Introduction: When a Data Center Becomes an Election Issue — The New Political Risk Around AI Infrastructure
On July 23, 2026, something happened in Henderson County, Texas, that would have been almost unthinkable two years earlier. A major data-center developer, Diode Ventures, formally withdrew its proposal to build a large, high-density AI facility near Cedar Creek Lake — not because it ran out of capital, not because it could not find a tenant, and not because the chips it planned to install had become obsolete. It withdrew because the project had failed a political test. Governor Greg Abbott announced that the development did not meet the state’s expectations for protecting local communities, natural resources, and the electric grid, and he framed the withdrawal as the enforcement of a new standard for anyone who wants to build AI infrastructure in Texas.[1]
“Data centers that want to do business in Texas must meet a clear standard.” — Governor Greg Abbott [1]
The story behind the withdrawal is even more instructive than the headline. Residents living along the western shore of Cedar Creek Lake — a reservoir that supplies drinking water to the Fort Worth area — organized a volunteer coalition called Save Cedar Creek Lake, gathered thousands of petition signatures, and raised alarms about reported water draws of as much as five million gallons per day, constant low-frequency noise from industrial cooling fans, potential on-site gas generation, and a development process they described as opaque.[2][3] In its notice to Henderson County, the developer conceded that the site’s close proximity to a rural residential community did not meet the higher standard that should guide data-center development, and it pledged to pursue closed-loop and air-cooled systems and earlier community engagement at future sites.[4] The local county commissioner who fought the project put the matter plainly:
“…was not an appropriate place for a project of this scale.” — Commissioner Wendy Spivey, Henderson County [5]
Eleven days later, the political ground shifted again. On August 3, 2026, Governor Abbott directed the Public Utility Commission of Texas and ERCOT to conduct a comprehensive verification and audit of every data-center project moving through the state’s interconnection process — effectively pausing new grid-connection approvals in the largest data-center growth market in America.[6] The scale of what triggered that pause is staggering: ERCOT’s interconnection queue had swelled to roughly 474 gigawatts of requests, more than five times the grid’s record peak demand, with data centers representing approximately 90 percent of the new load.[7] By August 14, 2026, state officials confirmed that the audit would cover roughly 250 to 300 projects representing about 200 gigawatts of future demand — more than twice ERCOT’s all-time peak — and would take “several months” to complete.[8] ERCOT suspended its “Batch Zero” large-load classification notifications, and lawyers immediately began advising developers, lenders, and investors to re-read the change-in-law, force majeure, and termination provisions in their project documents.[9]
Read that last sentence again, because it is the thesis of this paper in miniature. A governor’s letter — a political document — instantly forced billions of dollars of project finance to be re-examined clause by clause. Politics moved, and capital had to reprice.
At the same time, this dynamic went national. Reuters reported in August 2026 that banks and asset managers financing the U.S. data-center boom are now explicitly underwriting political and community opposition as a risk category, alongside the traditional technical, environmental, zoning, and financial reviews.[10] The numbers explain why. In the first quarter of 2026 alone, at least 75 projects worth roughly $130 billion faced local opposition, according to research firm Data Center Watch; a further 20 proposals worth roughly $98 billion across 11 states were blocked or delayed between April and June.[10][11] Against this, Goldman Sachs forecasts that big technology companies will spend more than $6 trillion on AI through 2030 — many times the capital deployed on internet infrastructure during the dotcom era.[10] The head of infrastructure finance at one of America’s largest banks summarized the new underwriting doctrine in a single line:
“…the second aspect I look for is the credit quality of the project.” — Karen Fang, Bank of America [10]
“Readiness” is the operative word. It now includes permits, approvals, and — remarkably — the support of the people who will live around the facility.
As the United States approaches the November 2026 midterm elections, electricity bills, water consumption, land use, noise, tax incentives, and infrastructure cost allocation have become live campaign issues for governors, legislators, county commissioners, and challengers of both parties. A Reuters/Ipsos poll in June 2026 found that only 33 percent of Americans agreed that building data centers at a rapid pace is mainly a good thing, while 77 percent feared higher electricity bills; respondents named data centers among their top midterm issues.[12] Gallup polling found that roughly 70 percent of Americans would object to an AI data center being constructed in their own community.[13] More than 300 state-level data-center bills were filed in the first six weeks of 2026, statewide moratorium proposals were introduced in 14 states from both sides of the aisle, and Maine came within a single House vote of enacting the first statewide data-center ban in U.S. history.[14] In Washington, Congress introduced the Ratepayer Protection Act in June 2026 to require large-load customers such as hyperscalers to bear the full costs of grid upgrades built to serve them.[15]
Something structural has changed. AI infrastructure — the physical substrate of the most capital-intensive technology buildout in history — has acquired a political price. This paper proposes a framework for understanding, measuring, and ultimately compressing that price.
Why I Call This Framework “Siting Beta”
In finance, beta measures sensitivity to risk — the degree to which an asset’s returns move with a source of systematic variation that cannot simply be diversified away by picking better securities. I use Siting Beta as a new AI-infrastructure concept describing how strongly a project’s value, timetable, financing cost, and probability of completion are affected by where it is built. The title fits this paper for six specific reasons, and it is worth stating them explicitly at the outset because they define everything that follows.
First, beta is the right metaphor because siting risk is systematic, not idiosyncratic. A single bad contractor is an idiosyncratic risk; a state where the governor pauses all interconnection approvals, where 14 states debate moratoria, and where 70 percent of voters oppose a facility near their homes is a systematic risk factor that loads onto every project in that jurisdiction simultaneously. Like market beta, it cannot be engineered away at the individual-asset level; it must be priced.
Second, beta implies sensitivity, and sensitivity is exactly what differs across locations. Two identical 1-gigawatt AI campuses — same GPUs, same tenant, same design — can have radically different values because one sits in a predictable jurisdiction with secured water, durable permits, and community consent, while the other faces political opposition, drought exposure, permitting disputes, contested elections, grid constraints, or tax rules that change after ground is broken. The facility is the same; the exposure is not. Beta is the standard financial language for exactly this kind of differential exposure.
Third, beta is measurable, and siting risk is becoming measurable. Data Center Watch now tracks blocked and delayed projects quarterly in dollar terms; pollsters track community sentiment; lenders track permitting status; S&P Global Market Intelligence estimates lenders committed $121.91 billion in credit for U.S. data-center properties in 2025 alone, with average reported financings of $1.2 billion — precisely the kind of concentrated exposure for which risk factors get formally quantified.[16] A concept only earns a financial name when it can be observed, and siting risk can now be observed.
Fourth, beta connects the local to the capital markets, which is the central transmission mechanism this paper documents. A zoning vote in a township of 2,400 people can move the pricing of a $12.3 billion bond. The word “beta” signals that we are not writing about land-use planning as such; we are writing about how land-use planning propagates into discount rates, spreads, contingencies, insurance, and valuation.
Fifth, beta implies that risk can be high or low — and therefore managed. A high-beta stock is not a bad stock; it is a stock whose risk must be compensated. Likewise, a high-Siting-Beta location is not unbuildable; it is a location whose political and physical uncertainty must either be compensated with return or compressed with community benefits, closed-loop cooling, self-supplied power, and durable political agreements. The framework is therefore constructive, not merely diagnostic.
Sixth, “Siting” restores the physical and political dimension that pure financial language misses. AI discourse is full of abstractions — parameters, tokens, FLOPs. But the binding constraints of 2026 are transformers, substations, aquifers, decibels, county boards, and ballots. Pairing the most physical word in infrastructure (“siting”) with the most financial word in asset pricing (“beta”) captures the core claim of this paper: geography has become a factor in the cost of capital for artificial intelligence.
Core Thesis
Siting Beta converts local political and infrastructure uncertainty into a capital-market variable. A simplified decomposition is:
Siting Beta = Power Risk + Permit Risk + Water Risk + Externality Risk + Community Risk + Political Risk
The remainder of this paper develops each term, shows how the composite enters the cost of capital, maps the emerging Siting Beta geography of the United States, examines how hyperscalers are learning to compress it, documents Wall Street’s discovery of it, traces its propagation through the Five-Layer AI Economy, and distills the lessons into seven pillars.
Section 1: When Location Becomes Financial Risk
1.1 From Site Selection to Capital Allocation
For most of the industry’s history, choosing a data-center site was a technical exercise delegated to real-estate teams: find cheap land, cheap power, fiber routes, low disaster risk, and a friendly tax abatement, then build. The facilities themselves were modest — tens of megawatts, a few hundred million dollars — and communities barely noticed them. That world is gone. The AI campus of 2026 is a multibillion-dollar industrial asset: the Stargate facility rising in Saline Township, Michigan, carries a $16 billion price tag and a 1.4-gigawatt power draw; the Homer City Energy Campus in Pennsylvania is billed as the largest capital investment in that state’s history; Amazon’s New Carlisle campus in Indiana is an $11 billion commitment; Meta’s Lebanon, Indiana campus is $10 billion.[17][18][19] Individual projects now rival the market capitalization of mid-sized public companies and consume the electricity of mid-sized cities.
When a single facility reaches this scale, site selection stops being a real-estate decision and becomes a capital-allocation decision — one of the largest a company will ever make. The hyperscalers themselves have made this explicit. Microsoft, Alphabet, Meta, and Amazon are expected to spend roughly $760 billion in combined capital expenditure in 2026, up from about $413 billion in 2025, with essentially all of the increase directed at AI data centers, chips, power, and networking.[20] Goldman Sachs now models roughly $5.3 trillion of combined capex for these four companies from fiscal 2025 through fiscal 2030.[21] Capital intensity at this level has a consequence that the industry is only beginning to internalize: when nearly all of your operating cash flow is being converted into immovable physical assets, the risk characteristics of the locations of those assets become the risk characteristics of the company.
1.2 Why Identical AI Factories Are Not Financially Identical
Consider two hypothetical 1-gigawatt AI campuses, identical in every engineering respect: the same Nvidia accelerators, the same hyperscale tenant on the same 15-year lease, the same closed-loop cooling design, the same construction contractor.
Campus A sits in a jurisdiction with a completed interconnection agreement, a signed water contract with a utility that has surplus reclaimed capacity, a community-benefits agreement negotiated before groundbreaking, a county board that approved the rezoning 7-0 after public hearings, and a state whose data-center rules were codified in statute three years ago and have not changed since.
Campus B sits in a jurisdiction where the interconnection queue has just been paused pending a state audit; where the water utility faces drought restrictions; where a residents’ coalition has gathered 4,000 signatures and filed a zoning appeal; where the governor faces a competitive re-election in November and has made data-center discipline a campaign theme; and where the legislature is debating whether to repeal the sales-tax exemption that underpinned the original pro forma.
The engineering net present value of these two campuses is identical. Their financial values are not, because the probability-weighted timeline of cash flows differs, the variance of that timeline differs, the probability of outright cancellation differs, and therefore the discount rate that a rational lender or equity investor applies differs. Campus B must either offer a higher expected return or absorb a lower valuation. This differential — location-driven, politically mediated, and increasingly quantifiable — is Siting Beta made visible.
1.3 The Emergence of the Siting Spread
I define the Siting Spread as the additional financing premium — in basis points of debt cost, in required equity return, in contingency reserves, or in guarantee requirements — associated with a higher-risk location relative to an otherwise identical project in a low-risk location. The Siting Spread is to AI infrastructure what the country-risk premium is to emerging-market sovereign debt: a systematic add-on that compensates capital for jurisdiction-specific uncertainty.
Reuters’ August 2026 reporting shows the spread being born in real time. Senior bankers described leaning toward projects in states that are “more welcoming” toward data centers; banks begin funding conversations at least a year before construction, so contested projects consume additional due-diligence cycles that themselves carry cost; and the largest financings — such as the $12.3 billion bond sale managed by JPMorgan and Morgan Stanley for BlackRock’s partnership with Meta in El Paso — now close against a backdrop in which other projects, such as Virginia’s Prince William Digital Gateway, are terminated outright after community pushback.[10][11] The market has not yet published a clean, quoted spread between high-beta and low-beta jurisdictions — Reuters is careful to note the shift is directional rather than a bright line — but the underwriting architecture that will eventually produce such a spread is now in place.[11]
A useful way to think about the components of the spread:
| Channel | Low Siting Beta | High Siting Beta |
| Debt pricing | Tighter spreads, longer tenors | Wider spreads, shorter tenors, more covenants |
| Equity return hurdle | Standard infrastructure IRR | Elevated IRR to compensate delay/cancellation risk |
| Contingency reserves | Standard construction contingency | Enlarged reserves for litigation, redesign, delay |
| Guarantees | Limited sponsor support | Completion guarantees, parent guarantees |
| Insurance | Standard builders’ risk | Political/regulatory riders, higher premiums |
| Land value | Premium for “risk-cleared” acreage | Discount for contested or unentitled land |
1.4 From Local Opposition to Wall Street
The transmission chain from a public hearing to a balance sheet runs through five links, each of which is now documented rather than hypothetical. First, opposition creates delay: rezoning denials, appeals, moratoria, and audits stretch timelines — Indianapolis has paused new approvals through the end of 2027; New York legislators passed a one-year permit moratorium for large sites; seventeen Indiana counties have enacted temporary moratoria and two have banned new facilities outright.[22][23][24] Second, delay creates cost: construction interest accrues, labor and equipment escalate, and — in the most expensive failure mode of the AI era — GPUs depreciate in warehouses while the buildings meant to house them await permits. Third, cost and delay create credit questions: lenders must now ask whether a project will ever reach commercial operation, which is why “readiness” has entered the underwriting lexicon.[10] Fourth, credit questions create pricing: wider spreads, larger reserves, stronger guarantees. Fifth, pricing creates valuation: the asset is worth less, the platform that owns it is worth less, and — at the limit — the equity story of an entire region changes.
The most striking feature of this chain is its speed. In the Cedar Creek Lake case, the interval between a grassroots petition and a governor’s statewide interconnection pause was measured in weeks, and the interval between the governor’s letter and formal legal advisories to lenders was measured in days.[2][9] Local opposition no longer takes years to reach Wall Street. It arrives before the next earnings call.
Section 2: The Six Components of Siting Beta
Siting Beta is a composite. Like any composite risk factor, it becomes useful only when decomposed into elements that can be separately observed, separately priced, and separately mitigated. This section develops each of the six components in turn, in each case beginning with the underlying physical or institutional reality and ending with its financial expression.
2.1 Power Risk
Power Risk is the probability-weighted cost of not getting electricity — or getting it later, in smaller quantities, on worse terms, or with politically explosive cost-shifting attached. It is the largest single component of Siting Beta because electricity is the one input for which there is no substitute, no import channel, and no inventory.
The magnitudes involved have broken every planning assumption in the utility industry. The International Energy Agency reports that electricity demand from data centers grew 17 percent globally in 2025 — and demand from AI-focused data centers grew roughly 50 percent — against total global electricity demand growth of only 3 percent; the IEA projects data-center consumption will double by 2030, reaching around 945 terawatt-hours, with AI-focused consumption tripling.[25] Brookings notes that if data centers were a country, their electricity consumption by 2026 could rank them fifth in the world, between Japan and Russia.[26] In the United States, the concentration is extreme: PJM’s long-term forecast attributes 94 percent of its projected 32 gigawatts of peak-load growth between 2024 and 2030 to data centers, and the Dominion zone alone expects more than 20,000 megawatts of data-center growth by 2037.[27]
Power Risk has several distinct faces. There is interconnection risk: the ERCOT queue reached roughly 474 gigawatts of requests before Abbott’s audit froze new approvals, and the audit itself will take months — months during which no developer in the queue can bank a connection date.[7][8] There is transmission risk: new lines take five to ten years, cross multiple jurisdictions, and generate their own opposition, as Pennsylvania’s contested 222-mile Kammer-Juniata line demonstrates.[28] There is generation risk: Microsoft’s own leadership has acknowledged holding GPUs in inventory for lack of electricity to energize them — the clearest possible statement that power, not silicon, is the binding constraint.[29] And there is the newest face, allocation risk: the political question of who pays. Texas Senate Bill 6 imposed disclosure and curtailment obligations on loads of 75 megawatts or more; Abbott’s June 2026 directive requires data centers to bear the full cost of the electric infrastructure that serves them; Virginia’s State Corporation Commission created a new rate class requiring large data-center customers to pay for at least 85 percent of contracted transmission and distribution demand and 60 percent of generation demand.[9][30] Every one of these rules changes project economics after the fact — which is precisely why Power Risk is a beta and not a budget line.
2.2 Permit Risk
Permit Risk is the risk that legal authorization to build and operate is denied, delayed, revoked, litigated, or rendered conditional in ways that change the project’s economics. It differs from Power Risk in a crucial way: it is discretionary. Electrons obey physics; permits obey politics.
The 2026 landscape shows Permit Risk operating at every level of government simultaneously. At the municipal level, moratoria have proliferated: at least 19 Michigan municipalities enacted data-center moratoria after the Saline Township episode; 17 Indiana counties have temporary moratoria and two — Marshall and Cass — have banned new facilities altogether; Indianapolis’s City-County Council voted 23-1 to pause new approvals through the end of 2027.[17][24][22] At the state level, more than 300 data-center bills were filed in the first six weeks of 2026, moratorium proposals appeared in 14 states, New York’s legislature passed a one-year permit pause for large-scale sites, and Maine came within one vote of a statewide ban.[14][23] At the federal level, the Clean Air Act has become an active battleground, as the xAI litigation discussed in Section 5 demonstrates.[31]
Two features of Permit Risk deserve emphasis. The first is permit durability: a permit granted is not a permit kept. Projects have been unwound after approval — QTS terminated the Prince William Digital Gateway after sustained community pushback — and audits like the Texas review can effectively reopen questions that developers considered settled.[11][6] The second is asymmetric optionality: opponents need to win only once, at any of a dozen procedural stages, while developers must win at every stage. This asymmetry means Permit Risk compounds with process length, and it explains why sophisticated developers now invest in pre-application community engagement: the cheapest permit is the one that is never contested.
2.3 Water Risk
Water Risk is the risk that cooling water is unavailable, restricted, contested, or so politically radioactive that it dominates the public narrative around a project. Of all six components, Water Risk is the one most underestimated by capital markets — and the academic evidence for it is now the strongest.
The definitive work comes from Professor Shaolei Ren of UC Riverside, together with co-authors including Caltech professor Adam Wierman. Their March 2026 study, “Small bottle, big pipe: Quantifying and addressing the impact of data centers on public water systems,” found that without new efficiencies, U.S. data-center cooling could require 697 million to 1.45 billion gallons per day of additional peak water capacity by 2030 — roughly the daily supply of New York City — and that the required community water infrastructure would cost between $10 billion and $58 billion.[32] The deeper insight is that averages deceive: it is the peak withdrawal on the hottest days of the year, precisely when municipal systems are most stressed, that breaks local water systems. And money alone cannot solve the problem:
“But reservoirs and snowpack are limited.” — Professor Shaolei Ren, UC Riverside [32]
Ren has also emphasized the mismatch between asset lives — water infrastructure serves for 50 or 60 years while a data center is designed for 15 or 20 — and the fundamentally local character of the resource:
“Water is a hyperlocal resource.” — Professor Shaolei Ren, in E&E News [33]
Water Risk is already reshaping outcomes. The Cedar Creek Lake project died in significant part on the reported five-million-gallon daily draw from a drinking-water reservoir during a drought year.[3][2] In Michigan, a regional water authority has refused service to proposed facilities.[17] In Arizona, water anxiety in Tucson, Chandler, and Marana catalyzed the state’s incentive moratorium.[34] The engineering response — closed-loop and air-cooled systems that trade water for electricity — is real and spreading, but it does not eliminate Water Risk; it converts part of it back into Power Risk, and it raises capital cost. There is no free thermodynamic lunch, only a choice of which scarce resource to consume and which community constituency to answer.
2.4 Externality Risk
Externality Risk covers the lived, sensory, and environmental impacts that fall on neighbors: noise from cooling fans and backup generators, emissions from on-site generation, the visual mass of windowless buildings and transmission corridors, construction traffic, and light. These impacts rarely appear in a pro forma, yet they are the raw material from which opposition movements are built, because they are what residents actually experience.
Noise deserves particular attention because it has become the signature grievance of the AI buildout. The low-frequency hum of large cooling plants carries far in quiet rural landscapes, and diesel or gas backup generators — tested regularly — produce episodic noise spikes that neighbors describe vividly at public hearings. In Jeffersonville, Indiana, residents near a Meta site testified that generator decibel testing caused severe noise pollution, contributing to a year-long municipal moratorium.[35] Cedar Creek Lake organizers put fan noise at the center of their campaign.[3] Saline Township’s settlement includes explicit noise caps.[36]
Emissions risk is best illustrated by the most aggressive buildout in the industry. xAI’s Colossus facilities in the Memphis area were powered in substantial part by dozens of unpermitted mobile gas turbines; in April 2026 the NAACP, represented by Earthjustice and the Southern Environmental Law Center, sued xAI over 27 unpermitted turbines at the Southaven, Mississippi site serving Colossus 2, alleging a clear violation of the Clean Air Act in a community already burdened by poor air quality.[31]
“The scale of it is astonishing.” — Patrick Anderson, Southern Environmental Law Center [37]
Externality Risk converts to financial risk through litigation exposure, retrofit requirements (sound walls, best-available-control-technology installations, enclosure redesigns), operating restrictions, and — most importantly — through its feeding of Community Risk. Every decibel and every plume is an organizing tool.
2.5 Community Risk
Community Risk is the probability that organized local opposition — public hearings, neighborhood associations, petitions, lawsuits, social-media groups, and local press — delays, conditions, shrinks, or kills a project. It is the component that has grown fastest, and it is the component that traditional infrastructure underwriting was least equipped to see, because it is made of people rather than of engineering.
The empirical record of 2025-2026 establishes three facts. First, community opposition is effective: 75 projects worth roughly $130 billion faced opposition in Q1 2026 and a further $98 billion was blocked or delayed in Q2; over a dozen Indiana projects were withdrawn after pushback; Google withdrew its $1 billion, 468-acre “Project Flo” rezoning in Franklin Township, Indiana, after council opposition and resident protests.[10][11][24][38] Second, it is bipartisan: Data Center Watch finds active opposition groups spanning 49 states, drawing from both parties, making data-center skepticism a rare zone of cross-political consensus in American infrastructure politics.[14] Third, it is networked: coalitions share playbooks, and a victory in one township — Cedar Creek’s petition, Saline’s settlement terms — becomes a template in the next.
Community Risk also has a subtler dimension: the jobs narrative has weakened. Consumer Reports notes that a typical 250,000-square-foot facility employs roughly 50 full-time workers, and that in November 2025 two business-school professors published research finding no clear evidence of local tech-employment gains attributable to data centers.[39] When the standard economic-development argument loses credibility, the remaining case for community acceptance must be built from real, negotiated benefits — which is why Section 3.4 treats community acceptance as a form of risk mitigation with a computable value.
2.6 Political and Election Risk
Political Risk is the risk that the rules of the game change — through gubernatorial directives, legislation, regulatory rulings, tax-incentive repeal, ballot measures, or a change of administration — after capital is committed. Election Risk is its calendar-driven special case: the predictable intensification of political risk as elections approach and incumbents respond to voter sentiment.
The 2026 cycle offers a natural experiment without precedent. In Texas, a Republican governor seeking re-election paused data-center approvals statewide and pledged to work with the legislature to repeal sales-tax exemptions for data centers in the 2027 session.[6][40] His Democratic challenger attacked the pause as inadequate rather than excessive:
“Abbott’s call for a ‘pause’ could be for one day.” — State Representative Gina Hinojosa [6]
When both candidates in a general election compete on who will be tougher on data centers, the political equilibrium for the industry has moved, whoever wins. In Pennsylvania, a Democratic governor in a swing-state re-election year built a “bring your own energy” doctrine and formal GRID Standards around the same voter anxieties.[28][41] In Arizona, a Democratic governor and a Republican legislature compromised on a three-year moratorium on data-center tax incentives.[34] In Washington, the administration pushed federal permitting acceleration and FERC authority over interconnection even as Congress introduced ratepayer-protection legislation pulling the other way.[42][15]
Political Risk has one property that distinguishes it from every other component: it is retroactive in effect even when prospective in form. A tax-incentive moratorium applies to new applications, but it devalues the option embedded in every site bank assembled on the assumption of incentive continuity — an effect the Data Center Coalition’s Chris Diorio described bluntly when Arizona’s pause was enacted:
“…it’s going to push development elsewhere.” — Chris Diorio, Data Center Coalition [34]
That sentence is Siting Beta operating as designed: capital observing a political signal in one jurisdiction and re-routing to another.
Summary Table: The Six Components
| Component | Core Question | Primary Evidence (2025-2026) | Financial Expression |
| Power Risk | Will electrons arrive, when, and who pays? | 474 GW ERCOT queue; Texas audit; VA rate class | Delay cost, tariff exposure, curtailment discounts |
| Permit Risk | Will authorization be granted and endure? | 300+ state bills; 14 moratorium states; NY pause | Approval probability, litigation reserves |
| Water Risk | Is peak cooling water available and acceptable? | $10-58B water infrastructure gap (Ren et al.) | Capex add-ons, cooling redesign, service denial |
| Externality Risk | Noise, emissions, traffic, visual impact? | Jeffersonville noise; xAI turbine litigation | Retrofit cost, operating restrictions, legal exposure |
| Community Risk | Will neighbors consent or mobilize? | $130B opposed Q1-2026; $98B blocked Q2 | Delay/cancellation probability, benefit payments |
| Political Risk | Will the rules survive the next election? | TX pause; AZ incentive moratorium; midterms | Regime-change discount, incentive durability haircut |
Section 3: How Siting Beta Enters the Cost of Capital
The previous section decomposed Siting Beta into its physical and political parts. This section traces how those parts flow into the numbers that actually govern investment: interest rates, return hurdles, reserves, guarantees, insurance, and valuation. The central claim is that Siting Beta is not a soft “ESG” consideration bolted onto the side of a deal; it is hard financial arithmetic that changes expected cash flows and their variance, and any pricing model that ignores it will systematically misprice AI infrastructure.
3.1 Delay Has a Price
Delay is the most common expression of Siting Beta, and it is expensive in five compounding ways.
First, construction interest. A multibillion-dollar project drawing on construction facilities accrues interest whether or not concrete is being poured; a year of delay on a $10 billion project financed at prevailing infrastructure rates adds hundreds of millions of dollars of carry before a single token is served.
Second, cost escalation. Labor, transformers, switchgear, and turbines are all supply-constrained; power-transformer lead times have stretched to well over two years, meaning that a delayed project does not simply resume where it left off — it re-enters queues for equipment at higher prices.[29]
Third, technology depreciation. This cost is unique to the AI era and dominates the others. GPUs improve on an annual cadence; a data center delayed eighteen months houses accelerators a generation behind the frontier, or forces the sponsor to hold inventory. Microsoft’s acknowledgment that GPUs have sat in inventory awaiting electricity converts this from a theoretical concern to an observed cost at the largest scale.[29]
Fourth, lost operating revenue. Every month of delay is a month of forgone lease or compute revenue at precisely the moment when demand exceeds supply — Google Cloud’s contract backlog alone has been reported in the hundreds of billions of dollars.[29]
Fifth, contractual exposure. Delay cascades through power purchase agreements, water contracts, tenant commitments, and equipment orders, triggering the change-in-law and force-majeure reviews that Texas lawyers advised immediately after the August 2026 pause.[9]
The delay channel explains an otherwise puzzling fact: why developers pay millions in community benefits for projects worth billions. Saline Township residents secured roughly $14 million in community benefits — fire-department funding, farmland preservation, environmental restrictions — attached to a $16 billion project.[17] Fourteen million dollars is less than one-tenth of one percent of project cost. If those payments shorten the expected timeline by even weeks, they are among the highest-return investments in the entire capital stack.
3.2 Cancellation Risk
Beyond delay lies the terminal outcome: the project never operates. Cancellation was once considered a tail risk in data-center finance; the 2025-2026 record has moved it into the body of the distribution. Diode’s Cedar Creek withdrawal, QTS’s termination of the Prince William Digital Gateway, Google’s withdrawal of Project Flo in Franklin Township, more than a dozen withdrawn Indiana projects, and Heatmap’s finding of at least 25 cancellations in 2025 in response to local objections together establish that cancellation is a realized, recurring outcome, not a hypothetical.[1][11][38][24][39]
For lenders, cancellation risk changes the fundamental question from “what is my spread?” to “will this asset exist?” This is why Reuters found banks assessing community sentiment a year or more before construction and why readiness now sits alongside credit quality in loan reviews.[10] For equity, cancellation risk truncates the upside distribution: a cancelled project returns land value minus sunk development cost, which is typically deeply negative once engineering, legal, deposit, and reputational costs are counted.
3.3 Higher Beta, Higher Financing Cost
Assembling the channels: higher perceived Siting Beta raises financing cost through wider debt spreads and shorter tenors; larger contingency reserves; stronger sponsor guarantees and completion support; more restrictive covenants tied to permitting milestones; higher insurance premiums; and elevated equity hurdles. The S&P Global figure — $121.91 billion of lender credit committed to U.S. data-center properties in 2025, with an average reported financing of $1.2 billion against a median of just $40 million — shows why the system is sensitive: a small number of massive projects create outsized, concentrated exposures for individual lenders and syndicates, exactly the structure in which a single high-beta failure can move an institution’s whole book.[16]
It is important to state what has not yet happened. No exchange quotes a “Texas audit spread.” Reuters explicitly notes that no bank has published a repriced loan attributable to community opposition, and that the shift should be treated as directional.[11] Siting Beta today is where climate risk was in the mid-2010s: visible in underwriting behavior, present in due-diligence checklists, priced implicitly through deal selection, but not yet standardized into a quoted premium. The trajectory, however, is unmistakable — and frameworks like the one proposed in this paper are intended to accelerate the standardization.
3.4 Community Acceptance as Risk Mitigation
If Siting Beta raises the cost of capital, then anything that verifiably lowers Siting Beta creates financial value — and this reframes community engagement from a public-relations expense into a risk-mitigation investment with a computable return. The toolkit now visible across the industry includes: community-benefit agreements with enforceable terms (Saline’s $14 million package with noise caps, water limits, and expansion restrictions); closed-loop and air-cooled cooling commitments that neutralize the water narrative (pledged by Diode for future sites and built into the Saline design); infrastructure cost self-funding, which answers the ratepayer grievance before it is raised (the Abbott doctrine, the Shapiro “bring your own energy” doctrine, and Virginia’s new rate class all convert this from voluntary virtue to mandatory rule); early and transparent engagement, whose absence was the explicit grievance at Cedar Creek; and local investment in fire departments, roads, schools, and farmland preservation.[17][36][4][28][30]
Stanford’s Anjney Midha, who teaches a widely followed course on AI infrastructure, has argued that opposition stems less from hostility to technology than from opacity about impacts and purpose — and has proposed nutritional-label-style disclosure of energy, water, and community impacts as the price of durable support:
“That is the level at which communities will want clarity.” — Professor Anjney Midha, Stanford University [43]
The financial translation is direct: transparency and benefit-sharing are premium payments on an insurance policy against delay and cancellation. The premium is small; the insured value is the timeline of a multibillion-dollar asset.
3.5 The Probability-Adjusted AI Factory
The conceptual endpoint of this section is a revised valuation model. Conventional project NPV discounts a deterministic cash-flow schedule at a rate reflecting market and credit risk. The probability-adjusted AI factory replaces the deterministic schedule with a distribution over three outcomes:
Expected Value = p(complete on time) x NPV(base) + p(delay) x NPV(delayed) + p(cancel) x (Salvage – Sunk Cost)
where the three probabilities are functions of Siting Beta, and NPV(delayed) itself embeds the five delay costs of Section 3.1. Under this model, two sites with identical NPV(base) diverge exactly as their Siting Beta components diverge — and the model makes explicit what practitioners already do intuitively when they “lean toward welcoming states.”[10]
An illustrative (deliberately stylized) example shows the leverage:
| Scenario | p(on time) | p(2-yr delay) | p(cancel) | Value vs. Base NPV |
| Low Siting Beta site | 85% | 13% | 2% | ~96% of base |
| Moderate Siting Beta site | 60% | 32% | 8% | ~85% of base |
| High Siting Beta site | 35% | 40% | 25% | ~62% of base |
The exact numbers matter less than the shape: modest shifts in completion and cancellation probabilities — shifts entirely consistent with the observed 2026 frequency of moratoria, audits, and withdrawals — produce valuation differences of a third or more between physically identical assets. That is the Siting Spread, derived from first principles.
Section 4: America’s Emerging Siting Beta Map
Siting Beta is not uniform across the United States; it is a map. This section reads that map state by state, choosing six jurisdictions that together span the full range of the 2026 experience: Texas, Virginia, Pennsylvania, Arizona, Indiana, and Michigan. Each state illustrates a different mechanism by which local conditions become financial conditions, and together they show that no political configuration — red or blue, deregulated or managed, water-rich or arid — is immune.
4.1 Texas — Abbott and the End of Automatic Growth
Texas entered the AI era as the archetypal low-friction jurisdiction: cheap land, an independent grid designed for speed, no state income tax, and a political culture built on welcoming industrial capital. It now hosts the nation’s second-largest concentration of data centers.[40] The 2026 story of Texas is therefore the most important single data point in this paper: it demonstrates that Siting Beta can rise fastest precisely where growth was most automatic, because unmanaged growth manufactures its own backlash.
The sequence bears restating as a timeline. In June 2026, Abbott issued a directive establishing standards: data centers must pay their own infrastructure costs, add grid capacity, use water-efficient cooling, respect neighborhood setbacks, and never raise residential electric bills.[9][44] In July, the Cedar Creek Lake project was withdrawn against that standard, with the governor publicly warning that other non-conforming projects should expect the same.[1] On August 3, the interconnection pause and audit followed, freezing ERCOT’s Batch Zero process with roughly 474 gigawatts in the queue.[6][7] By mid-August, the PUCT confirmed 250 to 300 projects would be audited over several months, and QTS — one of the largest operators in North America — publicly endorsed the audit, calculating that visible compliance with the new standard is now a competitive asset.[8][44] Abbott has further pledged to seek 2027 legislation codifying cost-bearing requirements, water-efficient cooling mandates, capacity-addition requirements, siting and setback rules, and repeal of data-center sales-tax exemptions.[40]
The financial reading: Texas has converted from a low-beta to a transitional-beta jurisdiction — near-term uncertainty is elevated (an audit of unknown duration, a legislature not in session until 2027, an election in November), but the end-state may be a codified, predictable regime. As Pillar 3 will argue, investors can price strict rules; what they struggle to price is the interregnum, and Texas is in one now. Notably, the pause also creates a perverse incentive gradient: projects with on-site generation can bypass the frozen queue entirely, accelerating the “bring your own power” architecture — and shifting risk from the Power component of Siting Beta to the Externality and Permit components, as the xAI experience shows.[6]
4.2 Virginia — From Data-Center Capital to Cost Allocation
Northern Virginia is the largest data-center market on Earth, and Virginia is therefore the jurisdiction where the ratepayer question matured first. The state’s own watchdog, the Joint Legislative Audit and Review Commission (JLARC), produced the analytical foundation: data centers currently pay their full cost of service, but unconstrained growth could roughly double statewide electricity demand within a decade, require generation and transmission buildout at rates the state has never achieved, and raise a typical Dominion residential bill by an estimated $14 to $37 per month.[45][46] Meanwhile, data-center demand contributed to an 833 percent increase in PJM’s capacity-auction price for 2025-2026, and wholesale prices in Virginia spiked above $1,000 per megawatt-hour during the July 2026 heat wave.[30][27]
Virginia’s response has been to re-engineer cost allocation rather than to prohibit growth. In November 2025 the State Corporation Commission approved a new rate class requiring large data-center customers, from January 2027, to pay for at least 85 percent of contracted distribution and transmission demand and 60 percent of generation demand.[30] Then, in a July 31, 2026 ruling, the SCC concluded that new large-load data centers are the cause of major transmission costs and must pay for transmission facilities constructed solely to serve them — a ruling Governor Abigail Spanberger’s administration had urged and celebrated as saving families and small businesses hundreds of millions of dollars.[47][48] Legislative pressure continues: SB 253 and successor measures would prohibit utilities from passing data-center-driven costs to other customers, and at least eight states introduced similar legislation in their 2026 sessions, making Virginia the exporter of the cost-allocation template.[27]
Virginia’s Siting Beta profile is thus distinctive: Community Risk remains high in saturated corridors (the Prince William Digital Gateway termination is the emblematic loss), but Political Risk is falling as rules crystallize.[11] A jurisdiction that tells developers exactly what they will pay — even if the number is large — offers something that a jurisdiction mid-audit cannot: a number.
4.3 Pennsylvania — Conditional AI Development
Pennsylvania illustrates the “conditional yes”: an energy-exporting state actively courting AI capital while constructing guardrails in real time. Governor Josh Shapiro has championed marquee projects — Amazon’s $20 billion, two-campus commitment, the largest private investment in state history; the conversion of the former Homer City coal plant into a gas-powered AI campus with the largest planned gas generation in the country; the restart of the Crane Clean Energy Center — while simultaneously acknowledging voter anxiety in a state where average household electricity rates jumped nearly 14 percent in a year.[41][28] His February 2026 budget address captured the balancing act:
“We need to be selective about the projects that get built here.” — Governor Josh Shapiro [49]
The machinery of selectivity arrived in May 2026 as the Governor’s Responsible Infrastructure Development (GRID) Standards — developed with local leaders, labor, industry, and environmental stakeholders, and specifying what developers must demonstrate on community value, impact mitigation, and responsible development — paired with the “bring your own energy” (BYOE) doctrine pushing data centers to fund or supply their own generation rather than lean on the ratepayer base.[41][28] In June, the Pennsylvania House passed legislation giving townships the option to pause new applications while codifying the Shapiro framework.[50] And in August 2026, the politics arrived at the ballot box: rallies targeting the governor by name, a Republican opponent promising a natural-gas-first fast-track agenda, and data centers functioning as a live barometer of swing-state sentiment ahead of November.[51]
Pennsylvania’s lesson for the Siting Beta map is that conditionality is a price schedule. GRID Standards and BYOE tell developers what consent costs. Whether that schedule survives the election — and whether the industry group’s warning that the standards create a complicated framework proves prophetic — will determine whether Pennsylvania lands as a medium-beta jurisdiction with high throughput or a high-beta jurisdiction with high rhetoric.[52]
4.4 Arizona — Water and Incentive Durability
Arizona compresses two components of Siting Beta — Water Risk and Political Risk — into a single case. The state built a top-ten national data-center position on a 2013 sales-tax exemption; nearly 98 facilities operate and 86 more are planned or under construction.[53] But intense 2025 backlash against proposed projects in Tucson, Chandler, and Marana — with Tucson’s contested “Project Blue” moving from land-use approval into litigation, utility-rate disputes, and water controversy — turned the 2026 legislative session into a referendum on the subsidy itself: data centers featured in at least 84 hearings and caucus meetings, more than 50 energy bills were introduced, and the budget deal signed by Governor Katie Hobbs imposed a three-year moratorium on new data-center tax-incentive applications, from July 1, 2026 through June 30, 2029.[34][54] Hobbs, who had sought outright repeal, framed the pause as policy hygiene rather than prohibition:
“Nobody’s talking about a moratorium on data centers themselves.” — Governor Katie Hobbs [53]
Developers rushed to file applications before the window closed — behavior that is itself a textbook demonstration of incentive-durability risk being priced in real time — and the Data Center Coalition warned that uncertainty about Arizona’s long-term commitment would push development elsewhere.[34] The contagion continued downstream: Pima County directed staff in August 2026 to develop a county-level moratorium, citing the state’s action as precedent.[55] Arizona thus teaches the general lesson of incentive durability: a tax exemption is an option written by the state, and 2026 demonstrated that the state can decline to renew the option. Every pro forma in every incentive-dependent jurisdiction now carries a durability haircut that did not exist two years ago — and Ohio and Illinois have joined Arizona in pausing exemptions while they study the question.[40]
4.5 Indiana — Development Versus Resource Constraints
Indiana is the clearest natural experiment in the divergence between state-level enthusiasm and county-level resistance. The state government, under Governor Mike Braun, has aggressively courted hyperscale investment and landed it: Meta’s $10 billion Lebanon campus, Amazon’s $11 billion New Carlisle campus, Google’s $2 billion Fort Wayne facility, with total announced investment exceeding $28 billion across 46 tracked projects.[19][19] Yet by mid-2026, Indiana University’s Environmental Resilience Institute counted 11 counties with data-center ordinances, at least 17 with temporary moratoria, and two — Marshall and Cass — with outright bans; over a dozen projects had been withdrawn after pushback; and the Indianapolis City-County Council voted 23-1 to pause new approvals in Marion County through the end of 2027 while it develops standards for noise, water management, and verified electrical capacity.[24][22] Even a member of Indiana’s congressional delegation introduced a federal AI Data Center Moratorium Act.[24]
The Indiana pattern — a green light at the statehouse and a patchwork of red lights at the county line — defines a specific and underappreciated form of Siting Beta: jurisdictional layering risk. A state incentive cannot overcome a county ban; a county approval cannot conjure water or transmission; and a project needs every layer to say yes. For site selectors, Indiana proves that state-level policy signals are necessary but radically insufficient information, and that genuine diligence must now reach down to the township ordinance and the county commission’s meeting calendar.
4.6 Michigan — AI Infrastructure Enters Electoral Politics
Michigan supplies the paper’s most complete single narrative, because the Saline Township saga contains every component of Siting Beta in one 250-acre story. A consortium of Oracle, OpenAI, Related Digital, Blackstone, and Walbridge selected a farming township of roughly 2,400 people, partly for its access to existing transmission capacity.[56][36] After months of contentious public meetings, the township board voted 4-1 in September to deny the rezoning. Two days later, the developer and landowners sued, alleging exclusionary zoning; within weeks the township settled, accepting roughly $14 million in community benefits, noise caps, water-use limits, farmland preservation, and expansion restrictions; seven weeks after the settlement, OpenAI and Oracle announced the site as part of Stargate; and by April 2026, Related Digital had secured financing for what had grown from a $7 billion proposal into a $16 billion, 1.4-gigawatt campus.[56][36][17]
The project was built — but the political outcome was the opposite of a victory for frictionless development. At least 19 Michigan municipalities have since enacted data-center moratoria; counties have passed resolutions; bipartisan state legislation has been introduced; and a regional water authority has refused to serve proposed facilities.[17] A resident’s assessment of the developer’s strategy — move so fast that the project is too far along to stop by the time anyone can challenge it — has become part of the national organizing playbook precisely because it circulated so widely.[17] Michigan thus demonstrates the boomerang property of Siting Beta: winning a single contested site through litigation can raise the beta of an entire state for every developer who follows. The Saline campus may prove to be simultaneously one of the most successful sitings and one of the most expensive precedents in the industry’s history.
4.7 The Geographic Repricing of AI
Assembling the map, the correct comparison between states is no longer electricity price per megawatt-hour; it is total Siting Beta. A stylized scoring of the six case-study states as of August 2026:
| State | Power Risk | Permit Risk | Water Risk | Community Risk | Political Risk | Overall Trajectory |
| Texas | High (audit/queue) | Rising | Moderate-High | High | High until 2027 codification | Transitional; repricing now |
| Virginia | Moderate (priced) | Moderate | Moderate | High in core corridors | Falling (rules crystallizing) | Predictable but costly |
| Pennsylvania | Moderate (BYOE) | Moderate | Moderate | Rising | Election-dependent | Conditional yes |
| Arizona | Moderate | Rising (county layer) | High | High | High (incentive durability) | De-rating |
| Indiana | Moderate | High (county patchwork) | Moderate-High | High | Split state/local | Layered risk |
| Michigan | Moderate | High (moratoria wave) | Moderate | Very High | Rising into midterms | Boomerang effect |
The deeper point of the map is dynamic, not static. Jurisdictions move. Texas moved from lowest-friction to mid-audit in ninety days. Virginia moved from open-ended exposure toward codified cost allocation. The Siting Beta of a state is not a fixed attribute of its geography or party registration; it is a function of how visibly, predictably, and fairly it manages the collision between AI capital and its own residents — which is why the winning strategy for both companies and states, developed in Sections 5 and 8, is the deliberate compression of beta rather than the search for a mythical zero-beta jurisdiction.
Section 5: Hyperscalers Learn to Compress Siting Beta
If Siting Beta is a cost, then the companies with the most capital at stake have the most to gain from compressing it — and 2025-2026 shows the hyperscalers converging, from very different starting points, on a common playbook: secure power early, neutralize the water narrative with closed-loop engineering, pay visibly into communities, diversify geographically, and treat political relationships as infrastructure. This section examines how each major player is adapting, and what their divergent strategies reveal about the trade-offs.
5.1 Meta
Meta’s posture is defined by scale and by the willingness to fund the ecosystem around its campuses. Mark Zuckerberg has framed the buildout in explicitly generational terms, guiding 2026 capital expenditure to a range that has been raised twice — toward $125-145 billion — and declaring the company’s intent to:
“build tens of gigawatts this decade” — Mark Zuckerberg, Meta [29]
Meta’s siting practice pairs very large campuses (the $10 billion Lebanon, Indiana facility; the El Paso project with BlackRock financed through a $12.3 billion bond sale) with structured financing that distributes risk to institutional capital, and with community programs and utility partnerships intended to blunt the ratepayer critique.[24][10] Yet Meta’s experience also demonstrates the limits of money without acceptance: the Jeffersonville, Indiana moratorium was enacted around an under-construction Meta site amid resident fury over generator noise and water-priority provisions — proof that Externality Risk can ignite even where the checkbook is open.[35]
5.2 Amazon and AWS
Amazon operates the largest single capex program in the industry — roughly $200 billion projected for 2026 — and its siting strategy emphasizes state-level partnership at record scale: the $20 billion Pennsylvania commitment across Luzerne and Bucks counties, negotiated with a governor who made it the centerpiece of his economic message, and the $11 billion New Carlisle campus in Indiana.[20][49][19] Amazon’s approach to Siting Beta is regional diversification plus utility negotiation: long-term energy contracts, participation in state incentive frameworks, and a portfolio spread wide enough that no single audit, moratorium, or election can strand a material share of its roadmap. Notably, when Texas paused approvals, the governor’s own campaign cited Amazon — alongside Google and Microsoft — as having embraced his standards, illustrating the new equilibrium: the largest players purchase predictability by conspicuously complying with the strictest rules.[57]
5.3 Google and Microsoft
Google and Microsoft represent the engineering-forward wing of beta compression. Alphabet raised its 2026 capex outlook to $195-205 billion at Q2 2026 earnings, allocating roughly 60 percent to servers and 40 percent to data centers and networking, and has leaned on energy procurement innovation — advanced nuclear agreements, geothermal, carbon-free energy matching — plus formal water stewardship targets to lower both Power and Water Risk at the source.[58] Microsoft’s fiscal 2026 was a physical sprint: 88 data centers opened across five continents during the year, 31 in the final quarter alone, with dock-to-live times cut nearly in half:
“We added nearly 1 GW of new capacity this quarter.” — Satya Nadella, Microsoft [58]
Both companies have also learned the negative lessons. Google’s withdrawal of the Franklin Township “Project Flo” rezoning after resident protests showed that even the most sophisticated developer cannot out-engineer a hostile council; its subsequent Indiana strategy has emphasized earlier disclosure and utility-verified capacity.[38] Microsoft’s admission that GPUs have waited in inventory for electricity reframed its entire siting calculus around energization dates rather than land dates.[29] For both, geographic diversification — across states, countries, and continents — functions as the portfolio-level hedge against any single jurisdiction’s beta.
5.4 OpenAI
OpenAI presents the newest and most instructive case, because Stargate-scale ambitions — a $500 billion infrastructure program with Oracle and SoftBank — collide with the fact that OpenAI does not own decades of utility relationships, state-government goodwill, or community-affairs infrastructure of the kind the incumbents built over twenty years. The Saline Township episode is the canonical result: a rejected rezoning, a two-day-later lawsuit, a settlement, a $16 billion campus under construction — and nineteen municipal moratoria, bipartisan legislation, and a statewide backlash following behind it.[17][56] The strategic lesson OpenAI’s trajectory teaches the industry is that community relations cannot be retrofitted after site control; at Stargate scale, they are a core development competency, co-equal with power procurement and construction management. Every future Stargate site will be negotiated in the shadow of Saline — by counterparties who have read the coverage.
5.5 xAI and Elon Musk
xAI is the limit case of speed-maximalist development: colossal clusters stood up in months by colocating self-supplied gas generation, bypassing both the interconnection queue and, according to federal litigation, the Clean Air Act permitting process. The approach delivered roughly a gigawatt of compute across the Memphis-area Colossus facilities on timelines no competitor matched — and generated the highest Externality and Permit Risk profile in the industry: an EPA determination that the “temporary” turbines require permits, a Clean Air Act lawsuit by the NAACP represented by Earthjustice and SELC over 27 unpermitted turbines at Southaven, allegations of dozens more, and demands for injunctions, best-available-control retrofits, and daily penalties.[31][37] xAI thus embodies the central trade of Siting Beta management: it converted Power Risk (queue delay) into legal and community risk (litigation, environmental-justice opposition), betting that compute delivered now outruns liabilities assessed later. Whether that bet clears depends on courts — a discount-rate input no engineer can control.
5.6 Nvidia and the Indirect Geography of GPUs
Nvidia builds no data centers, yet no company has more exposure to Siting Beta, because every delayed or cancelled campus is delayed or cancelled demand for accelerators. The point became explicit in August 2026 when Nvidia moved to orchestrate roughly $500 billion of Wall Street capital toward AI infrastructure buildout — a chipmaker underwriting the physical and financial conditions of its own demand.[59] Nvidia’s revenue recognition happens at shipment, but its growth narrative happens at energization: a GPU in a warehouse in a paused jurisdiction is revenue already booked and demand not yet created. The five-layer analysis of Section 7 formalizes this transmission, but the summary is simple: the geography of permits has become an input to the valuation of semiconductors — a sentence that would have been unintelligible in 2022 and is a research-desk commonplace in 2026.
Section 6: Wall Street Discovers Geography
6.1 Banks Begin Underwriting Community Risk
The August 2026 Reuters investigation marks the moment the financial system formally acknowledged Siting Beta. Senior bankers described a two-part test — readiness and credit quality — in which readiness explicitly includes permitting status and the support of the surrounding community; banks reported leaning toward welcoming states; and the due-diligence burden of contested projects was described as a cost in itself, since funding conversations begin a year or more before construction.[10] This is a structural change in underwriting doctrine, not a mood. Community sentiment has moved from the “soft factors” appendix of a credit memo to the front page, alongside debt-service coverage — because the empirical base rates changed: $130 billion opposed in one quarter, $98 billion blocked or delayed in the next, and marquee projects like the Prince William Digital Gateway terminated outright.[10][11]
6.2 Private Credit and Infrastructure Funds
Behind the banks stands the deeper pool: private credit and infrastructure capital. Blackstone is an equity participant in the Saline consortium; BlackRock partners with Meta in El Paso behind a $12.3 billion bond; Morgan Stanley and KKR have led financings for CyrusOne in the $9.7 billion range; Apollo, Brookfield, and Blue Owl have each built dedicated digital-infrastructure strategies measured in the tens of billions.[56][10][11] For these investors, Siting Beta is double-edged. On one side, it is a risk to be diligenced — and private credit’s longer holding periods make cancellation risk more dangerous than it is for a syndicating bank. On the other side, it is a source of return: capital that can accurately price political and community risk earns the spread that capital fleeing uncertainty leaves behind. The emergence of specialist advisory practices, political-risk trackers like Data Center Watch, and law-firm rapid-response teams around events like the Texas pause is the infrastructure of a market learning to price a new factor.[9][11]
6.3 Insurance and Completion Guarantees
Insurance is the slowest-moving but ultimately most standardizing part of the capital stack, and it is beginning to absorb Siting Beta through three channels: builders’ risk and delay-in-start-up coverage that must now contemplate politically induced delay; sponsor completion guarantees demanded by lenders where permit durability is questionable; and the early exploration of political-risk-style wrappers for domestic regulatory reversal — a category historically reserved for emerging markets, now discussed for American counties. The direction of travel is clear from the underwriting logic: wherever a probability can be estimated from base rates (and Data Center Watch now publishes quarterly base rates), an actuary can price it, and once an actuary prices it, the Siting Spread stops being implicit and becomes a line item.[11]
6.4 Risk-Cleared Acreage
A new asset class is crystallizing at the intersection of land and law: risk-cleared acreage — sites with secured power allocations, water contracts, durable entitlements, executed community-benefit agreements, and demonstrated political support. The Texas audit makes the value of clearance literal: a project inside the audited queue holds an uncertain claim of unknown duration, while a project with completed interconnection holds a scarce, dated right.[8] The same logic explains why QTS publicly embraced Abbott’s standards — certified compliance converts a regulatory burden into a moat — and why Diode, exiting Cedar Creek, advertised closed-loop cooling and early engagement as the criteria for its next site: both companies are marketing their future acreage as low-beta.[44][4] Land without clearance is dirt; land with clearance is infrastructure. The spread between the two is widening every quarter, and it is the purest observable price of Siting Beta.
6.5 Siting Beta and M&A
The final Wall Street expression is corporate. As the buildout matures, acquisition targets will increasingly be valued not by megawatts operating but by megawatts deliverable — pipelines of entitled, powered, watered, community-accepted sites. A development platform whose portfolio sits in codified-rule jurisdictions with executed benefit agreements deserves — and will command — a premium over a nominally larger platform whose pipeline is trapped in audited queues and moratorium counties. Expect Siting Beta diligence to become as formalized in digital-infrastructure M&A as environmental diligence became after the 1980s: dedicated workstreams, specialist advisors, representations and warranties on community agreements, and purchase-price adjustments tied to permitting milestones. The companies assembling de-risked platforms today are, knowingly or not, manufacturing the premium acquisition currency of 2028.
Section 7: Siting Beta Across the Five-Layer AI Economy
The Five-Layer AI Economy — Energy, Chips, Data Centers, Models, and Applications — is the organizing schema of this research series, and Siting Beta touches every layer, though it is physically concentrated in only one. This section traces the propagation.
7.1 Layer 1 — Energy
Energy is where Siting Beta originates, because generation, transmission, fuel, and water are the site-specific inputs that no amount of capital can teleport. The IEA’s finding that data-center electricity demand will double by 2030 while AI-focused demand triples defines the aggregate pressure; the IMF’s modeling in its working paper “Power Hungry: How AI Will Drive Energy Demand” finds the AI boom producing manageable but policy-dependent increases in energy prices and emissions — with the crucial qualifier that outcomes vary with infrastructure constraints, which is to say, with siting.[25][60] Layer 1 is also where the political feedback loop closes: rising bills produce voter anger, voter anger produces the directives and rate classes documented throughout this paper, and those rules re-enter Layer 1 as cost-allocation regimes.
7.2 Layer 2 — Chips
Chips are globally mobile, but their deployment is not. Every audited queue position and moratorium county postpones the installation of Nvidia, AMD, and custom accelerators, and the industry’s own disclosures — GPUs in inventory awaiting electricity — prove that Layer 2 revenue can be booked while Layer 2 utility is stranded by Layer 3 delays.[29] Nvidia’s mobilization of Wall Street capital toward infrastructure is Layer 2 reaching down the stack to compress the beta that constrains it.[59]
7.3 Layer 3 — Data Centers
Layer 3 is where Siting Beta is physically concentrated: the buildings, the fans, the substations, the water mains, the county hearings. Everything in Sections 1 through 6 lives here. The layer’s defining 2026 statistic pair: roughly $760 billion of hyperscaler capex flowing in, and roughly $228 billion of projects opposed, blocked, or delayed across just the first two quarters — an opposition ratio no other infrastructure class in America approaches.[20][10][11]
7.4 Layer 4 — Models
Model training runs and inference fleets are scheduled against capacity that exists, not capacity that is promised. A delayed campus delays the training cluster it was to house; an audit of unknown duration makes multi-year compute roadmaps probabilistic. The EPRI working-paper debate over whether data centers raise or, under some structures, lower residential rates matters here too: the perception of rate impact drives the politics that gates the capacity that gates the models.[61] Frontier labs have responded rationally — multi-cloud contracts, geographic spread of training sites, and (in OpenAI’s case) direct co-development of Stargate campuses — which is Layer 4 vertically integrating downward to manage a Layer 3 risk.
7.5 Layer 5 — Applications and Agentic Systems
At the top of the stack, applications and agentic systems inherit whatever capacity survives the layers below. The IEA notes that while per-task energy efficiency is improving at a rate perhaps unprecedented in energy history, total demand rises anyway because usage — especially agentic usage — is exploding.[25] Inference capacity constraints translate directly into rate limits, latency, regional service quality, and enterprise deployment schedules. The most abstract layer of the AI economy is therefore chained, through four intermediating layers, to the most concrete of all facts: whether a county board in Indiana or a utility commission in Austin said yes.
7.6 The Five-Layer Transmission Effect
| Layer | What Siting Beta Does There | 2026 Evidence |
| 1. Energy | Originates: grid, fuel, water, cost allocation | ERCOT audit; VA rate class; IMF/IEA demand outlook |
| 2. Chips | Postpones deployment; strands utility of shipped GPUs | GPUs in inventory awaiting power |
| 3. Data Centers | Concentrates: delay, cancellation, spread | $228B opposed/blocked in H1 2026 vs. $760B capex |
| 4. Models | Converts capacity risk into roadmap risk | Multi-site training diversification; Stargate co-development |
| 5. Applications | Surfaces as capacity, latency, rollout limits | Agentic demand growth against constrained inference |
The transmission effect can be stated in one sentence: a local zoning decision propagates from land and electricity all the way to models and applications, which means the marginal unit of usable intelligence in the AI economy is now co-produced by engineers and by county commissioners.
Section 8: What Have We Learned? Seven Pillars
Pillar 1 — Geography Is Becoming Financial
The location of AI infrastructure is no longer merely an engineering decision; it is a determinant of financing cost, completion probability, and valuation. The evidence is now institutional rather than anecdotal: lenders formally screen for readiness and community support; a state audit forced immediate contractual review across billions in commitments; land with durable entitlements trades at a widening premium to land without them.[10][9] The correct mental model for an AI campus is no longer “a building with servers” but “a thirty-year claim on local electrons, local water, and local consent” — and claims are priced by their jurisdiction.
Pillar 2 — Community Acceptance Has Economic Value
Political consent reduces delay, litigation, cancellation risk, and financing uncertainty, and its price is astonishingly low relative to the value it protects: $14 million of community benefits against a $16 billion campus; early engagement against months of hearings; a closed-loop cooling commitment against a drinking-water controversy.[17][4] The industry’s historical instinct — secrecy through land-banking LLCs and non-disclosure agreements — is now demonstrably value-destroying, because opacity was the explicit grievance at Cedar Creek, in Tucson, in Jeffersonville, and in Saline. Consent is cheaper than litigation, and it compounds: a developer’s reputation for keeping community agreements is a portfolio asset that lowers the beta of every future site.
Pillar 3 — Predictability Can Matter More Than Deregulation
Investors can price strict rules. They have far greater difficulty pricing rules that change after billions are committed. Virginia’s trajectory — expensive, codified, explicit — is becoming more financeable even as its nominal costs rise, because the SCC’s rate class and transmission ruling convert open-ended ratepayer politics into a number.[30][47] Texas’s trajectory — historically the cheapest and freest — is temporarily less financeable, because an audit of unknown duration is unpriceable until it ends.[8] The deepest implication inverts a generation of site-selection orthodoxy: the question is not “which state has the fewest rules?” but “which state has the most durable ones?”
Pillar 4 — Scale Can Compress Siting Beta
The largest hyperscalers can afford self-supplied generation, closed-loop water systems, community-benefit programs, elite legal teams, multi-state diversification, and conspicuous compliance with the strictest standards — capabilities smaller developers cannot match. When Texas paused approvals, the companies cited as already embracing the governor’s standards were Amazon, Google, and Microsoft; when audits separate what regulators call responsible developers from the rest, incumbency deepens.[57][44] Siting Beta is therefore quietly concentrating the AI infrastructure industry: it functions as a fixed cost of consent that the largest balance sheets amortize easily and marginal entrants cannot, with consequences for competition policy that deserve their own future paper.
Pillar 5 — The AI Map Will Follow Both Megawatts and Political Consent
The winning AI regions will combine: Power + Water + Permitting + Community Acceptance + Political Durability + Capital. Any five of the six are insufficient. Texas has power and capital but is mid-audit; Arizona has capital and permits but contested water and revocable incentives; Michigan has transmission capacity but a mobilized citizenry. The regions that assemble all six — and can prove it — will absorb a disproportionate share of the trillion-dollar-per-year capex trajectory analysts now project for 2027 and beyond.[20][21]
Pillar 6 — Transparency Is Becoming the Currency of Consent
A consistent thread runs from Stanford’s classroom to the Cedar Creek petition: communities are not rejecting technology; they are rejecting mystery. Polling shows 43 percent of Americans already believe data centers raise their bills and roughly 70 percent would object to one nearby — numbers that opacity feeds and disclosure can contest.[43][13] The emerging remedy — nutritional-label-style disclosure of energy use, water use, noise profiles, cost allocation, and community impact — is simultaneously good politics and good finance, because standardized disclosure is precisely what allows Siting Beta to be measured, compared, and compressed rather than feared. The industry that invented the dashboard should not be losing public hearings for lack of one.
Pillar 7 — Elections Are Repricing Events
The 2026 cycle has established that data centers can decide, and be decided by, elections — and therefore that every election calendar is now a repricing calendar for AI infrastructure. Abbott’s directives arrived in a re-election year; Shapiro’s GRID Standards arrived in a swing-state cycle; Hobbs’s incentive moratorium emerged from a divided-government budget; congressional ratepayer legislation arrived ahead of the midterms; and polling shows voters ranking data centers among their top issues.[6][41][34][15][12] Sophisticated investors already model central-bank meeting dates and OPEC calendars; they must now model gubernatorial cycles, legislative sessions (Texas, 2027), and county election dates along the path of every project. Political time has joined construction time on the Gantt chart.
Conclusion: The Price of Place
The next phase of the AI infrastructure race will not be determined only by who owns the best GPUs, raises the most capital, or secures the cheapest electricity. It will increasingly depend on who can turn a location into a politically acceptable, physically sustainable, and financially bankable AI factory.
That is why I call this concept — and this paper — Siting Beta, and the name deserves a final defense now that the full argument is on the table. Beta, because the risk is systematic: it loads onto every project in a jurisdiction at once, cannot be diversified away by better engineering, and must therefore be priced — in spreads, hurdles, reserves, guarantees, and valuations — exactly as the capital markets price every other systematic factor. Beta, because the risk is differential: identical assets carry different exposures depending on where they stand, and finance has no better word for differential exposure. Beta, because the risk is measurable: quarterly opposition trackers, polling series, moratorium counts, audit scopes, and rate-class rulings now generate the base rates from which probabilities — and premiums — can be estimated. Beta, because the risk is manageable: it can be compressed by consent, transparency, self-supplied infrastructure, and durable rules, which makes the framework a strategy and not merely a lament. And Siting, because the binding constraints of artificial intelligence in 2026 are not mathematical but municipal: transformers, aquifers, decibels, ballots, and the patience of people who live near the fence line.
Permit Populism — the companion concept in this research series — explains why communities resist AI infrastructure: rising bills, strained water, industrial noise in rural quiet, subsidies without visible jobs, and decisions made about places without the people of those places. Siting Beta explains what happens when that resistance reaches the balance sheet: delay acquires an interest rate, cancellation acquires a probability, consent acquires a value, predictability acquires a premium, and geography acquires a price.
The empirical record of a single summer makes the case. Between June and August 2026, the governor of the largest data-center growth state in America established standards, celebrated a withdrawal, and paused an entire interconnection queue pending an audit of three hundred projects.[1][6][8] The regulator of the largest data-center market on Earth ruled that data centers must pay for the transmission built to serve them.[47] The financial press documented banks underwriting community sentiment.[10] A quarter-trillion dollars of projects met organized opposition in six months, while three-quarters of a trillion dollars of hyperscaler capex went looking for places to land.[10][11][20] Those two flows — capital seeking sites, and communities pricing consent — now meet in every county in America, and the clearing price of that meeting is Siting Beta.
As data centers become election issues, electricity issues, water issues, and ratepayer issues, the future winners of the Five-Layer AI Economy will be the companies — and the states — that learn to compress Siting Beta before billions of dollars are committed rather than litigate it afterward. The losers will discover, at the cost of stranded queues and boomerang moratoria, that in the age of artificial intelligence the scarcest input is not compute, not capital, and not even power.
It is a place that says yes — and keeps saying yes.
Footnotes and Endnotes:
[1] Office of the Texas Governor (Greg Abbott), “East Texas Data Center Withdraws After Falling Short of Governor Abbott’s Standards,” July 23, 2026. https://gov.texas.gov/news/post/east-texas-data-center-withdraws-after-falling-short-of-governor-abbotts-standards
[2] Alejandra Martinez and staff, The Texas Tribune, “Data center company decides not to build in East Texas,” July 23, 2026. https://www.texastribune.org/2026/07/23/east-texas-data-center-withdrawn/
[3] KSST Radio, “Data Center Developer Pulls Proposal in East Texas Following Fierce Local Pushback and Governor’s Directive,” July 23, 2026. https://www.ksstradio.com/2026/07/data-center-developer-pulls-proposal-in-east-texas-following-fierce-local-pushback-and-governors-directive/
[4] FOX 4 Dallas-Fort Worth, “Henderson County data center plans scrapped by developer,” July 2026. https://www.fox4news.com/news/henderson-county-data-center-plans-scrapped-developer
[5] The Daily Signal, “Major Data Center Proposal Withdrawn Following Talks With Texas Officials,” July 24, 2026. https://www.dailysignal.com/2026/07/24/major-data-center-proposal-withdrawn/
[6] Alejandra Martinez et al., The Texas Tribune, “New Texas data center projects frozen until state audits them,” August 3, 2026. https://www.texastribune.org/2026/08/03/texas-data-center-project-audit-greg-abbott/
[7] Robert Walton, Utility Dive, “Facing an estimated 474 GW of interconnection requests, Texas hits pause on data centers,” August 5, 2026. https://www.utilitydive.com/news/texas-hits-pause-data-center-interconnections/827046/
[8] The Texas Tribune, “Texas will audit up to 300 projects, mostly data center proposals,” August 14, 2026. https://www.texastribune.org/2026/08/14/texas-data-center-approval-pause-ercot-power-grid/
[9] Troutman Pepper Locke LLP, “Texas Hits Pause on Data Center Grid Connections Amid Growing Oversight Push,” August 2026. https://www.troutman.com/insights/texas-hits-pause-on-data-center-grid-connections-amid-growing-oversight-push/
[10] Reuters (Karen Fang interview; Goldman Sachs and Data Center Watch data), republished via Virginia Business, “Lenders scrutinize US data center financing as community opposition builds,” August 10, 2026. https://virginiabusiness.com/us-banks-scrutinize-data-center-financing-community-opposition/
[11] Alexis Dufresne, AI Weekly, “US Banks Add Community Opposition to Data Center Loan Checks” (summarizing Reuters and Data Center Watch Q1-Q2 2026 tracking), August 10, 2026. https://aiweekly.co/alerts/us-banks-add-community-opposition-to-data-center-loan-checks
[12] Reuters/Ipsos poll (Valerie Volcovici and Jason Lange), via Allwork.Space, “AI Data Center Boom Faces Public Backlash As 77% Fear Higher Electricity Bills,” June 2026. https://allwork.space/2026/06/ai-data-center-boom-faces-public-backlash-as-77-fear-higher-electricity-bills/
[13] Quartz (citing Gallup and Data Center Watch), “AI data center backlash grows as electricity costs rise,” May 19, 2026. https://qz.com/ai-data-center-backlash-electricity-costs-public-opinion-051926
[14] Telecoms.com (citing Data Center Watch), “The AI data center backlash is costing Big Tech billions,” August 2026. https://www.telecoms.com/ai/the-ai-data-center-backlash-is-costing-big-tech-billions
[15] Atlantic Council, EnergySource, “Backlash against data centers could cost the US its AI edge,” July 2026. https://www.atlanticcouncil.org/blogs/energysource/backlash-against-data-centers-could-cost-the-us-its-ai-edge/
[16] Credit and Collection News (citing S&P Global Market Intelligence), “Lenders scrutinize US data center financing as community opposition builds,” August 2026. https://www.creditandcollectionnews.com/lenders-scrutinize-us-data-center-financing-as-community-opposition-builds/
[17] Tom’s Hardware (citing Bridge Michigan and Forbes), “After a $16 billion Stargate AI data center was built despite being voted down, Michigan towns rush to block new buildouts,” May 7, 2026. https://www.tomshardware.com/tech-industry/michigan-towns-rush-to-block-ai-data-centers-after-16-billion-stargate-project-overrode-local-opposition
[18] The Allegheny Front, “Largest natural gas power plant in the U.S., data center planned for Homer City,” 2025. https://www.alleghenyfront.org/homer-city-pennsylvania-gas-plant-data-center-ai/
[19] AI Law Tracker, “Indiana Data Center Map — AI Infrastructure Projects” (project inventory and investment totals), 2026. https://ailawtracker.org/data-centers
[20] Statista, “Big Tech’s AI Spending to Reach $760 Billion in 2026” (Q2-2026 earnings of Microsoft, Alphabet, Meta, Amazon), August 2026. https://www.statista.com/chart/35046/capital-expenditure-of-meta-alphabet-amazon-and-microsoft/
[21] Brian Sozzi, Yahoo Finance (citing Goldman Sachs), “Meta, Microsoft, Amazon, and Alphabet are about to spend a shocking amount of money to dominate the AI era,” June 3, 2026. https://finance.yahoo.com/sectors/technology/article/meta-microsoft-amazon-and-alphabet-are-about-to-spend-a-shocking-amount-of-money-to-dominate-the-ai-era-115359575.html
[22] Farrah Anderson, WFYI Indianapolis, “Indy council advances pause on data centers, sends to commission for final approval,” August 10, 2026. https://www.wfyi.org/wfyi-news/2026-08-10/indy-council-advances-pause-on-data-centers-sends-to-commission-for-final-approval
[23] Bloomberg Tax, “Arizona Data Center Tax Incentive Pause Signed by Governor Hobbs” (also covering New York S 10642 permit moratorium), June 2026. https://news.bloombergtax.com/daily-tax-report-state/arizona-data-center-tax-incentive-pause-signed-by-governor-hobbs
[24] WFYI Indianapolis (citing Indiana University Environmental Resilience Institute; Janet McCabe), “Nearly a third of Indiana counties have moved to restrict data centers,” July 6, 2026. https://www.wfyi.org/statewide/2026-07-06/indiana-counties-data-center-moratoriums-bans-2026
[25] International Energy Agency (IEA), “Key Questions on Energy and AI” — news release: “Data centre electricity use surged in 2025, even with tightening bottlenecks driving a scramble for solutions,” 2026. https://www.iea.org/news/data-centre-electricity-use-surged-in-2025-even-with-tightening-bottlenecks-driving-a-scramble-for-solutions
[26] Cameron F. Kerry et al., The Brookings Institution, “Global energy demands within the AI regulatory landscape,” updated April 2026. https://www.brookings.edu/articles/global-energy-demands-within-the-ai-regulatory-landscape/
[27] Introl Research, “Virginia SB 253: Data Center Rate Shift Could Set National Precedent” (citing PJM Long-Term Load Forecast, JLARC, NVTC), February 2026. https://introl.com/blog/virginia-sb-253-data-center-electricity-rate-shift-2026
[28] Inside Climate News via WESA, “Pennsylvania’s governor has a plan to make data centers bring their own energy. Now comes the hard part,” June 5, 2026. https://www.wesanews.org/politics-government/2026-06-05/pennsylvania-gov-shapiro-ai-data-center-plan
[29] Introl Research, “Hyperscaler CapEx Hits $690B in 2026” (Mark Zuckerberg remarks; Microsoft power constraints; transformer lead times; Google Cloud backlog), February 2026. https://introl.com/blog/hyperscaler-capex-690-billion-microsoft-azure-power-bottleneck-2026
[30] American Action Forum, “Virginia’s New Data Center Electricity Rate Class,” April 2026. https://www.americanactionforum.org/insight/virginias-new-data-center-electricity-rate-class/
[31] Earthjustice and Southern Environmental Law Center (for the NAACP), “NAACP Sues xAI for Illegal Pollution from Data Center Power Plant,” April 14, 2026. https://earthjustice.org/press/2026/xai-sued-for-illegal-power-plant
[32] UC Riverside News, “Data center water spikes could cost billions” (Professor Shaolei Ren, UC Riverside; Professor Adam Wierman, Caltech; Yuelin Han; Professor Pengfei Li, RIT — “Small bottle, big pipe”), March 9, 2026. https://news.ucr.edu/articles/2026/03/09/data-center-water-spikes-could-cost-billions
[33] Jason Plautz and Christa Marshall, E&E News by POLITICO, “Thirsty data centers fuel local angst over water infrastructure” (quoting Professor Shaolei Ren), March 19, 2026. https://www.eenews.net/articles/thirsty-data-centers-fuel-local-angst-over-water-infrastructure/
[34] Arizona Capitol Times, “Data centers dominated 2026 session — lawmakers answered with 3-year tax incentive pause” (Chris Diorio, Data Center Coalition), July 9, 2026. https://azcapitoltimes.com/news/2026/07/09/data-centers-dominated-2026-session-lawmakers-answered-with-3-year-tax-incentive-pause/
[35] WAVE 3 News, “Jeffersonville passes year-long data center moratorium amid resident outcry and major utility disclosures,” August 4, 2026. https://www.wave3.com/2026/08/04/jeffersonville-passes-year-long-data-center-moratorium-amid-resident-outcry-major-utility-disclosures/
[36] Fortune, “A Michigan farm town voted down plans for a giant OpenAI-Oracle data center. Weeks later, construction began,” May 6, 2026. https://fortune.com/2026/05/06/ai-data-center-michigan-saline-politics-farmland/
[37] Technology.org, “xAI Ran 59 Unpermitted Gas Turbines for Colossus 2 Near Memphis” (Patrick Anderson, SELC; EPA determination), July 15, 2026. https://www.technology.org/2026/07/15/xai-59-unpermitted-gas-turbines-southaven-colossus-2/
[38] Data Center Watch, “Briefing 09/26/2025” (Google Project Flo withdrawal; Howell Township; weekly political-risk tracking), September 26, 2025. https://datacenterwatch.substack.com/p/briefing-09262025
[39] Consumer Reports (citing Ari Peskoe, Harvard Law School Electricity Law Initiative; Heatmap Pro; academic employment research), “AI Data Centers: Big Tech’s Impact on Electric Bills, Water, and More,” March 2026. https://www.consumerreports.org/data-centers/ai-data-centers-impact-on-electric-bills-water-and-more-a1040338678/
[40] MultiState Insider, “State Data Center Policy: Governors Restrict Tax Exemptions” (Texas, Arizona, Illinois, Ohio actions), June 22, 2026. https://www.multistate.us/insider/2026/6/22/state-data-center-policy-shifts-as-governors-impose-new-restrictions
[41] Commonwealth of Pennsylvania, Office of Governor Josh Shapiro, “Gov. Shapiro Releases Full GRID Standards to Protect Pennsylvanians,” May 27, 2026. https://www.pa.gov/governor/newsroom/2026-press-releases/gov-shapiro-releases-full-grid-standards-to-protect-pennsylvania
[42] The Hill (opinion, energy and environment), “Data center backlash could impact AI advancements,” April 28, 2026. https://thehill.com/opinion/energy-environment/5851140-data-center-power-grid-crisis/
[43] Pareesa Afreen, The News International, “43% say data centers are raising energy bills, Stanford prof warns” (Professor Anjney Midha, Stanford University), May 2026. https://www.thenews.com.pk/latest/1401869-43-say-data-centers-are-raising-energy-bills-stanford-prof-warns
[44] KVUE (ABC Austin), “Gov. Abbott touts data center company support for audit, pause of new data center approvals” (QTS statement), August 2026. https://www.kvue.com/article/news/local/texas/greg-abbott-data-center-audit-pause-new-approvals/269-df910bc8-9b59-40b5-b4e0-2a323a8885cb
[45] Joint Legislative Audit and Review Commission (JLARC), Commonwealth of Virginia, “Data Centers in Virginia” (with University of Virginia Weldon Cooper Center and E3 consulting analyses). https://jlarc.virginia.gov/landing-2024-data-centers-in-virginia.asp
[46] State Senator Glen Sturtevant, Cardinal News, “Before Virginia approves another data center, it should decide whether more are justified,” August 3, 2026. https://cardinalnews.org/2026/08/03/sturtevant-on-data-center-moratorium-before-virginia-approves-another-data-center-it-should-decide-whether-more-are-justified/
[47] The Piedmont Environmental Council, “The Virginia State Corporation Commission takes important first step of requiring large-load data centers to pay for transmission facilities constructed solely to serve them,” August 3, 2026. https://www.pecva.org/resources/press/press-release-the-virginia-state-corporation-commission-takes-important-first-step-of-requiring-large-load-data-centers-to-pay-for-transmission-facilities-constructed-solely-to-serve-them/
[48] Northern Virginia Magazine, “Virginia Data Centers Must Pay for New Transmission Infrastructure” (Governor Abigail Spanberger statement), August 7, 2026. https://northernvirginiamag.com/news/2026/08/07/virginia-data-centers-must-pay-for-new-transmission-infrastructure/
[49] Pennsylvania Capital-Star, “Pa.’s approach to joining the AI race must put people first, Shapiro said in budget address,” February 4, 2026. https://penncapital-star.com/campaigns-elections/pa-s-approach-to-joining-the-ai-race-must-put-people-first-shapiro-said-in-budget-address/
[50] Pennsylvania Capital-Star, “Pa. House passes data center ‘pause’ along with Shapiro’s plan for ‘responsible’ development,” June 25, 2026. https://penncapital-star.com/economy/pa-house-passes-data-center-pause-along-with-shapiros-plan-for-responsible-development/
[51] Marcus Baram, Capital & Main via Pennsylvania Capital-Star, “In Pennsylvania, the data center wars hit the ballot box,” August 6, 2026. https://penncapital-star.com/energy-environment/in-pennsylvania-the-data-center-wars-hit-the-ballot-box/
[52] Justin Sweitzer and Harrison Cann, City & State Pennsylvania, “Gov. Josh Shapiro’s new data center standards reflect new reality,” May 28, 2026. https://www.cityandstatepa.com/policy/2026/05/gov-josh-shapiros-new-data-center-standards-reflect-new-reality/413789/
[53] The Cooldown, “Arizona puts data center tax breaks on hold for 3 years after water, power backlash” (Governor Katie Hobbs statements), July 13, 2026. https://www.thecooldown.com/green-tech/arizona-data-centers-tax-breaks-hold/
[54] Server Country, “Arizona Data Center Policy” (Project Blue litigation, Tucson large-water-user code, municipal ordinances), 2026. https://servercountry.org/policy/arizona/
[55] AZPM News, “Pima County will develop moratorium for data centers,” August 13, 2026. https://news.azpm.org/p/newsheadlines/2026/8/13/230837-pima-county-staff-will-develop-moratorium-for-data-centers/
[56] The Detroit News, “The fight over AI data centers is playing out in Michigan communities,” July 16, 2026. https://www.detroitnews.com/story/news/politics/2026/07/16/the-fight-over-ai-data-centers-is-playing-out-in-michigan-communities/90941978007/
[57] Natalie Weber, Houston Public Media, “Gov. Greg Abbott pauses new data centers until ERCOT, PUCT audit energy, water usage,” August 3, 2026. https://www.houstonpublicmedia.org/articles/news/energy-environment/2026/08/03/558529/gov-greg-abbott-pauses-new-data-centers-until-ercot-puct-audit-energy-water-usage/
[58] Data Center Knowledge, “Microsoft, Alphabet, Meta Pivot from Buy to Build in AI” (Satya Nadella Q2-2026 remarks; Alphabet capex guidance of $195-205 billion), July-August 2026. https://www.datacenterknowledge.com/data-center-construction/hyperscalers-say-ai-race-has-entered-a-new-phase
[59] Northeast Times, “Wall Street now treats data center backlash as a credit risk” (Nvidia $500 billion Wall Street mobilization; Reuters summary), August 11, 2026. https://northeasttimes.com/2026/08/11/wall-street-now-treats-data-center-backlash-as-a-credit-risk/
[60] International Monetary Fund (IMF) Working Paper WP/25/81, “Power Hungry: How AI Will Drive Energy Demand” (IMF-ENV model), 2025. https://www.imf.org/-/media/Files/Publications/WP/2025/English/wpiea2025081-print-pdf.ashx
[61] Fortune (citing Electric Power Research Institute working paper and Goldman Sachs electricity-cost projections), “Data centers are actually making your electric bill cheaper — but sinking AI demand could change that,” July 26, 2026. https://fortune.com/2026/07/26/data-centers-electricity-costs-cheaper-7billion-buildout-ai-demand/
