Half of planned data centers will miss 2028, and the bottleneck is permits, not chips

Goldman Sachs expects only about half of new US AI compute capacity to be running on schedule by 2028, against a historical rate of roughly 72% of planned data-center capacity delivered on time, according to CNN's account of the bank's work ^[1][2]. That is a 22 point drop in delivery reliability ^[3], and it means any compute roadmap anchored to announced megawatts is overstated by construction, not by chip supply.

The pipeline arithmetic is worse than the schedule slippage implies. Stanford's AI Index Report counted 5,427 data centers operating in the United States at the end of last year, while AI companies announced 3,969 more, of which only 802 are actually under construction ^[4][5][6]. That is about 20% of the announced set in the ground ^[7], with the announcements alone representing a 73% addition to the existing stock ^[8]. Goldman's explanation is mundane: developers file in several jurisdictions at once and later pick the site with the best terms ^[9]. The same project can appear in multiple local pipelines and never net out of the headline total.

JPMorgan puts this year's AI infrastructure investment at $750 billion, but about 60% of the capacity meant to be live by the end of 2027 has not entered construction, and a further 7% of projects already started are behind schedule ^[10][11][12]. A typical build runs 18 to 24 months ^[13], so the un-started portion of that 2027 cohort is already out of calendar.

Chips have not stopped mattering: TSMC makes essentially all leading AI processors, including Nvidia Blackwell and AMD MI300X ^[14]. But the queue has moved. Goldman argues the main obstacle is permitting duration rather than outright bans, even though roughly ten states have weighed moratoriums and New York and Texas have imposed temporary restrictions ^[15][16]. Data centers consume about 8% of US electricity, which the American Edge Project projects rising to 12% by 2028, a 50% increase in share ^[17][18][19]. JPMorgan records a tripling of wait times for step-up transformers ^[20]; GE Vernova says generator orders have doubled in five years to $200 billion ^[21]; inflation in transformers and power regulators has been the second fastest of 47 categories tracked by the Bureau of Labor Statistics since 2020 ^[22]. The American Edge Project estimates the announced plans need an additional 500,000 electricians, 300,000 welders and 550,000 plumbers ^[23]. Gallup found 71% of Americans do not support these facilities being built ^[24].

Columbia Business School real estate professor Stijn Van Nieuwerburgh puts it in portfolio terms: of the 565 gigawatts AI companies currently plan, perhaps 180 gigawatts get built over the next decade ^[25], about 32% ^[26], with two thirds of the portfolio "implausible" in his words ^[27]. Even that reduced case implies roughly $10 trillion of investment, 50% more than the 19th century railroad buildout, the largest previous US investment boom ^[28][29]. Spread over ten years that is about $1 trillion a year, above the current $750 billion run rate ^[30]. The modest scenario is still an acceleration.

Watch the conversion rate, not the announcements: the 802 figure is the only number in this set that reflects committed capital and a poured foundation. Watch transformer and generator lead times, which set the physical floor on delivery dates regardless of financing. Watch component pricing bleed into consumer goods, where Reuters reports memory demand from AI data centers pushed Apple to raise prices on some MacBooks and iPads in July, with personal computers and peripherals up 3.5% in a month ^[31][32][33]. And note the Federal Reserve's position, per Reuters, that the direct effect of AI on US inflation and employment remains limited and contradictory ^[34], which is a reminder that the constraint story and the macro story are still running on different clocks.