How to read a capacity projection

Every projection of nuclear capacity rests on an anchor, and the anchor decides the way it will be wrong. At one end is concrete already poured: take the fleet operating today, add the units under construction at their expected grid-connection year, and subtract the ones retiring. At the other end is a policy target — a five-year plan, a government goal — spread over time. In between sit announced and planned units, which already have a name, a site and an estimated date but no construction yet. All three show up in headlines without saying which is which, and they go wrong so differently that the first question to ask of any forward number is what it is anchored to.

A projection anchored to construction fails on timing, almost never on existence: the capacity shows up, later than the bar said. In the atlas dataset, across the 111 units connected between 2010 and 2026 construction ran a median of 7.2 years from structural concrete to commercial operation — but the mean is 9.5 and the ninth decile reaches 13.8. That gap between median and mean is the whole story: the distribution has a long tail that drags the average. Olkiluoto 3 took 17.7 years, Flamanville 3 took 18.4, Mochovce 3 took 35.9 with the site idle for nearly two decades, and Watts Bar 2 took 43.1 years between 1973 and 2016. None of those projects evaporated. They all finished, late.

The tail is not evenly spread either, which breaks any projection that applies one world average. Over that same 2010-2026 window, the 49 units connected in China have a median of 5.7 years and a ninth decile of 8.9; the 62 connected everywhere else have a median of 8.8, a mean of 12.3 and a ninth decile of 18.4. South Korea and the UAE look more like China than like Europe, for the same reason: repeated series of one design, built by crews that have built it before. India sits at the other end, with Kakrapar 3 and 4 at 12.6 and 13.4 years and Rajasthan 7 at 13.8.

A projection anchored to a target fails another way: it does not arrive late, it arrives small, sometimes by multiples. In 2004 India set a goal of 20 GWe by 2020; actual capacity in 2020-21 was 6,780 MW by the Indian government's own count, close to three times less, and parliament had already been given revised figures of 14,600 MWe for 2020-21 back in December 2011. China is the interesting case, because it is usually cited as the one that delivers: its 12th Five-Year Plan set 58 GW operating by 2020 and delivered around 50, and it has missed its nuclear target in each of the last three planning cycles even as it overshoots its renewable ones. Building fast is not the same as hitting your own targets.

One last source of confusion is not about method but about basis, and it explains much of the arguing over a single country's numbers: gross capacity at the generator terminals runs 6 to 8 percent above the net capacity that reaches the grid. A source reporting 66 GW gross and another reporting 61 GW net for the same fleet, on the same date, agree with each other.

Reactor Atlas publishes net capacity, as recorded by the IAEA's PRIS, and its forward view is a hybrid that is worth reading as one. Of the 74.2 GWe it adds to the 2031 frame, 54.7 come from units under construction and 19.5 from planned units, 26 percent that has no construction behind it yet. For the construction half, what is said above holds: the likely error is the bar sliding right. For the planned half it does not, because an announced project that never gets built does not arrive late, it disappears. Anyone who trusts only poured concrete should read the 2031 total about 19 GWe below what the atlas shows.