NextFin News - The two largest data-center markets in the United States have slammed on the brakes within weeks of each other. New York Governor Kathy Hochul signed the nation's first statewide moratorium on hyperscale data centers on July 14, 2026, halting new projects while regulators study their impact on the power grid, water supplies and ratepayers. On August 3, Texas Governor Greg Abbott ordered a "comprehensive verification and audit" of every data center advancing through ERCOT's interconnection queue — roughly 474 gigawatts of pending requests, more than five times the grid operator's record peak demand of 91,089 megawatts set in July, with data centers representing about 90 percent of the total. The message from Albany to Austin is the same: the AI buildout is running into the one constraint no amount of chip innovation can fix overnight, the physical limits of the electric system.
The Situation: Two States, One Constraint
For three years, the AI investment thesis has been written in chip counts, model parameters and hyperscaler capital-expenditure guidance. The constraint that finally forced state governments to intervene is far less glamorous: electrons, water and local opposition.
New York's pause is the more formal of the two. Executive Order No. 62, signed July 14, stops new hyperscale data centers in their tracks until the state's Department of Public Service completes a Generic Environmental Impact Statement covering energy demand, water use and quality, air quality and noise. The order sets a 50-megawatt threshold and carries no grandfathering for projects still awaiting state environmental approval as of that date. Behind it sits legislation, A11560, now on the back burner, that would have gone further, requiring data centers with peak loads of at least 5 megawatts to source at least a third of their electricity from renewables by 2030, with rising targets after that, plus energy-efficiency standards set by the state's energy research authority.
"These hyperscale AI data centers consume enormous amounts of power, truly threatening to outpace our grid's capacity," Hochul said at the July 14 signing. "Progress shouldn't arrive with a higher utility bill, depleted water supply, or noise pollution. That is why today I'll be signing the nation's first-ever statewide moratorium on hyperscale data centers."
The numbers behind the decision show why the politics turned. The New York Independent System Operator's large-load queue swelled from six projects totaling about 1,045 megawatts in 2022 to 48 proposals totaling roughly 12 gigawatts by the end of last year, with more than two-thirds of that capacity landing in the queue during 2025 alone. That is not a gradual trend; it is a step change.
Texas took a different legal route but arrived at the same conclusion. Governor Abbott's August 3 directive to the Public Utility Commission of Texas and ERCOT does not create a statutory moratorium; it freezes the queue pending an audit with no stated deadline. Any project that fails the verification process "must be denied" a grid connection. The order immediately derailed ERCOT's newly approved "Batch Zero" interconnection study process, which the commission had signed off on June 18 and which was supposed to start classifying projects in early August. ERCOT has instead sought a good-cause exception and postponed the study.
"ERCOT is reviewing Governor Abbott's letter concerning data centers and will work with the Public Utility Commission of Texas to implement the Governor's directive, including postponement of the Batch Zero transmission planning study," ERCOT spokesperson Trudi Webster said.
The scale of what is being paused is difficult to overstate. ERCOT was already tracking approximately 410 gigawatts of large-load interconnection requests as of late March 2026, about 87 percent of them data centers. For context, ERCOT's record peak demand sits at 91,089 megawatts. The queue is not a pipeline of near-term construction; it is a wish list that assumes generation, transmission and transformers will materialize on command.
The Analysis: Why the Grid Is the Bottleneck
The constraint is physical, not financial
The first thing to understand is that these pauses are not really about data centers. They are about the lead time of the electric system. A gas-fired peaker plant, a new transmission line or a large power transformer takes years to permit, manufacture and commission. A hyperscale data center can be financed and designed in months. That asymmetry has been papered over during the low-demand era, when the grid had spare capacity and efficiency gains kept data-center power draw flat even as workloads nearly tripled between 2015 and 2019.
That era is over. The International Energy Agency now projects global data-center electricity consumption could reach 945 terawatt-hours by 2030 and climb to around 1,200 TWh by 2035. Deloitte's forecast runs similar, at 1,065 TWh by 2030. Goldman Sachs Research expects data-center power demand to grow roughly 160 percent by 2030 compared with 2023, with AI alone adding about 200 TWh a year over that span and accounting for around 19 percent of total data-center demand by 2028. The United States, which accounted for 45 percent of global data-center electricity consumption in 2024, is expected by the IEA to see its data-center energy demand rise 130 percent by 2030.
This is the structural heart of the story: the bottleneck is in the grid's build rate, and build rates are governed by permitting, supply chains and capital discipline, none of which respond to software-cycle speed.
Cyclical speculation meets structural scarcity
It is important to separate two forces that are being conflated in the political debate. The long-term demand for compute is structural. AI training and inference loads are not a bubble that deflates when sentiment turns; they are embedded in the operating models of the largest companies in the world. Deloitte estimates US AI data-center power demand could grow more than thirtyfold by 2035, reaching 123 gigawatts from about 4 gigawatts in 2024.
But the interconnection queue is cyclical, and in parts speculative. When 90 percent of every new power request in ERCOT comes from one industry, and when that queue is five times peak system demand, a large share of those projects are options, not commitments. Developers queue cheaply, shop sites across multiple states, and abandon the rest. The audit that Texas ordered is, in effect, a forced purge of that speculation.
"I will be pleasantly surprised if in April of 2027, the industry has a clear picture of which data center loads are going to be able to move forward," said Beth Garza, who previously served as the independent market monitor for ERCOT.
That timeline, an audit with no deadline pushing clarity well into 2027, is itself the point. Regulators are no longer willing to let the queue sort itself out through private optionality while ratepayers underwrite the grid upgrades that winning projects will need.
The second-order effect: bring your own power becomes the price of entry
The immediate consequence everyone can see is delay. The second-order consequence is a change in who gets to build. New York's order explicitly directs regulators to consider making data centers make upfront capital contributions to finance grid improvements, join demand-response programs, support new clean-energy procurement and fund an insurance pool, a proposed "New York Grid Acceleration Fund." At the federal level, the White House has promoted a "Ratepayer Protection Pledge" under which hyperscale AI companies and their utility and developer partners "will build, bring, or buy every kilowatt their facilities need and cover every dollar of the infrastructure that delivers it."
"It's only fair that the cost of building the new infrastructure required to meet this demand should be borne by the corporation themselves and not by the American consumers," President Donald Trump said in late July.
That framing, whether it becomes binding policy or remains rhetoric, redraws the competitive map. It advantages the hyperscalers with balance sheets large enough to contract directly for generation, including dedicated gas turbines, battery storage and small modular reactors. It advantages the equipment vendors that sell into that buildout: gas-turbine manufacturers, transformer makers, electrical-component suppliers and the engineering firms that can deliver turnkey power. It disadvantages smaller developers who relied on the grid to deliver reliable, subsidized power to their door, and it disadvantages the ratepayer-subsidy model that underpinned the first wave of the boom.
The geography shifts too. If New York and Texas make queue access harder, capital flows to jurisdictions that have kept the door open, and to sites where developers can build generation behind the meter. That is not a slowdown in the buildout; it is a rerouting, and it concentrates activity in places with weaker grids and thinner regulatory capacity, which is its own risk.
The counter-thesis: this is political theater, not a real brake
The strongest argument against reading too much into these pauses is that they do not actually stop demand. Data-center demand is inelastic: the models are being trained, the inference loads are growing, and the hyperscalers have committed to hundreds of billions of dollars in capital expenditure that they cannot simply walk away from. Projects will route around the moratoriums. New York's pause does not cover facilities that proceed solely through local permitting without state environmental approval, and it expires once the environmental review is complete. Texas's audit has no deadline, which means it can be completed quickly once political pressure eases.
There is also the efficiency card. Between 2015 and 2019, data-center workloads nearly tripled while power demand stayed flat, because efficiency improvements absorbed the growth. If chip efficiency accelerates again, and if AI inference becomes cheap enough per query, the demand curve could bend down faster than the current forecasts assume. On that view, the moratoriums are a speed bump that will look excessive in hindsight, and the equipment vendors and power developers pricing in a decade of scarcity are the ones getting ahead of themselves.
The problem with that argument is timing and magnitude. Efficiency gains have slowed since 2020 even as model size and query volume have grown; the IEA's doubling forecast is made in spite of expected efficiency improvements, not because forecasters forgot about them. And even if demand growth moderates, the queue that exists today is real: 474 gigawatts in ERCOT, 12 gigawatts in New York, hundreds of gigawatts more across other grid operators. The moratoriums are a response to commitments and requests that are already on the books, not to a hypothetical demand surge.
The falsifying signal is specific. If ERCOT's verified data-center load falls materially, say a verified drop of more than 30 percent in committed large-load requests after the audit, and if the major hyperscalers cut capital-expenditure guidance for AI infrastructure by double-digit percentages in consecutive quarters, then the scarcity thesis is overstated and the moratoriums are indeed political theater. Until then, the burden of proof has shifted to the developers.
Outlook: A Regime Change, Not a Reversal
The moratoriums are best read as a regime change in how data-center growth gets permitted, not as a rejection of the growth itself. That distinction matters for who wins and who loses across time horizons.
In the short term, the queue gets purged. Projects that were speculative, under-capitalized or dependent on ratepayer-funded grid upgrades will stall or relocate. Expect delays to stretch into 2027 in Texas, given the audit's open-ended timeline, and into late 2026 or 2027 in New York as the environmental review runs its course. Data-center developers with assets concentrated in paused jurisdictions face execution risk; hyperscalers with diversified site portfolios can absorb the friction.
Over the medium term, the cost of power becomes a first-order competitive variable for AI, not an overhead line item. Companies that locked in cheap, firm power early, through long-term renewable contracts, gas-fired dedicated generation or nuclear offtake, gain a structural cost advantage over rivals that assumed the grid would absorb them at regulated rates. The equipment supply chain, turbines, transformers, switchgear and cooling systems, remains the binding constraint and the clearest beneficiary.
Over the long term, the structural demand for compute is unlikely to be legislated away. What changes is the route: more behind-the-meter generation, more direct procurement, more buildout in states that welcome it, and a higher hurdle for any project that expects the public grid to carry its cost. The states that figure out how to say "yes, if you pay for it" will capture the investment; the states that say "not yet" will delay it, not cancel it.
The reality these moratoriums deliver is not that the AI boom is ending. It is that electricity has finally become the scarce input, and scarcity always gets priced, one way or another.
Explore more exclusive insights at nextfin.ai.
