Why America is struggling to power the AI boom

An FT data-led explainer by Martha Muir uses the 102-mile Cardinal-Hickory Creek transmission line from Iowa to Wisconsin, which took 13 years and was completed in 2024, to show how fragmented US grid governance has become a bottleneck for AI. ITC chief executive Krista Tanner, asked whether the US can build fast enough: 'Increasingly, I worry that the answer is no.' Hyperscalers are expected to spend about $900bn on AI projects this year (JPMorgan), and Grid Strategies expects US demand to rise by 166 GW by 2030, equal to 15 New York Cities. Yet the US built nearly 4,000 miles of transmission in 2013 and under 900 miles in 2024; major projects take about a decade on average (IEA puts new lines at 5-15 years in Europe too), while China has built more lines in 15 years than the US in its history. Utilities, grid operators, states, FERC, environmental reviews and litigation all overlap: the Niskanen Center finds more than 25% of major projects under federal review in 2010-2020 were delayed or cancelled, though 16 of 18 lawsuits failed. S&P says utilities plan $1.3tn of generation and transmission investment by 2030, largely recovered through bills, and Pew finds half of Americans think data centres are bad for household energy costs. Responses include advanced conductors, sensors, demand flexibility (Google's March deals in Tennessee, Arkansas and Michigan to curtail load for faster grid access), behind-the-meter power (Cleanview counts 59 sites, over a quarter of planned US data centre capacity, such as AWS beside the Susquehanna nuclear plant) and a bipartisan Senate bill from September to speed federal reviews. Counterweights in the piece: Jigar Shah says flexibility plus modest upgrades could meet US needs for a decade while AEP's David Stader says it does not give that scale; RFF's Jesse Buchsbaum warns demand may be overestimated, leaving ratepayers with the bill; data centre operators argue they could lower rates; PJM's market monitor raises rationing or blackouts as the bleakest case. Chart values were not captured here.