Investment in US artificial intelligence data centers is still moving forward, but whether projects can connect to the grid on schedule—and how much of the proposed demand will actually materialize—has become a central issue for utilities, regulators and investors. Community opposition, shortages of construction workers, long delivery times for power equipment and uncertainty over demand for AI services have delayed or disrupted some projects. At the same time, grid planners must avoid overbuilding for demand that may not materialize while ensuring that insufficient preparation does not undermine reliability.
Energy consultancy Relae, formerly known as Carbon Direct, said in June that more than $170 billion of AI data center capacity had been blocked, withdrawn or put on hold because of community opposition since January 2024. That figure represents part of the roughly $581 billion Goldman Sachs expects major cloud providers to invest in US AI infrastructure this year, but the two estimates use different definitions and should not be treated as equivalent measures of lost investment.
Goldman Sachs estimated in May that US data center power demand would more than double from 2025 levels to 66 gigawatts by 2027. However, it expects only about half of the data center capacity scheduled to come online over the next one to two years to begin operations on time, after accounting for delays and cancellations. The Electric Power Research Institute estimates that data centers could account for 9% to 17% of US electricity demand by 2030, rising to as much as 20% by 2035.
Texas Reviews a 474-Gigawatt Interconnection Queue
The pressure on supply and demand planning is particularly acute in Texas. The Electric Reliability Council of Texas, or ERCOT, has paused new data center interconnection applications while it reviews a queue totaling 474 gigawatts. That is more than five times the state grid's record summer peak load of slightly over 90 gigawatts this year. Regulators and grid operators have stressed that applications in the queue do not necessarily represent load that will ultimately be built.
ERCOT President and Chief Executive Pablo Vegas said in April that current forecasts exceeded the organization’s expectations for future load growth. He also said uncertainty remained over how to identify and validate large loads and incorporate them into long-term planning. Thomas Gleeson, chairman of the Public Utility Commission of Texas, said in June that the forecasts appeared too high and that some applications were speculative.
The size of the queue is not the only source of pressure. Residential, commercial and industrial electricity demand is also rising in Texas, and some areas that once had spare capacity are now experiencing grid congestion. Gleeson said regulators wanted customers with signed interconnection agreements and projects ready to begin construction to receive the power they need to make reliable business decisions. At the same time, regulators need to establish that proposed projects are viable so that costs and reliability risks are not shifted to other customers.
Texas is using policy changes to screen large loads. Governor Greg Abbott signed Senate Bill 6 last summer, requiring large electricity users to demonstrate control over their project sites, cover a significant share of grid upgrade costs, disclose duplicate interconnection applications in other parts of the state and accept curtailment when the grid needs fixed loads to be reduced. On June 18, the PUCT approved a cluster-study framework for projects larger than 75 megawatts, intended to replace lengthy and duplicative individual studies. ERCOT said at the time that the final transmission plan for the first group, known as “Batch Zero,” was expected by the end of 2027.
Abbott later instructed the grid operator in August to review every data center in the queue first. ERCOT consequently delayed the initial cluster-study process and said it aimed to complete the review by December. BloombergNEF estimates that the Texas interconnection pause could delay about 20% of the US data center project pipeline. If the pause continues into next year, the share of affected projects could be higher.
Regulated Utilities Screen Projects Too
In markets where vertically integrated utilities oversee generation, transmission and distribution, distinguishing viable projects from speculative applications is also becoming more important. Brad Viator, president of the utility industry group Power for Tomorrow, said utilities were raising interconnection-study fees as applications increased. They are also adding provisions for upfront infrastructure payments and load-ramp schedules to power supply agreements, reducing the risk that other customers will be left paying for unused infrastructure if a project stalls.
Viator said utilities generally wait until a power supply contract has been signed or is close to being signed before including new generation capacity for a large customer in their load forecasts. In his view, the ability of vertically integrated utilities to coordinate generation, transmission and distribution can help secure supply after an interconnection is completed. He said interconnection queues in these markets typically take one to five years, compared with eight years or more in some deregulated markets within the PJM Interconnection region.
Karl Rábago, a former member of the Public Utility Commission of Texas and head of Rábago Energy, cautioned that the return structure for regulated utilities can encourage infrastructure investment, making it important to examine whether companies are adequately validating large-load forecasts. Utilities that own power plants and transmission and distribution assets also have substantial influence in regulatory proceedings, he said, meaning public forecasts should be assessed alongside community opposition and evidence of projects reaching completion. Grid Strategies said in November last year that aggregate data center load forecasts submitted by utilities to the Federal Energy Regulatory Commission were about 40% above the industry's credible forecast through 2030, in its assessment.
Rábago said cooperation across states and regulatory jurisdictions could improve forecasting accuracy, although he said he was not aware of such coordination currently taking place. He also noted that the industry still lacks a clear picture of how data center projects are distributed by size. Actual projects could be smaller than the thresholds used in large-load rate programs, he said.
Smaller Sites Add Complexity to AI Demand Forecasts
Some utilities are beginning to focus on smaller data centers, which are typically easier to connect to the grid and may face less community opposition. In late March, NorthWestern Energy submitted a large-load tariff proposal to the Montana Public Service Commission that would exclude loads below 50 megawatts from direct oversight. Chief Executive Brian Bird said the proposal was intended to help Montana attract high-quality economic development projects.
Anne Hedges, executive director of the Montana Environmental Information Center, argued in April that the approach could theoretically allow NorthWestern to connect multiple 49-megawatt data centers without subjecting them to the same level of oversight or providing comparable protections for utility customers. In Florida, Governor Ron DeSantis signed legislation in May requiring data centers larger than 50 megawatts to pay the full cost of connecting to the grid. The rule is consistent with a large-customer tariff previously approved for Florida Power & Light, the vertically integrated utility owned by NextEra.
Project forecasts also depend on actual demand for AI and cloud services. In May, xAI said it would rent servers already operating at its Memphis data center to Anthropic. Meta is planning to launch a cloud computing business similar to Microsoft Azure and Amazon Web Services. These developments suggest that some technology companies may have computing capacity beyond their near-term internal needs, although that does not mean other customers will necessarily be available to use it.
Bain & Company described the shift among large cloud providers in October last year as a move from competing for AI capacity to pursuing more selective, execution-focused growth. Advait Arun, a senior associate for capital markets at the Center for Public Enterprise, said forecasts based on speculative demand for “wildcat” data centers were losing support. Projects backed by upfront funding from major cloud providers are more likely to overcome construction obstacles, he said, while adding that this alone does not establish that evidence-based load forecasts are too high.
The need to expand the grid does not depend solely on the number of projects. In 2024, a transmission failure in Virginia caused 1.5 gigawatts of data center load to disconnect. The North American Electric Reliability Corporation has also recently warned that the sudden loss of large computing loads could create system-wide reliability risks. The final scale of data center construction remains uncertain, but Texas’s interconnection review, state rules on cost allocation and the grid’s ability to absorb changes in demand now directly affect who pays for future infrastructure and when that investment is made.