Grid bottlenecks put half of US AI data center plans at risk, Berkeley Lab data show
More than 70% of US grid interconnection requests are withdrawn, threatening the $5.2 trillion AI infrastructure buildout before it completes.
More than 70% of grid interconnection requests in the United States are ultimately withdrawn because the grid cannot accommodate them, according to Berkeley Lab data — a constraint that one analyst believes will prevent half of all currently planned US data centers from ever being built.5
Industry forecasts place AI data center capital expenditure at roughly $5.2 trillion between now and 2030, with Goldman Sachs Research projecting global data center power demand will surge up to 165% over 2023 levels by that date.5 Those figures describe an investment programme that depends on a grid that is already rejecting the majority of new connection requests.
The infrastructure gap is physical, not regulatory. Energy companies and public utilities must invest not only in replacing or retrofitting aging transmission lines and distribution infrastructure, but also in building new generation, storage, and transmission projects.3 Upgrading the grid takes years. AI data center timelines are measured in months.
That asymmetry shapes where projects survive. Those with the strongest sponsors, best locations, and clearest utility partnerships will likely move forward. Others face delays, cost overruns, or extended interconnection timelines, and the same pressure extends to the broader power generation buildout feeding these facilities.4
The spending imbalance between the US and China makes the grid constraint more acute for American operators. Last year China's tech firms were estimated to have spent $24bn on AI infrastructure including data centers; American ones spent over $350bn.2 That 15-to-1 gap concentrates grid strain in the states absorbing the bulk of US deployment.
China sidesteps some of those bottlenecks through price. Industrial power in China costs around half the rate many American businesses pay, according to official figures, and because the government sets residential power prices separately, there is little public opposition to power-hungry infrastructure.2 Ken Liu, an analyst at UBS, expects China to build another 25GW of AI data centers by 2029, having built just 5GW over the past two years.2
Yet China's chip position limits what those data centers can do. Less than 0.1% of its chips are capable of the intense calculations needed to train AI models, according to a manager at one facility, and some data centers reportedly run at utilisation rates as low as 20%.2 Cheap electrons and willing planners do not substitute for compute density.
NVIDIA's trajectory captures the demand side of this story. Two and a half years ago the company was a $300 billion gaming chip manufacturer; by late June 2026 it had crossed $4 trillion in market value, the most valuable company in history at that point.5 ChatGPT launched in November 2022 and reached 100 million users in two months — the pace that set off the infrastructure race now running into physical grid limits.5
The market is already pricing uncertainty into grid-adjacent equities. Fluence Energy, a grid-scale battery storage company, traded across a 52-week range from $4.40 to $33.51, with a five-year monthly beta of 2.62 — volatility that reflects deep disagreement about which projects actually get built and on what timeline.1
If 50% of planned US data centers are never completed, the demand surge Goldman projects would be materially smaller, and the power, gas and carbon curves that have partly priced that buildout would need to adjust.5 The interconnection data suggests this is not a tail scenario. For energy traders, the forward indicator is how quickly US utilities and transmission operators can accelerate queue processing — and whether any of the projects now stalled begin to formally withdraw rather than simply age in the pipeline.4