Energy intelligence firm Currence published an assessment on 27 July projecting that 30% to 50% of large-scale data centre capacity due online in 2026 will slip — and, more usefully, showing its working on why.
The gap
Since 2024, roughly 190 GW of AI data centre capacity has been announced. Of that, 16 GW is scheduled to be operational during 2026. Currence finds visible construction progress on approximately 5 GW — meaning around 11 GW is scheduled to open this year with no construction the firm can see. The track record supports scepticism: in 2025, more than 25% of expected capacity missed its completion date, and a further 10% had its commercial operation date postponed.
The finding that connects to everything else
Projects with on-site generation account for under 10% of announced facilities but nearly half of announced capacity. The largest projects are the ones going off-grid, because the interconnection queue is where gigawatt-scale demand goes to wait. That is the same configuration the EPA removed a federal compliance step from on the same day — a coincidence of timing that reads less like coincidence the longer you look at it.
How much to trust the number
Two caveats belong on this. "Visible construction progress" is a proprietary methodology from a commercial vendor, not a count derived from building permits or interconnection filings — it cannot be independently reproduced. And the date of the underlying report is not stated in the write-up, so what is verifiably new is the publication rather than necessarily the research. The 190 GW figure, meanwhile, is exactly the sort of number that should never be treated as a pipeline: it is the sum of press releases.
Why announce capacity you cannot build
Because announcements are free and they move share prices, land options and utility negotiating positions. A developer who announces 2 GW and delivers 400 MW has still shaped a market. The discipline this report imposes is simple and rare — asking whether anyone has poured concrete.