan a 200 MW multitenant AI factory at Sejong, South Korea, by 2028, with no basis stated. In Japan, RHAELM and JERA's Chiba AI project, with Dell Technologies, plans power capacity of up to 400 MW.
The IAEA matches microreactors to data centres under 200 MW and 300 MWe units, in multiples, to gigawatt campuses

The IAEA report's Table 3 pairs reactor classes with data centre sizes.
Microreactors of 1 to 20 MWe suit data centres under 200 MW.
Small SMRs of about 100 MWe suit data centres of 200 to 500 MW, several units at a time.
Mid-size SMRs of about 300 MWe suit gigawatt campuses, several units at a time, and regional clusters.
Large reactors of 800 to 1,600 MWe suit multigigawatt campuses and multisite clusters.
Kairos Power's Hermes 2 is due to generate up to 50 MW from a single reactor from 2030. Hermes 2 sits between the first two classes, the Xe-100 near the second, and the BWRX-300 in the third. Rolls-Royce SMR is developing a 470 MWe design, between the mid-size class and the large one. As with data centre figures, a figure without its class cannot be placed.
With a turbine each, four 80 MWe units can carry 200 MW of hyperscale IT capacity with one out

A campus fed directly by its own reactor, behind the meter, must cover that reactor's outages. The light water BWRX-300 is designed for a 10 to 15 day refuelling outage every 12 to 24 months.
Take the campus with 200 MW of hyperscale IT capacity and a 220 to 240 MW connection. One 300 MWe unit carries it while running, and nothing while stopped.
A second unit covers the outage, but such a spare could be oversized, the IAEA notes. Four 80 MWe reactors make 320 MWe, and any three make 240 MWe, the full connection.
That holds only if the turbines are spared as well as the reactors. China's HTR-PM, a high-temperature gas-cooled plant at Shidao Bay, shows why. It entered commercial operation in December 2023 with two small reactors driving one turbine.
Long Mott would pair four Xe-100 units with Dow's chemical site at Seadrift. World Nuclear News calls the permit the NRC's first for a design that refuels online, without stopping.
In Johannesburg, South Africa, Cassava Technologies has space for a 20 MW AI factory installation. If that is single-tenant hyperscale IT load, it needs 22 to 24 MW. That is about two dozen 1 MWe microreactors before a spare. Equinix has pre-ordered 20 of Radiant's Kaleidos microreactors.
If a reactor fed your campus directly, what would carry the load while it was down?
Cascade and Hermes 2 will deliver through the grid, which carries the spare

Most current deals, especially in the United States, deliver through the grid, the IAEA observes. The grid's spread of demand then absorbs one unit's outage.
Energy Northwest plans the Cascade plant outside Richland, Washington, with Amazon funding. Its first phase is four Xe-100 units and 320 MWe. Hermes 2's output will go to the Tennessee Valley Authority grid that powers Google's data centres in two states.
For data centres in the Netherlands, Equinix has a letter of intent with ULC-Energy for up to 250 MWe. ULC-Energy has chosen Rolls-Royce SMR, so that is about half of one unit's output.
In the UK, Poland's SGE, formerly Synthos Green Energy, leads a team planning 14 BWRX-300 units at three sites. Some data centres are planned alongside them.
Smaller units can follow a campus's growth, but today arrive after its first phases. Radiant targets its first commercial units for 2028. A campus built now runs its first phases on the grid or other generation.
The NRC lets applicants scale the planning zone, and KHNP wants rules for private supply

An NRC rule published in November 2023 lets SMRs and other new designs use a scalable emergency planning zone. For a campus beside its reactor, that zone becomes part of the siting case.
In the IAEA report, Korea Hydro and Nuclear Power (KHNP) compares grid, behind-the-meter and hybrid supply for its Innovative SMR. Behind the meter could be faster, it writes. But market rules must change before nuclear power can reach private users.
Through the grid, the reactor can sit where nuclear sites and transmission exist. The campus can sit where land and fibre are.
Four questions test any pairing before a site is chosen

The utility's engineers and the campus's power team answer them together.
What does each figure measure: net MWe, nameplate, IT capacity or peak?
What carries the campus while a unit is down?
Does the reactor sit on the campus, or deliver through the grid to several sites?
Will the units arrive in time for the campus phases they serve?
Our Atlas shows nuclear facilities and AI factories / data centres on one globe.
The takeaway

A campus with its own reactor must cover that reactor's outages, with several smaller units or a grid backup. When a reactor feeds the grid, the grid carries the spare, and most deals so far take that route. Either way, convert the megawatts first. A 300 MWe reactor is one unit in a fleet, rarely a campus's exact match.
Next week: Handover data now runs the AI factory, not only the paperwork.
Frontier models perform better in Attio’s agent harness.
We gave GPT-6 Astra the same job twice: analyze 100 sales calls and draft the follow-ups, once in Ask Attio and once in ChatGPT.
Ask Attio finished in 4 minutes 49 seconds. ChatGPT took 26 minutes 11 seconds, about 5.4 times as long.
The most ambitious GTM work requires more than frontier intelligence. It requires architecture that can put that intelligence to work at scale.



