Building up the US power grid won't be wasted, even if the AI bubble bursts

The Register's piece turns on a McKinsey argument that the asymmetric risk for US utilities is building too little, not too much, for the coming datacenter load. The consulting firm's arithmetic implies a 30 to 55 GW nationwide capacity gap by 2030 against roughly 120 GW of projected demand growth, after subtracting 50 to 75 GW of expected coal and gas retirements and adding the ~40 GW of dispatchable headroom and ~100 GW of committed new capacity the article cites. Only one feed is carrying the story and the underlying report is not named in the material provided, so the figures should be treated as single-sourced. The conceptual move McKinsey is making is that capital deployed for AI compute is not a one-way bet, because the same assets can serve other loads and harden the grid.

The cushion that makes overbuilding tolerable, in McKinsey's framing, is that datacenters are not the only source of new US electricity demand. The piece names electric vehicles and the electrification of industrial processes as concurrent pressures that would absorb capacity freed by a slower AI buildout, and notes that 71% of organizations report 'negative implementation outcomes' in their AI efforts, which the firm uses to justify caution about extrapolating current datacenter pipelines. For an engineer planning multi-year capacity, the relevant question is how substitutable that demand actually is across end uses, because a megawatt pulled by a hyperscale hall is not interchangeable with a megawatt pulled by residential EV charging in transmission planning terms. McKinsey's claim is that the demand stack is broad enough that the infrastructure will find users either way, which is a planning assumption worth scrutinizing rather than accepting by default.

The most operationally concrete shift described in the article is the move toward on-site generation, driven by grid-connection delays rather than by economics. McKinsey's survey of power sector leaders finds nearly 60% expect datacenters to keep permanent on-site generation by 2030 even after grid connections become available, with 64% of those relying on natural gas and fuel cells plus battery storage as adjacent options. The Register also reports that the Trump administration authorized up to $500 million to keep 13 coal-fired plants running, a separate but parallel signal that thermal capacity is being held on life support to bridge the gap. The practical effect for site selection is that 'where the grid is strong' may matter less than 'where a gas turbine can be permitted and fueled,' which changes emissions accounting, water use assumptions, and the calculus on whether a site can ever run off grid power alone.

Looking past the immediate shortfall, the article lays out McKinsey's expected technology mix: solar paired with battery storage and gas as the main on-site and grid-scale additions, nuclear limited to life extensions and upgrades of existing plants rather than new builds, and small modular reactors and next-generation geothermal unlikely to add material capacity before the mid-2030s. A Bain & Company report from roughly two years earlier warned of the same underbuilding risk, so the framing is not new, and the absence of independent feed corroboration for the McKinsey piece is worth flagging. The takeaway for planning horizons past five years is that the firm is treating today's thermal additions and solar-plus-storage as the realistic near-term mix, which is a more conservative posture than the SMR-and-geothermal narrative that has been common in industry coverage.