The Siting Crisis: When the Grid and the Neighbors Both Say No
Data centers now draw 6% of US power, and the binding constraint on the AI buildout is shifting from chips to the politics of where the boxes can physically go.
The binding constraint has moved
For three years the story of the AI buildout has been told in the language of silicon: HBM allocations, CoWoS capacity at TSMC, Nvidia’s grip on the accelerator margin pool, the scramble for a million-GPU cluster. That story is still true, but it is no longer the binding constraint. The binding constraint is now where you put the building — and increasingly, whether the people who live near the proposed site, and the utility that would have to feed it, will let you.
Two items in this week’s feed look superficially unrelated. Microsoft has walked away from a 244-acre data center campus in Caledonia, Wisconsin after sustained community pushback. A 20-megawatt crypto operator’s proposal in La Pine, Oregon has residents in revolt. Behind them sits a single number that finally landed in the public conversation this month: data centers now consume roughly 6% of US electricity, with the backlash arriving at the same time. The thesis of this piece is simple: the AI capex cycle has reached the point where the marginal gigawatt is procured not from a grid operator but from a county commission, and the industry has not yet built the political machinery for that procurement to succeed at scale.
What the 6% number actually means
The 6% figure — drawn from the trajectory the Department of Energy’s Lawrence Berkeley lab projected at the end of 2024, when it warned data centers could reach 6.7% to 12% of US power demand by 2028 — is the kind of statistic that becomes a fulcrum precisely because it is legible. Six percent is, in round terms, more electricity than the entire state of Pennsylvania consumes. It is roughly double the share data centers held in 2018. And it is concentrated: ERCOT, PJM, and a handful of nodes in Virginia, Ohio, Arizona, and Oregon absorb the bulk of it.
The thing the headline number obscures is the shape of the demand. Conventional industrial loads grow with the underlying economy at a percent or two a year. AI training campuses arrive in 200- to 1,000-megawatt chunks, sometimes announced with eighteen months of lead time against transmission upgrades that take seven to ten years in PJM’s queue. FERC’s standard interconnection queue data — the basis here is the commission’s annual reports and PJM’s own published timelines — shows backlogs that have roughly tripled since 2020, with the vast majority of queued megawatts being some combination of solar, storage, and now, increasingly, behind-the-meter and co-located generation purpose-built for data centers.
That last category is where the 6% number breaks the old mental model. A growing fraction of the AI buildout is no longer asking the grid for power at all. It is asking for a gas turbine on-site, or a power purchase agreement with a specific nuclear plant — the kind of arrangement Amazon attempted at Susquehanna and FERC pushed back on in 2024. The industry’s response to interconnection paralysis has been to route around the grid, which works at the level of one project but creates a different political problem at the level of the county: the neighbors see a turbine, not a server.
Caledonia and La Pine are the same story
Caledonia is the more important data point because of who walked away. Microsoft had already committed billions across Wisconsin — Mount Pleasant alone is a multi-phase, multi-gigawatt anchor — and yet the 244-acre Caledonia parcel could not survive an organized local opposition that, by all reporting, focused on water use, noise from cooling infrastructure, traffic during construction, and the more diffuse sense that the community was being asked to host industrial load on behalf of users elsewhere. The hyperscaler with the most political capital in the region, the most experience navigating siting, and the cleanest balance sheet in the industry, decided the fight wasn’t worth it.
La Pine is the same dynamic in miniature. A 20-megawatt crypto facility is a rounding error against the 6% national figure, but the local newspaper coverage and the DCD report document a familiar pattern: residents object first to noise and water, then to the broader question of whether the load serves the community at all. The crypto framing makes the opposition louder and easier — bitcoin mining is a softer target than “AI” — but the underlying objections transfer cleanly to any high-density compute load.
What’s striking is how little the industry’s public messaging has adapted. The standard pitch — jobs, tax base, community investment — is calibrated for industrial loads that employ thousands. A hyperscale campus, once built, employs dozens. The construction boom is real and temporary; the permanent footprint is a building that hums, draws water, and sends electrons elsewhere. That asymmetry is becoming the central political fact of the buildout.
The capex math still assumes siting works
The four largest US hyperscalers — Microsoft, Alphabet, Amazon, Meta — have collectively guided to capital expenditure in the range of $320–360 billion for calendar 2026, with the AI-attributable share well over half. Those numbers are disclosed in their most recent 10-Qs and earnings calls and represent the highest absolute capex of any industry in American economic history, exceeding the peak years of the railroad buildout in real terms and rivaling the interstate highway program.
The financial models behind those commitments make an implicit assumption: that for every dollar of GPU spend, the matching dollar of shell, power, cooling, and fiber can be procured on a roughly parallel timeline. That assumption is what’s now in question. A GPU that sits in a warehouse for nine months waiting for a building is a depreciating asset against a four-to-six-year useful life — Nvidia’s accelerator generations now ship roughly every twelve to eighteen months, which means time-on-shelf erodes residual value faster than time-in-rack. Pat Gelsinger’s old line about “every fab year lost is a generation lost” applies in reverse to the customer side: every quarter a Blackwell or Rubin-class system waits for a building, its competitive value against the next generation declines.
This is the lens through which to read The Next Platform / Register Intelligence’s coverage of the hardware crunch: infrastructure teams report extended lead times and accelerated platform timelines arriving simultaneously. Both pressures bear on the siting problem. Faster platform refresh means the window in which a given chip generation pays back is shorter, which means the time-cost of a delayed building is higher, which means hyperscalers are willing to pay more for shovel-ready sites — which, in a market with finite shovel-ready sites, bids up land and power and triggers the next round of community opposition because the prices being paid become news.
And the fiber problem nobody priced
DCD’s editorial on “AI factories have a fiber problem” flags a second-order constraint that the siting debate has barely touched. Training clusters are increasingly distributed across multiple buildings on a campus, and inference fleets are distributed across regions. Both depend on dark fiber and metro/long-haul wavelengths that were sized for a pre-AI internet. The build cycle for new long-haul fiber is on the order of years; the build cycle for a metro ring around a new data center campus can be eighteen to thirty months on its own, often through the same right-of-way negotiations that bedevil power transmission. Fiber and power share a permitting problem because they share the trench.
The implication: even if a hyperscaler successfully sites a campus over local opposition, the connectivity needed to make it useful as part of a global training or inference fabric has its own siting and permitting overhead that doesn’t shrink with capital. You can’t pay your way past a county road department in less than its own clock speed.
The steelman: this has happened before, and it resolved
The strongest counter-argument is historical. Every major US industrial buildout has hit a wall of local opposition and grid constraint, and most resolved within a decade. The railroad land grants of the 1860s were politically toxic; they happened. Rural electrification ran into landowner objections in the 1930s; it happened. Cell tower siting fights in the late 1990s spawned federal preemption under the Telecommunications Act of 1996, which gave the FCC and the industry a framework to overrule the most obstructive localities. The historical pattern is: industry hits the wall, federal preemption follows, buildout resumes.
There are signs that mechanism is already loading. The Trump administration’s AI executive orders in early 2026 floated federal preemption of state and local data center permitting on national-security grounds; several state-level “AI economic zone” bills are working through legislatures in Texas, Virginia, Ohio, and Georgia that streamline approval and limit the role of local boards. The DOE has fast-tracked transmission corridors under the National Interest Electric Transmission Corridor authority it has held since 2005 but rarely used. The legal and political machinery for resolving the siting crisis exists in latent form.
A second steelman: the 6% figure may overstate the binding constraint because efficiency gains compound silently. Nvidia’s Blackwell-to-Rubin transition is projected — by Nvidia’s own disclosures at GTC — to deliver roughly 2-3x performance per watt on inference workloads. If model efficiency on the algorithmic side improves at the rate it has since GPT-4 (call it 3-4x per year on capability-per-FLOP for frontier reasoning tasks, per the Epoch AI and Stanford AI Index tracking), then the megawatts-per-unit-of-useful-AI-output is falling faster than headline demand is rising. The siting crisis may be a transitional phenomenon that resolves itself when the efficiency curve catches up with the demand curve.
Both arguments are real. Neither is dispositive. The cell tower precedent is closest, and it took six years from the start of organized opposition to the 1996 Act, during which deployment was substantially slowed in major metros. The efficiency argument requires the demand side to behave — and the demand side is currently expanding faster than supply at every layer of the stack, which is the definition of a bubble or a genuine technological shift, take your pick.
What to watch
Preemption fights at the state level. Watch Texas SB-level proposals to constrain ERCOT large-load interconnection rules, the ongoing Virginia legislative session debates over Loudoun County moratoria, and any federal NEPA carve-out for data center transmission. Each of these is a leading indicator of whether the political resolution comes in 2027 or 2030.
Behind-the-meter generation approvals at FERC. The post-Susquehanna order set the framework; what matters now is whether co-located gas, SMR-PPAs, and dedicated solar+storage configurations get standardized templates. If they do, the binding constraint moves from grid interconnection to land. If they don’t, hyperscalers will quietly relocate marginal capacity to friendlier jurisdictions — Alberta, the Nordics, the Gulf states — and the US share of frontier training capacity drifts downward.
Hyperscaler capex guidance for FY2027. The current $320–360B run-rate assumes siting works. If the Q3 and Q4 2026 calls show capex flat or declining while AI demand commentary stays bullish, that’s the tell that the industry is constraint-binding on something other than chips. Conversely, sharply rising capex with stable margins suggests the constraint is being bought through — which means costs are rising faster than disclosed.
Local ballot initiatives. Caledonia is the early signal. By the 2026 November ballot, expect a handful of county-level referenda explicitly on data center moratoria or zoning. Their results — particularly turnout — will tell you whether community opposition is the loud minority the industry assumes or the durable majority it fears.
Supermicro and the export-compliance overhang. Tangential to siting but related to capex predictability: Jensen Huang’s public pressure on Supermicro to fix compliance after the $2.5B smuggling bust signals that the GPU supply chain itself is becoming more politically constrained, which raises the cost of any given megawatt of compute and intensifies the cost of delayed siting.
The deeper point
The AI infrastructure debate has spent three years treating the GPU as the scarce resource. The GPU is the scarce resource that gets the headlines. The actually-scarce resources are now: a county commission willing to approve a 244-acre rezoning; a utility willing to fast-track a 500-MW substation; a stretch of dark fiber on a route that was sized in 2015. None of these are produced by Nvidia, none are addressed by anyone’s capex line, and all of them have political constituencies that the chip supply chain does not.
What Caledonia means is that the most capable, best-resourced, most-experienced actor in the industry decided a fight wasn’t winnable. That is not how industries behave when their binding constraint is upstream. It is how they behave when the binding constraint has migrated, and the playbook hasn’t caught up. The next twelve months of US AI buildout will be decided in zoning meetings, FERC dockets, and state legislatures more than in any fab in Taiwan or any GPU launch in San Jose. The 6% number was the warning shot. The shots that matter come next.