The pause that got everyone's attention
Texas has paused new large load grid interconnections pending an audit of a queue that has swelled to 474 gigawatts, more than five times ERCOT's own summer peak demand of roughly 90 gigawatts. Senate Bill 6 now requires large load customers to prove site control, pay the upfront cost of grid upgrades their connection requires, and accept curtailment during tight periods. The Public Utility Commission of Texas adopted a cluster study framework for loads above 75 megawatts, but ERCOT delayed the first cluster, known as Batch Zero, with the underlying audit now targeted for completion by December.
BloombergNEF estimates the pause alone could delay 20 percent of the entire US data center pipeline, which is a remarkable figure for a single state's regulatory action to carry. ERCOT CEO Pablo Vegas has been direct about why the queue ballooned in the first place, saying the commission believes the forecast is higher than expected future load growth, a polite way of saying much of the 474 gigawatts was never going to be real demand.
Texas is not acting alone
Florida passed a law requiring data centers over 50 megawatts to cover the full cost of their grid connection, removing the option to socialize that cost across other ratepayers. Virginia governor Abigail Spanberger signed an executive order on data center accountability on September 18, a notable move given Virginia's status as the largest data center market in the country. Montana's NorthWestern Energy has proposed exempting loads under 50 megawatts from direct regulatory oversight, though critics warn this could let developers route around scrutiny by splitting large projects into multiple 49 megawatt sites.
The common thread across all four states is utilities and regulators building screening tools, large load tariffs with minimum contract lengths, collateral requirements, upfront impact study payments, and exit fees, specifically designed to separate serious projects from speculative ones before they consume engineering and planning resources that could otherwise go toward projects that will actually get built on schedule. Expect this pattern to keep spreading to any state with a large interconnection queue and a utility under pressure to explain it.
The number that should worry capacity planners more than the pause
Goldman Sachs expects roughly 581 billion dollars in US hyperscaler AI infrastructure spending this year and projects US data center power demand will reach 66 gigawatts in 2027, more than double 2025 levels. The firm's more consequential estimate is that only about half of capacity scheduled to come online in the next one to two years is likely to arrive on time. That is the figure that matters most to anyone building a capacity plan around a provider's public roadmap, since it implies a systematic, not occasional, gap between announced and delivered capacity.
EPRI data puts data centers at roughly 4 to 5 percent of US electricity demand today, projected to reach 9 to 17 percent by 2030 and as much as 20 percent by 2035 in the highest-demand states. Virginia alone could see data centers reach 39 to 57 percent of state electricity demand by 2030, which gives a sense of how concentrated the strain already is in the markets enterprises have historically defaulted to for colocation.
Why the forecasts themselves are part of the problem
Grid Strategies found that utility data center forecasts submitted to the Federal Energy Regulatory Commission ran roughly 40 percent higher than credible industry projections through 2030. That is not a rounding error, and it has a structural cause. Former Texas PUC commissioner Karl Rábago points to the return-on-equity incentive built into utility regulation, which rewards utilities for building infrastructure regardless of whether the load behind it ever materializes, and said plainly that he gets no sense utilities are excited about serving this load so much as building toward it.
PUCT chairman Thomas Gleeson was equally direct about it, calling a large share of the queue speculative in his own public comments. Advait Arun of the Center for Public Enterprise argues the most aggressive, speculative load forecasts are now dead in the water as a planning basis for utilities and regulators alike, though he is careful to note that baseline grid investment is still genuinely needed, just not at anywhere near the scale the inflated forecasts had implied to planners.
The cost of getting this wrong has already been paid once
Community opposition has blocked, withdrawn, or stalled more than 170 billion dollars in AI data center capacity since January 2024, according to Relae, formerly Carbon Direct. That figure predates and partly explains the current wave of regulatory screening. Utilities and state governments are not inventing friction out of nowhere. They are responding to a pattern where projects announced without adequate power and community groundwork ended up canceled anyway, after consuming months of planning and generating local backlash that made the next project in that market harder to site.
Brad Viator of Power for Tomorrow frames the shift from the utility side as wanting assurance they are dealing with a serious project rather than, in his words, a tire kicker. That is a rational response to the 170 billion dollar figure, and it means the screening friction enterprises are now encountering when evaluating new regions is a direct consequence of speculative activity that occurred well before any individual company got involved.
What this means for your next capacity decision
If a provider's roadmap for a specific region depends on grid capacity that has not yet cleared interconnection, treat the published timeline as the optimistic case, not the planning case your team actually builds around. Build contractual slack into any colocation or dedicated capacity agreement to cover delay, and ask providers directly whether their stated capacity has already cleared an interconnection queue or is still waiting behind an audit like the one currently underway in Texas.
The underlying constraint on AI infrastructure has shifted over the past year from chip supply, which hyperscalers have largely worked through by now, to grid interconnection, which is a slower-moving, more politically contested process that chip supply never had to navigate in the same way. Any enterprise still planning around 2024-era assumptions about how fast new capacity shows up is effectively planning against a bottleneck that no longer sits where they expect it to.



