Bernstein told the Financial Times that roughly 35% of the announced US data centre pipeline should be treated as credible. Separately, the tracker RegLayer counted 47 or more large US projects delayed, withdrawn or blocked during 2026. Both figures landed on 18 October, the same week the Department of Energy confirmed $4.42 billion in loans to restart three nuclear plants largely to serve AI load.
The causes named are unglamorous and mostly physical: grid interconnection queues, permitting, equipment lead times - transformers and switchgear in particular - and local opposition. Several states now have active moratorium proposals. None of these is a demand problem. The compute demand is real; the question is whether the power, the paperwork and the parts arrive in the order the announcements assume.
For anyone reading capacity announcements as forecasts, the discount implied here is severe. If Bernstein is roughly right, an announced gigawatt is worth about 350 megawatts of expectation. That reframes the nuclear restarts from a bullish signal into something closer to a necessary one - firm power with an existing site and licence is attractive precisely because the greenfield pipeline keeps slipping.