Data center opposition is no longer background noise. It is becoming an execution risk for the capex cycle Wall Street has largely priced as inevitable.
Takeaways
-
The AI capex story now has a fourth bottleneck: chips, power and capital have been joined by permission to build.
-
Moratoriums are becoming a genuine timing risk: Wells Fargo counts 374 active local measures, with active pauses averaging close to a year.
-
This looks more like a repricing of execution risk than the end of the AI boom: good projects will probably still get built, but investors may need to stop treating every announced gigawatt as though it arrives on schedule.
AI’s New Bottleneck Is Permission to Build
Data center opposition is no longer background noise. It is becoming an execution risk for the capex cycle Wall Street has largely priced as inevitable.
For most of the AI boom, the bottleneck conversation has revolved around chips, power and capital. Nvidia (NASDAQ:) could not make GPUs fast enough, utilities could not connect generation quickly enough, and hyperscalers seemed willing to throw almost unlimited amounts of money at the problem.
A fourth bottleneck is now moving rapidly up the list: permission to build.
That is becoming harder to dismiss as a few angry town halls or another round of NIMBY politics. Morgan Stanley estimated that roughly $156 billion of data center projects were cancelled or delayed during 2025. In the first quarter of 2026 alone, another $130 billion ran into delays or cancellation. Data Centre Watch separately counted at least 75 disrupted projects that quarter, with opposition groups spanning 49 states.
Those numbers change the conversation. The AI buildout may still have mountains of capital behind it, but money cannot bulldoze its way through every county commission, electricity regulator and water authority in America.
And the political clock is making that problem louder.
With the midterms approaching, data centers have moved from a relatively obscure infrastructure debate into a broader argument over electricity bills, water, industrial development and who ultimately pays for the grid expansion needed to feed the AI machine. President Donald Trump has firmly backed the buildout on competitiveness, investment and jobs grounds, while Alexandria Ocasio-Cortez and Bernie Sanders have pushed for a nationwide pause on new construction. The important market point is not which political argument wins. It is that data centers have become visible enough to become campaign material.
For investors, that is when infrastructure risk begins to acquire a political volatility premium.
Wells Fargo’s latest tracker, according to research from analyst Shahriar Pourreza, takes the acceleration a step further. The bank now counts 374 active local moratoriums, up from just 92 reported in June. The objections cluster around the same pressure points again and again: electricity costs, grid capacity, water use and the impact of giant industrial developments on surrounding communities. Michigan, Ohio, North Carolina, Iowa and Tennessee are among the states showing particularly heavy activity.
This is the picture that matters.
The map turns what can sound like another Washington talking point into something much more tangible. This is not one state throwing sand into the gears. The resistance is scattered across large parts of the country, which means the industry cannot simply pick up a proposed campus and move it across the county line every time local opposition surfaces.
More importantly, these pauses have duration.
Wells Fargo says active moratoriums are averaging roughly 346 days, while the average duration across all tracked measures has risen to 311 days from 277. Some can stretch for years. In AI infrastructure, a year is not a rounding error. It can push power contracts, financing, equipment orders and anticipated revenue streams further down the runway.
That is where this starts moving from politics into the spreadsheet.
The market has spent much of the past two years valuing the AI infrastructure cycle as though the principal uncertainty was how high capital spending could climb. The next phase may be less about willingness to spend and more about how quickly that spending can be converted into operating assets.
That distinction matters enormously.
If a hyperscaler delays a project because a GPU shipment is six weeks late, much of the economic value is simply pushed slightly to the right. If the project loses local approval, requires a redesigned power solution, has to absorb more grid costs or gets moved to another state entirely, the economics themselves can change.
The financing market is already noticing. Reuters reported last month that lenders are paying much closer attention to permitting, community acceptance and regulatory readiness when underwriting data center projects. With AI-related infrastructure increasingly tapping both public and private credit markets, political permission is starting to become part of credit analysis rather than an afterthought.
This does not mean the AI infrastructure boom is about to hit a brick wall.
In fact, Wells Fargo’s own conclusion is considerably more nuanced. Pourreza expects some of the political noise to fade after November and argues that credible projects should remain manageable. His essential point is that the freewheeling stage of the buildout is ending, not the demand for additional computing capacity. Projects that protect ratepayers, offer genuine local benefits and arrive with believable power and water plans should still get built.
That sounds about right.
The Golden Goose is probably not being killed. But towns are starting to ask who is feeding it, who is paying the electricity bill and what happens to the water table after it moves in next door.
For markets, that makes the AI trade a little more complicated than simply drawing another straight line through hyperscaler capex.
The first leg of the boom was about compute scarcity. The next may increasingly be about scarcity of social and political permission.
And unlike GPUs, you cannot simply order more of that from Taiwan.
