Introduction: When the Megawatts Move Away From the Metropolis
In Middleton Township, Wood County, Ohio, about twenty-five miles south of Toledo, a woman named Breanne Kidd used to watch the sun come up over farmland while she drank her coffee and waited for the toddlers to arrive at the daycare she runs out of her home. Over the course of roughly a year, that view was replaced by cranes, steel, and dust as crews built out Meta’s eight-hundred-acre Bowling Green data center. Then something appeared that nobody had told her about: the beginnings of a large natural-gas power plant, sited to serve the data center and nothing else. She described the distance with the precision of somebody who has measured it.[4]
It’s not like we’re two streets away. We’re literally across the street. I’m living next to a threat.
— Breanne Kidd, resident, Middleton Township, Ohio [4]
The facility she was pointing at is the Apollo Generating Station — approximately 350 megawatts of natural-gas-fired generation plus roughly 120 megawatts of battery storage, designed to operate behind the meter, serving only the adjacent Meta campus and not physically connected to the public grid. The Ohio Power Siting Board approved it on 3 February 2026 without public hearings, under an expedited review process intended for generation dedicated to a single customer; plans had been submitted less than three months earlier, and the state’s draft air permit did not become publicly available until March, after construction had already started. In the paperwork, the client was listed not as Meta but as a subsidiary called Liames LLC.[4,5]
Reuters, working with data from the research firm Cleanview, found that Apollo is not an anomaly. There are at least fifty-seven off-grid power plants proposed or under construction in the United States to serve individual data centers, with a combined capacity of roughly 73,000 megawatts — enough, in aggregate, for tens of millions of homes. More than a dozen won approval in under a year, with little or no notice to residents.[4]
On 19 August 2026, Reuters reported a number that may prove more consequential to the geography of artificial intelligence than another benchmark score, another accelerator launch, or another record capital-expenditure forecast. Drawing on data shared by JLL, one of the world’s largest commercial real-estate and property-services firms, the report found that new European hyperscale data centers due to come online between 2026 and 2028 are being developed an average of approximately 109 miles — 175 kilometers — from a major hub. Facilities delivered between 2022 and 2025 averaged only about 29 miles, or 46 kilometers, from the same kinds of metropolitan centers. The new generation of European AI infrastructure will sit more than three times further from major cities than the generation that immediately preceded it.[1]
The same analysis found that greenfield sites now account for approximately 39 percent of Europe’s forward pipeline against only about 8 percent of previously delivered projects, and that the share of pipeline projects in inner-city locations is expected to fall to roughly 5 percent from about 13 percent, with the remainder in industrial or edge-of-city locations. Separately, DC Byte’s tracking of early-stage projects found that of nine proposed European developments exceeding one gigawatt, only one is planned near a major city — Paris — with the others distributed from rural Spain to northern Sweden.[1,3]
Figure 1. Three views of the same migration. European hyperscale projects are moving outward, onto undeveloped land, and away from inner-city locations simultaneously. Source: JLL data reported by Reuters, 19 August 2026.[1,3]
The explanation JLL offers is deceptively simple, and it is worth quoting precisely because it inverts thirty years of received wisdom about digital infrastructure.
The determining factor is increasingly where sufficient power can be secured, rather than simply where demand exists. Data centers are being brought to where the power is, not the other way around.
— Assad Noori, Head of Data Centres, Europe, Middle East and Africa, JLL [1]
Noori’s colleague Martin Jensen, president of JLL’s EMEA data-center division, framed the same finding as a bifurcation rather than an abandonment, and this distinction will matter throughout this paper.
Europe’s core markets will remain critical because enterprise demand isn’t going anywhere. Hyperscale AI infrastructure requires a completely different scale of power and land.
— Martin Jensen, President, EMEA Data Centres, JLL [1]
The economics reinforce the point with unusual clarity. JLL estimates that powered land costs an average of €2.36 million per megawatt of IT load in Europe’s core markets, approximately €978,000 per megawatt in secondary cities including Copenhagen, Warsaw, and Milan, and approximately €512,000 per megawatt in tertiary areas — with some locations around Bordeaux falling as low as roughly €200,000 per megawatt. Amsterdam remains the most expensive market at roughly €2.7 million per megawatt, followed by London at approximately €2.6 million and Frankfurt at approximately €2.5 million. A developer who moves from Amsterdam to a tertiary French market is not saving a marginal percentage. The developer is changing the order of magnitude of the input cost.[1]
Rupert Duckworth of Savills, describing the London market specifically, identified the mechanism that produces this spread — a mechanism that is fundamentally about competition for scarce inputs rather than about technology at all.
London has already seen significant digital infrastructure development driven by cloud and has other asset classes competing for space leading to high land prices. Power is now constrained in the key cloud locations across the market.
— Rupert Duckworth, Associate Director, EMEA Data Centre Advisory, Savills [1]
Figure 2. The price of powered land across European market tiers, expressed in euros per megawatt of IT load rather than per hectare (about 2.5 acres) — a shift in the unit of valuation that is itself part of the story. Source: JLL data reported by Reuters, 19 August 2026.[1]
Why the term, and why now
The European numbers are striking because they render statistically visible something that has already been happening physically across large parts of the United States. Meta’s Hyperion supercluster is rising in Richland Parish, Louisiana — a sparsely populated agricultural parish in the northeastern corner of the state, roughly as far from a traditional technology capital as it is possible to be within the continental United States. On 13 July 2026 Meta confirmed the expansion of that project to five gigawatts of compute capacity at a total cost exceeding fifty billion dollars, up from a $27 billion, roughly two-gigawatt plan earlier in the year and an original $10 billion announcement in December 2024.[15,16] In northern Indiana, Amazon Web Services committed approximately eleven billion dollars to a campus near New Carlisle, described at the time as the largest capital-investment announcement in the state’s history; an estimate presented to Indiana utility regulators put the site’s eventual service capability at approximately 2,250 megawatts at full development.[20,21] In Pike County, Ohio, on the grounds of a Cold War uranium-enrichment site, Nvidia agreed on 17 August 2026 to provide credit support of as much as $105 billion connected with OpenAI’s twenty-year lease of an SB Energy campus that could ultimately reach eight gigawatts of IT load, supported by at least ten gigawatts of newly built generation.[23,25]
These are not incremental additions to an existing estate. They are the industrial colonization of places that the digital economy previously treated as consumers of its services rather than as producers of its inputs. That inversion is why the term Powered Periphery seems to me both necessary and precise.
Periphery has a settled meaning in economic geography. It describes places outside the economic center: rural counties, former industrial districts, agricultural regions, secondary cities, mining territories, inexpensive land along transmission corridors, municipalities distant from the headquarters, universities, venture-capital networks, and highly paid technical workforces that define the digital economy’s core. For the whole of the internet era, these places consumed applications and cloud services whose highest-value intellectual and corporate functions were located elsewhere. Artificial intelligence complicates that hierarchy because the physical infrastructure required to manufacture intelligence at extraordinary scale needs something that London, Frankfurt, Amsterdam, New York, San Francisco, and many other mature metropolitan markets increasingly struggle to supply in very large increments: available power, at the right location, on the right timetable, across enough contiguous land to permit continued expansion.
Powered is the operative qualifier, and I use it deliberately. The emerging periphery is not valuable simply because its acreage is inexpensive. Cheap acreage without electricity can remain cheap indefinitely; there is a great deal of inexpensive land in the world and almost none of it is being bid for by hyperscalers. The valuable property is powered land — land connected to, adjacent to, or capable of developing generation, transmission, substations, water systems, fiber routes, and the political permissions necessary to operate an AI factory. Artificial intelligence can therefore reverse parts of the old real-estate hierarchy. A rural tract beside a major transmission corridor, a natural-gas pipeline, a nuclear plant, a hydroelectric system, or a large renewable resource may be strategically more valuable to an AI infrastructure investor than far more expensive commercial land near a global metropolis.
The Powered Periphery is the emerging geography in which regions outside traditional technology and metropolitan centers acquire disproportionate strategic value because they possess the energy, land, water, connectivity, expansion capacity, and political durability required for industrial-scale artificial intelligence.
The phrase does not mean that artificial intelligence is abandoning cities, and any reading of this paper that reaches that conclusion has misread it. Latency-sensitive inference, enterprise connectivity, financial applications, cloud interconnection, regulatory proximity, and large populations will keep substantial computing infrastructure close to metropolitan areas. The European industry’s own analysis makes precisely this distinction: AI training increasingly favors regions with abundant electricity, including Nordic and remote southern European locations, while inference continues to place pressure on metro-proximate facilities.[2,66] Live capacity across the FLAP-D markets has reached approximately 3.8 gigawatts, more than doubling since 2019, with a further 1.4 gigawatts under construction and roughly 2 gigawatts in the planned pipeline.[2] The core is not shrinking. It is simply no longer where the largest new increments go.
The more precise argument is that the marginal gigawatt of AI capacity is becoming geographically different from the marginal megawatt of the cloud era. Training clusters, frontier-model infrastructure, large-scale batch inference, and increasingly autonomous computational systems can tolerate geographic distance far more easily than they can tolerate unavailable electricity.
That shift has consequences well beyond data-center real estate. It changes the value of farmland. It changes the bargaining power of utilities. It changes where transmission lines are built and who pays for them. It changes county tax bases. It changes demand for construction workers, electricians, and power engineers. It changes water politics. It changes the relationship between governors and hyperscalers. It changes the political meaning of nuclear plants, gas fields, renewable-energy regions, and industrial brownfields. And ultimately it begins to redraw the physical geography of the Five-Layer AI Economy itself.
Section 1 — From Cloud Geography to Power Geography
1.1 The metropolitan logic of the cloud era
To understand why the geography is inverting, it is necessary first to be honest about how thoroughly the previous geography was determined by a small number of engineering constraints that no longer bind in the same way.
The early geography of cloud computing was overwhelmingly a geography of proximity. Large concentrations of enterprises, internet exchanges, telecommunications infrastructure, and end users favored major metropolitan areas, and the infrastructure followed. Northern Virginia became one of the world’s largest data-center concentrations partly because of extraordinary network connectivity and partly because of its proximity to government and commercial customers. Frankfurt, London, Amsterdam, Paris, and Dublin formed the FLAP-D geography that has organized European data-center investment for two decades. Silicon Valley and the Pacific Northwest accumulated their own ecosystems of fiber, power, and talent.
In that system, distance imposed a genuine economic penalty. It imposed latency on interactive workloads. It imposed connectivity complexity and cost. It imposed friction on enterprise procurement, on physical maintenance, on sales relationships, and on the informal institutional trust that governs where large firms are willing to place their production systems. The result was a self-reinforcing agglomeration: talent attracted capital, capital attracted companies, companies attracted infrastructure, and infrastructure reinforced the city.
Artificial intelligence changes the weighting of every variable in that equation. Training a frontier model is not equivalent to serving an interactive web request. A cluster performing weeks or months of densely parallelized computation does not need to sit a dozen miles from millions of consumers. Its critical requirements are accelerators, an enormous and stable electrical supply, cooling infrastructure adequate to remove the resulting heat, very high-speed internal networking, sufficient external fiber connectivity to move checkpoints and datasets, physical security, and — perhaps above all — the ability to expand on the same site for a decade or more without renegotiating the entire premise of the project.
As those requirements grow from tens of megawatts toward hundreds of megawatts and then gigawatts, the geography begins to invert. The relevant question ceases to be where are the users? and becomes where can we actually energize the machines?
1.2 Training separates compute from population
Artificial intelligence therefore introduces an important and historically unusual separation between the place where intelligence is manufactured and the place where intelligence is consumed. A user in Paris may interact with an application whose inference is processed nearby while the underlying frontier model was trained six hundred miles away in Scandinavia. A company in London may build an agentic workflow on models trained in Iberia. A corporation in New York may consume intelligence manufactured by accelerator clusters operating in rural Ohio or northeastern Louisiana.
This resembles earlier industrial transformations far more than it resembles conventional software geography. Steel production does not happen where steel is consumed. Petroleum refining does not occur beside every filling station. Semiconductor fabrication does not occur beside every smartphone owner. Aluminium smelting migrated historically toward stranded hydroelectric power in Iceland, Quebec, and the Pacific Northwest for exactly the reason that AI training is now migrating toward stranded and developable electrical capacity: the input is heavy, continuous, and expensive to transport, while the output is light, valuable, and cheap to transport.
Once computation becomes sufficiently industrial, the same logic applies with the same force. AI training increasingly resembles an industrial production process in which electricity, silicon, cooling, and physical infrastructure are converted into a new economic output — machine intelligence — that can then be distributed globally at negligible marginal transport cost. The scale is no longer metaphorical. Stanford’s 2026 AI Index reports that the United States now hosts more than 5,400 data centers, over ten times the count of any other country, and that AI data-center power capacity has reached approximately 29.6 gigawatts — comparable to the peak electricity demand of New York State.[11,12]
1.3 The Marginal Gigawatt Principle
This leads to what I will call the Marginal Gigawatt Principle, and it is the analytical core of the paper’s first half.
Existing metropolitan data centers will continue to operate. Cities will remain critical to cloud computing and to inference. The question that determines the new geography is not where capacity already sits but where the next enormous block of capacity gets built. And the answer to that question changes discontinuously as the size of the block increases.
Table 1 — The Marginal Gigawatt Principle: how the siting question changes with scale
| Increment of new capacity | What the developer is actually selecting | Binding constraint |
| ≈ 10–50 MW | A parcel, or space within an existing campus | Land availability and fit-out capital |
| ≈ 100–500 MW | A utility service territory and a substation position | Interconnection queue and transformer supply |
| ≈ 1–2 GW | A regional transmission system and a state regulatory posture | Generation adequacy and transmission headroom |
| ≈ 5–8 GW | An industrial region, and often a bespoke generation program | Political durability, fuel access, water, and community consent |
Table 1. As the increment grows, the unit of site selection escalates from parcel to region and the binding constraint migrates from real estate to politics. Author’s framework.
If a hyperscaler requires another fifty megawatts, an established metropolitan campus may still accommodate it, and the traditional geography survives intact. If it requires another five hundred megawatts, the geography becomes difficult, and the developer begins consulting interconnection queues rather than brokers. If it requires another two gigawatts, the geography becomes strategic, and the conversation moves from a leasing team to a governor’s office. If it requires five or eight gigawatts, the developer is no longer selecting a data-center parcel at all. It is selecting an industrial region — and, increasingly, commissioning the electricity system that region will need.
Europe’s changing site distances are therefore not an incidental real-estate trend to be filed alongside vacancy rates and rental yields. They are evidence that AI’s physical scale has exceeded the assumptions on which the previous geography of cloud infrastructure was constructed. The 29-to-109-mile shift is what it looks like, in aggregate, when a large number of independent developers each solve the same constrained optimization problem and each arrive at the same conclusion.
1.4 Reading the 109-mile signal properly
A movement from an average of 29 miles to an average of 109 miles deserves interpretation rather than mere citation, because it means that infrastructure is crossing several distinct boundaries simultaneously, and each crossing carries its own economics.
It leaves the normal commuting geography of the city, which changes the labor market for construction and operations and forces developers to think seriously about workforce housing, transport, and training pipelines rather than assuming an available metropolitan labor pool. It frequently crosses utility service territories, which changes the counterparty, the tariff, the interconnection process, and the regulator. It encounters different municipalities and planning authorities, which changes the political process from one dominated by professional planning departments to one in which a county commission of five people may hold decisive authority over billions of dollars of capital. It reaches cheaper land, which changes the capital structure of the project. It reaches regions where larger contiguous parcels can be assembled, which changes the expansion option value embedded in the site. And, critically, it may reach electricity systems with unused or developable capacity that is simply unavailable in the traditional core.
The geography of compute, in short, begins to follow the geography of energy — and the geography of energy has never been the geography of population. Coal was where coal was. Hydro was where the fall line was. Gas is where the basins and the pipelines are. Nuclear is where mid-century siting decisions and cooling water put it. Wind is where the wind is. Solar is where the irradiance and the cheap land coincide. None of those maps look like a map of where people live, and the AI map is beginning to look like them rather than like the cloud map.
1.5 Rural does not mean disconnected
The Powered Periphery should not be mistaken for digital isolation, and this misunderstanding is common enough to be worth addressing directly.
A modern AI campus can be geographically remote while remaining computationally central. Fiber collapses informational distance even where physical distance persists. A training cluster in northern Sweden participates in the global frontier of machine intelligence as fully as one in Santa Clara; the difference is that it does so at a lower cost of electricity and with a longer expansion runway. What is remote is the building, not the system. The distinction matters because much of the political and journalistic discussion of rural data centers implicitly assumes that peripheral location implies peripheral importance, and the opposite is closer to the truth.
The new location equation can be summarized compactly, and I will return to it in the closing section:
The site-value equation
AI Site Value = Power Certainty + Land Expandability + Fiber Reach + Cooling and Water Capacity + Political Durability — Note what is absent: traditional real-estate prestige, proximity to a central business district, and, at the frontier, latency to end users.
That absence is an extraordinary change. A rural county without a major airport, a research university, or a venture-capital community could nevertheless become one of the most important computational locations on earth if several gigawatts of advanced accelerators operate there for twenty years. Nothing in the traditional apparatus of regional economic development anticipated this, which is part of why the policy response has been so uneven.
1.6 The counter-forces: why the metropolis does not simply lose
Intellectual honesty requires acknowledging the forces pulling in the opposite direction, because the migration described in this paper is real but bounded.
First, inference is growing faster than training as a share of total workload, and inference is far more latency-sensitive. As AI moves from research artifact to embedded product — into search, into productivity software, into customer service, into agents that call tools in real time — the serving footprint must move closer to users. The industry’s own forecasts anticipate that AI workloads could account for around half of global data-center capacity by 2030, but that share includes both the training clusters that migrate outward and the inference estate that does not.[1,2]
Second, enterprise demand is sticky and regulated. Financial institutions, healthcare systems, and government agencies have data-residency requirements, audit expectations, and existing interconnection relationships that are expensive to relocate. The European Commission’s own AI infrastructure program illustrates the point: Iceland, despite an ideal cooling climate and abundant renewable power, faces limits as a destination for certain EU workloads because data-residency rules favor locations inside the Union’s territory. Digital sovereignty is now a siting variable that can outrank thermodynamics.
Third, the periphery’s advantage is contingent, not permanent. If a rural region’s transmission headroom is consumed by the first two gigawatts, the third gigawatt faces the same queue that drove developers out of Frankfurt. The migration relocates the constraint; it does not dissolve it. Several European analysts have already observed that the shift does not resolve the physical constraints around traditional hubs so much as move some of them to new regions.[3,65]
These qualifications matter. The claim of this paper is not that geography has been abolished or that the metropolis is in decline. It is that a second geography has emerged alongside the first, organized by a different logic, and that this second geography is where the largest increments of frontier compute are now going.
Section 2 — The European Signal: Documenting the Migration
Europe deserves a section of its own, not because the phenomenon originated there — it did not — but because Europe is where it has first been measured cleanly. The American build-out has been faster, larger, and more chaotic; the European build-out has been slower, more constrained, and consequently more legible. When a system is constrained, its priorities become visible.
2.1 Why Europe’s constraints made the pattern measurable
The FLAP-D markets did not become congested by accident. They became congested because they are simultaneously the most connected, the most regulated, the most land-constrained, and the most electrically mature data-center markets on the continent. Frankfurt, London, Amsterdam, Paris, and Dublin remain the largest and continue to be in demand, but they increasingly face land shortages, planning restrictions, and lengthy waits for grid connections.[1] Each of those three constraints is independently sufficient to slow a project. Together they are close to disqualifying for a facility that needs several hundred megawatts on a three-year timetable.
The consequence is that European developers were forced to articulate, explicitly and early, a trade-off that American developers could for several years resolve implicitly by simply building somewhere large and empty. The European decision had to be defended to planning authorities, to transmission system operators, and to boards. That defense generated documentation, and the documentation is what JLL and DC Byte have now aggregated into the numbers that opened this paper.
The scale of the capital involved makes the constraint binding rather than theoretical. JLL estimates that the world’s four largest hyperscale cloud providers will spend approximately $725 billion in 2026, up roughly 77 percent from about $410 billion in 2025, the overwhelming majority of it on AI computing and data-center infrastructure.[1] By the close of the second-quarter 2026 reporting season, after Alphabet, Microsoft, Amazon, and Meta had each revised guidance upward, the combined 2026 figure for those four companies was tracking toward approximately $760 billion against roughly $413 billion in 2025.[13] Capital of that magnitude does not wait politely in an interconnection queue.
Figure 3. Combined capital expenditure of the four largest U.S. hyperscalers, 2020 through 2026 guidance. The 2026 figure reflects upward revisions announced during the second-quarter 2026 reporting season. Sources: company earnings disclosures compiled by Statista and the Financial Times.[13,14]
2.2 The price of powered land, and what it reveals
The most analytically interesting feature of the JLL dataset is not the distance figure at all. It is the unit of measurement. JLL prices European land not in euros per hectare but in euros per megawatt of IT load. That is a quiet but profound change in how a class of real estate is understood, and it deserves to be treated as a finding in its own right rather than as a convenient normalization.
When land is priced by area — per hectare in Europe, per acre in the United States — the buyer is purchasing ground, and the relevant comparables are other parcels of similar area, zoning, and location. When land is priced per megawatt of IT load, the buyer is purchasing the right and the ability to consume electricity at a specific place, and the relevant comparables are entirely different: interconnection agreements, substation positions, generation queues, and utility tariffs. Two adjacent parcels of identical size and identical agricultural quality can differ in price by an order of magnitude under the second convention and not at all under the first.
The observed spread is dramatic. Core markets average €2.36 million per megawatt. Secondary markets — Copenhagen, Warsaw, Milan — average approximately €978,000. Tertiary markets average approximately €512,000, and specific locations around Bordeaux can fall to approximately €200,000.[1] The ratio between the most expensive core market and the cheapest tertiary location approaches thirteen to one. For a one-gigawatt campus, that spread is the difference between roughly €2.7 billion and roughly €200 million in powered-land cost alone — a difference large enough to fund a substantial fraction of the generation the site will need.
Table 2 — Powered land in Europe: cost per megawatt of IT load by market tier, 2026
| Market tier | Representative markets | Cost per MW of IT load | Implied cost, 1 GW campus |
| Core — most expensive | Amsterdam | ≈ €2.70 million | ≈ €2.70 billion |
| Core | London | ≈ €2.60 million | ≈ €2.60 billion |
| Core | Frankfurt | ≈ €2.50 million | ≈ €2.50 billion |
| Core — tier average | FLAP-D average | ≈ €2.36 million | ≈ €2.36 billion |
| Secondary | Copenhagen, Warsaw, Milan | ≈ €978,000 | ≈ €978 million |
| Tertiary | e.g. Bordeaux region | ≈ €512,000 | ≈ €512 million |
| Tertiary — low end | Specific Bordeaux locations | ≈ €200,000 | ≈ €200 million |
Table 2. Source: JLL estimates reported by Reuters, 19 August 2026. The final column is the author’s arithmetic extension and is illustrative rather than a market quotation, since very large sites are not priced by simple linear extrapolation.[1]
2.3 Where Europe’s gigawatt projects are actually going
The DC Byte finding — that of nine proposed European developments above one gigawatt, only one is planned near a major city — is worth unpacking geographically, because the destinations are not random. They cluster around four distinct energy endowments.
- Nordic hydro and cold-climate cooling. Norway, Sweden, and Finland offer abundant hydroelectric and nuclear generation, low ambient temperatures that reduce cooling load, and, in several cases, transmission capacity built decades ago for heavy industry that has since contracted. Stargate Norway in Narvik — a facility backed by OpenAI, Nscale, and Aker, targeting approximately 230 megawatts rising toward 520 megawatts, running on Norwegian hydropower with direct-to-chip liquid cooling and waste-heat recovery to local industry — is the archetype.[62]
- Iberian solar and available grid headroom. Spain and Portugal combine very large solar resources, comparatively fast permitting in certain autonomous communities, and grid capacity that the FLAP-D markets lack. Aragón in particular has absorbed multi-billion-euro commitments precisely because it had available grid capacity when the traditional queue did not.[65] In Portugal, the Start Campus site at Sines is permitted for up to 1.2 gigawatts.[62]
- Nuclear-adjacent French and Central European locations, where existing baseload generation and heavy-industrial transmission make large loads technically tractable without waiting for new build.
- Central and Eastern European secondary markets, where land, labor, and power costs are materially lower and EU data-residency requirements are satisfied.
One market analysis projects that Europe’s secondary markets — the Nordics, Iberia, and Central and Eastern Europe — will grow by approximately 110 percent between 2024 and 2030 against roughly 55 percent for the traditional hubs, and that by 2035 half of European data-center capacity will sit outside those hubs.[65] Whether that specific ratio proves accurate is less important than the direction, which is corroborated independently by JLL’s pipeline data, by DC Byte’s project tracking, and by the disclosed siting decisions of the hyperscalers themselves.
2.4 The connection queue as the real constraint
Underneath every one of these siting decisions lies a single operational variable that rarely appears in headlines: the grid connection timeline. In February 2026, Amazon Web Services’ head of energy markets and regulation for EMEA stated the problem in terms unusually blunt for a regulated-industry counterparty.
There’s a misalignment. We want to expand and grow within two years. These delays are challenging our growth aspirations.
— Head of Energy Markets and Regulation, EMEA, Amazon Web Services [65]
A two-year corporate planning horizon colliding with a five-to-ten-year interconnection process is not a negotiation. It is a redirection. The developer does not wait; the developer moves. And when enough developers move for the same reason at the same time, the aggregate result is an 80-mile shift in the mean distance from a major hub.
The systemic implications have begun to worry the institutions responsible for grid stability. On 30 April 2026, ENTSO-E — the body coordinating Europe’s transmission system operators — published a warning that reframed the concern entirely. The problem was no longer that data centers could not obtain power. It was that data centers might absorb so much of Europe’s available generation headroom that grid operators would be forced to curtail renewable penetration across the continent.[65] That is a different order of policy problem, and it is one that no amount of peripheral siting solves by itself.
2.5 Europe’s institutional answer: gigafactories and sovereign compute
Europe’s policy response has been to attempt to convert the constraint into an industrial strategy. In January 2026 the EU Council amended the regulation governing the European High-Performance Computing Joint Undertaking to extend its mandate to the development and operation of AI gigafactories — facilities defined in the amended regulation as state-of-the-art large-scale installations capable of handling the complete lifecycle of very large AI models, and equipped with on the order of 100,000 advanced AI processors each, roughly four times the capacity of the current generation of AI factories.[63,64]
The program sits inside the InvestAI initiative and the broader AI Continent Action Plan, and it is capitalized through a public-private structure: approximately €10 billion in public funding drawn from EU budgets and national contributions, targeting roughly €20 billion in private investment, against a headline ambition to mobilize €200 billion in total AI investment across the bloc. Public money is intended to cover early-stage costs — land, planning, grid connections, initial infrastructure — with private capital entering at a later phase, at a target ratio of approximately one euro of public funding to every two euros of private capital.[64]
Note what the public money is explicitly designated for. It is not designated for chips or for models. It is designated for land, planning, grid connections, and initial infrastructure — which is to say, for manufacturing powered land where the market has not produced it. That is as clear an institutional confirmation of this paper’s thesis as one could reasonably ask for: the European Union has concluded that the binding constraint on its AI ambitions is the availability of energized sites, and has structured a €30 billion program around relieving it.
Meanwhile the private sector has not waited. Thirteen smaller AI factories were selected across seven European countries in 2024 and 2025, entering ramp-up through 2026, while commercial gigafactory-scale projects — Stargate Norway, the Sines campus in Portugal, Nebius’s 310-megawatt Finnish facility and 240-megawatt site near Lille — proceed on private capital and private siting logic.[62,63,67] The competitive pressure runs in both directions: EU-funded centers now compete for the same land, the same grid connections, and the same turbine and transformer supply chains as the hyperscalers they were designed to counterbalance.
One structural disadvantage bears mention because it shapes the entire European calculation. European industry pays roughly double the electricity rate of its United States counterparts, according to data published by the EU Agency for the Cooperation of Energy Regulators in 2026.[67] Electricity accounts for only a single-digit percentage of the total cost of a frontier training run, but when training runs cost hundreds of millions of euros, every percentage point compounds, and the compounding is one of the reasons that European AI infrastructure has migrated so decisively toward the continent’s cheapest-power regions rather than merely toward its emptiest ones.
Section 3 — The New Economics of Rural Land
If the first half of this argument concerns where compute goes, the second concerns what happens to the places it goes to. And the first thing that happens is that the land itself changes meaning — not physically, but in the way it is measured, priced, financed, taxed, and fought over. This section is about that change of meaning, because almost everything else in the Powered Periphery follows from it.
3.1 From acres to megawatts
Agricultural property has traditionally been evaluated through a stable and well-understood set of attributes: acreage, soil quality and crop productivity, water access, drainage, development rights, and proximity to grain elevators, processors, and transport. Industrial property adds roads, rail, labor sheds, and utility service. Commercial property adds foot traffic, demographics, and visibility. Each of these valuation systems is mature, has an established appraisal profession behind it, and produces comparables that a lender will accept.
AI infrastructure introduces a valuation metric that none of those systems was built to handle: how many megawatts can this land support, and how quickly can those megawatts arrive?
This is exactly why the JLL analysis measured European powered land in cost per megawatt of IT load rather than cost per hectare. Electricity access has become embedded in property value to such a degree that it dominates every other attribute. A parcel worth relatively little as farmland can command radically different strategic value once it possesses a credible, documented, time-bounded path to hundreds of megawatts of electrical service. Conversely, a parcel with superb agricultural characteristics and no realistic interconnection prospect is, to an AI developer, worth nothing at all.
The American market has begun to register this in prices that make no sense under the older convention. The American Society of Farm Managers and Rural Appraisers has observed that farmland values in parts of the Southeast are becoming increasingly disconnected from the farm economy as land transitions to other uses, with some properties reaching between $250,000 and $1 million per acre depending on the infrastructure and resources available with the land.[61] That final clause — depending on the infrastructure and resources available with the land — is the whole thesis compressed into ten words. The dirt is not what is being bought.
3.2 The land–power conversion and energization probability
The Powered Periphery therefore creates what might be called a land–power conversion: a mechanism by which land acquires value not from its own productive characteristics but from its position inside an infrastructure system.
Consider two physically similar rural properties. One sits near high-capacity transmission, close to natural-gas infrastructure and long-haul fiber, inside a jurisdiction with a demonstrated willingness to permit large industrial development quickly. The other sits forty miles from adequate transmission and would require years of grid reinforcement, a new substation, and a contested route across neighboring properties. To a farmer, and to a conventional appraiser, these two properties may have nearly identical characteristics. To an AI developer they are not merely different in degree; they are different in kind. The first is an asset. The second is an option that will probably never be exercised.
What distinguishes them is something invisible in conventional real-estate analysis, which I will call energization probability: the joint likelihood that a given parcel can obtain a specified quantity of firm electrical service within a specified time, at a specified cost, with the necessary regulatory and community permissions intact. Energization probability is not a physical property of land. It is a compound of engineering, regulation, politics, and supply chain — and it is increasingly the dominant determinant of AI infrastructure value.
Key idea
Energization probability — the joint likelihood that a parcel can obtain a specified quantity of firm electrical service, within a specified time, at a specified cost, with regulatory and community permissions intact. It is not a physical attribute of land, yet it now dominates the valuation of land at the AI frontier.
Four observations follow from treating energization probability as the operative variable. First, it explains why developers will pay extraordinary premiums for brownfield industrial sites with legacy interconnections — the Portsmouth Gaseous Diffusion Plant in Ohio being the most spectacular current example — even when those sites carry environmental liabilities. Second, it explains why behind-the-meter generation has become so attractive: a developer who builds its own power converts an uncertain, externally controlled variable into a capital-expenditure line item under its own control. Third, it explains why community consent has migrated from a public-relations concern to a financial one, since a project that cannot obtain local approval has an energization probability approaching zero regardless of its engineering merits. Fourth, it explains why the same parcel can be worth $2,000 an acre on Monday and $60,000 an acre on Friday: nothing about the soil changed, but the probability distribution did.
3.3 The gigawatt campus as a new kind of industrial estate
The projects that result from this logic are not ordinary data centers, and continuing to call them data centers has begun to obscure more than it reveals.
Meta’s Richland Parish campus illustrates the magnitude. The site now spans thousands of acres — reported at roughly 3,200 to 4,000 acres depending on the measurement date and on how adjacent acquisitions are counted — with approximately four million square feet of building and a confirmed five-gigawatt build-out costing more than fifty billion dollars.[15,17,18] The expanded project includes more than one billion dollars in local infrastructure improvements covering roads, water, and wastewater systems. Since breaking ground in December 2024, local Louisiana businesses have received more than $1.6 billion in contracts.[16]
An installation on this scale does not function like a building. It functions like an industrial estate, and it requires the full apparatus of one: substations, transmission expansion, dedicated generation, road construction, water and wastewater infrastructure, construction staging areas, equipment warehousing, worker accommodation, security, telecommunications, expanded emergency services, and a local supplier ecosystem deep enough to service continuous operations for decades. A gigawatt AI campus therefore imposes a new economic geography onto a county that may previously have been organized around agriculture, manufacturing, or resource extraction — and it does so on a construction timetable that gives local institutions very little time to adapt.
The scale of the construction itself is worth pausing on. Economists at ConstructConnect reported that United States data-center construction spending stood at $58.1 billion year-to-date through May 2026 — more than four times the previous record.[17] This is no longer a segment of the non-residential construction market. In several regions it is the market.
3.4 Land appreciation, and the unequal distribution of gains
The economic benefits will not distribute themselves evenly, and the pattern of unevenness is predictable enough that it can be planned for.
Owners of strategically positioned parcels may receive extraordinary premiums. Construction firms may experience years of demand at rates they have never previously commanded. Counties may gain tax revenue on a scale that transforms their fiscal position — the New York Times has reported that sales-tax collections in Richland Parish surged after construction began and that some teachers in the district received bonuses of $50,000 or more.[16] Utilities may substantially expand their rate base. Restaurants, hotels, and service businesses may benefit during the construction phase. Schools and public infrastructure may receive new funding streams. Meta has committed $5 million to Louisiana Delta Community College to fund scholarships training local residents for data-center work, and all graduates of Richland Parish high schools beginning with the class of 2026 are eligible for full scholarships toward data-center-related trade certificates or courses.[16]
Yet permanent data-center employment is characteristically modest relative to the capital intensity of the facility. Louisiana expects roughly 1,000 operational jobs from a project involving more than $50 billion of investment. The Ohio PORTS-Pike campus is projected to create approximately 35,000 construction jobs through 2032 and approximately 2,500 long-term operating positions.[25,26] These are real jobs and in many cases very good ones, but the ratio of capital to permanent employment is unlike anything in the history of industrial development.
This creates an unusual political economy: enormous capital investment without proportionately enormous permanent employment. The Powered Periphery can therefore become wealthy in infrastructure without becoming equally wealthy in household income. That distinction is not merely an economic observation. It is the fault line along which the politics of the next decade will run, and Section 7 examines what the rigorous evidence actually says about it.
3.5 What the land transaction looks like from the farm
Aggregate statistics conceal the texture of these transactions, and the texture matters because it explains the political response.
In Cumberland County, Pennsylvania, Mervin Raudabaugh, an eighty-five-year-old lifetime grower, was offered $15 million for 261 acres — roughly $60,000 per acre — by data-center developers. He turned it down in 2026, accepting approximately $2 million instead by selling to the Lancaster Farmland Trust to ensure preservation.[59] In Mason County, Kentucky, Ida Huddlestone and Delsia Bare, a mother and daughter raising cattle on 1,200 acres held in the family since the Civil War, were offered $26 million for 600 acres and rejected it.[59] In Kentucky, a diversified farmer named Tim Grosser turned down a $10 million offer for his 250-acre farm — five times what he had paid for it thirty years earlier — after being told he would need to sign a nondisclosure agreement simply to learn who was buying.[60]
I said, ‘We ain’t signing any nondisclosures and that was the end of it.’ … I spent half my life working that farm, and now my son and I do it together. He built a house on it.
— Tim Grosser, farmer, Kentucky [60]
Raudabaugh’s assessment of the trajectory was blunter still, and it captures a sentiment that appears repeatedly in the rural record of 2026.
Only the land that’s preserved here is going to be here. The rest of it, every square inch is going to get built on. The American farm family is definitely in trouble.
— Mervin Raudabaugh, farmer, Cumberland County, Pennsylvania [59]
Two structural features of these transactions deserve analytical attention. The first is the nondisclosure agreement, which converts what is ordinarily a public land-use process into a private one and which — as Section 6 shows — has now become the specific target of state executive action. The second is the tax treatment. Under section 1031 of the Internal Revenue Code, a landowner may defer capital-gains tax by exchanging into like-kind property, which means a farmer who sells to a data-center developer can reinvest in farmland elsewhere in the state or country. The American Farm Bureau Federation has noted that this mechanism can transmit data-center land premiums into farmland markets far from any data center, bidding up prices for farmers who never had any contact with the industry at all.[58]
The macro context is that American farmland is under conversion pressure from many directions simultaneously. The American Farmland Trust estimates the loss of approximately 2,000 acres of farmland each day to non-agricultural uses, and the USDA records a decline of roughly 75 million acres between 1997 and 2022. Farm electricity expenditures, meanwhile, are forecast by USDA’s Economic Research Service to rise 48 percent — approximately $2.8 billion — from $5.75 billion in 2019 to $8.5 billion in 2026.[58] Data centers are not the cause of that trend. But they are arriving into it, and they are arriving as the most conspicuous and best-capitalized buyer in the room.
3.6 The Rural AI Bargain
Communities consequently need to negotiate what I will call the Rural AI Bargain — and the quality of that bargain, far more than the size of the announced investment, determines whether the Powered Periphery becomes rural revitalization or rural extraction.
The instinct of most local officials, understandably, is to ask a single question: how many jobs will the data center create? That is the wrong first question, because as Section 7 will show, the honest answer is ‘fewer than promised, though more than critics claim,’ and because employment is the dimension on which this class of investment is structurally weakest. The right questions concern infrastructure, cost allocation, risk, and reversibility. The following framework is offered as a practical instrument rather than as a rhetorical list.
Table 3 — The Rural AI Bargain: a negotiating framework for host communities
| Domain | The question to ask | What a strong answer looks like |
| Electricity cost allocation | Who pays for generation, transmission, and substation build-out, and what happens if forecast load never materializes? | A large-load tariff with minimum demand commitments, financial assurances, and exit provisions that protect existing ratepayers |
| Residential rate impact | What is the modeled effect on residential and small-commercial bills over ten and twenty years? | Independently reviewed modeling filed publicly with the utility commission, not a developer summary |
| Water | What is the peak daily water demand, not the annual average, and from what source? | Peak-demand disclosure, a named source, and developer contribution to capacity that exceeds its own use |
| Fiscal | What tax revenue is guaranteed versus abated, and over what term? | A payment-in-lieu floor that does not fall to zero if the abatement schedule is renegotiated |
| Employment | How many permanent positions, at what wage, with what local hiring and training commitment? | A community benefits agreement with enforceable hiring targets and funded training pipelines |
| Local supply chain | Can local contractors and manufacturers compete for work, and how is that measured? | Published local-content reporting during construction and operations |
| Transparency | Are nondisclosure agreements being used with the county or the state? | No NDAs; a public permitting tracker; identified rather than shell-company applicants |
| Reversibility | What happens if the project is canceled mid-build, and who is responsible for decommissioning? | Bonded decommissioning obligations and staged infrastructure triggered by verified milestones |
| Land | Which land is being converted, and is prime agricultural soil being steered away from? | Zoning overlays that direct development to brownfields and marginal land |
Table 3. Author’s framework, synthesized from the regulatory filings, community-benefit literature, and state executive actions reviewed for this paper.[32,37,43,45,47]
This is not an anti-development framework. It is a pro-durability framework. Every item on it is a term that a well-capitalized counterparty can meet, and several of the largest projects reviewed for this paper meet many of them already. What separates a strong bargain from a weak one is not the developer’s willingness but the community’s preparation — and preparation is precisely what a county receives least of when a project arrives under an NDA on a ninety-day timetable.
Section 4 — The American Powered Periphery Is Already Here
Europe measured the migration. The United States is living it. Four cases — in Louisiana, Indiana, Ohio, and Pennsylvania — demonstrate that the Powered Periphery is not a single development model but a family of them, each shaped by the particular energy endowment, regulatory culture, and industrial history of the place. Read together, they show the mechanism operating under four quite different sets of conditions, which is the strongest available evidence that the mechanism is real.
4.1 Louisiana: from agricultural parish to frontier training geography
Richland Parish may become the clearest single example of how artificial intelligence can reorder geographic importance. Northeastern Louisiana does not resemble Silicon Valley, Seattle, or Northern Virginia in any respect that traditional economic-development theory would consider relevant. It has no major research university, no venture ecosystem, no dense professional labor market, and no history in the digital economy. What it has is land, water, a cooperative state government, and — critically — a utility willing and able to build generation at extraordinary speed.
Meta broke ground on Hyperion in December 2024 with an initial budget of roughly $10 billion for a facility of approximately four million square feet. In October 2025, a joint venture with Blue Owl Capital valued the buildings and infrastructure at roughly $27 billion for a 2.06-gigawatt facility. By February 2026 Meta had acquired approximately 1,400 additional adjacent acres, signaling a second phase. On 13 July 2026 the company confirmed the expansion to five gigawatts and total investment exceeding $50 billion, with the first two gigawatts expected online by 2030 and the full five gigawatts around 2032.[15,16,18] In eighteen months the project’s announced cost multiplied roughly fivefold.
The energy program surrounding it is as significant as the campus itself, and arguably more so. Entergy Louisiana is building natural-gas-fired generation delivering more than seven gigawatts of capacity — a build-out representing more than a thirty percent increase in Louisiana’s entire grid capacity. Meta is funding up to 2.5 gigawatts of renewable generation, roughly 240 miles of new transmission, grid-scale battery storage, and nuclear plant uprates. The company states that it is paying the full costs of the energy, water, and related infrastructure, and Entergy projects more than $2 billion in customer savings over twenty years.[18] The first tranche of gas plants received expedited approval from the Louisiana Public Service Commission in August 2025, on a timetable far shorter than the several years of planning and hearings such projects typically require.[19]
The conceptual significance exceeds the dollar value. For decades, regions attempting to enter the knowledge economy pursued a well-worn strategy: build a research park, recruit a university partnership, cultivate a software cluster, and hope for spillovers. The success rate of that strategy has been poor and its results slow. Hyperion suggests an entirely different route of entry — through physical compute production rather than through knowledge production.
Louisiana does not need to become Silicon Valley in order to become indispensable to Silicon Valley’s machines. That is a genuinely new proposition in American regional economics, and whether it proves to be a durable one is among the most consequential open questions of the next decade.
4.2 Indiana: New Carlisle and the industrialization of compute
Northern Indiana offers a second model, distinguished by the fact that its constraint was never land and always electricity.
Amazon Web Services announced an approximately $11 billion data-center investment near New Carlisle in St. Joseph County, described at the time by Governor Eric Holcomb as the largest capital-investment announcement in Indiana’s history.[20] The site was selected in part because it already adjoined extra-high-voltage transmission and a very substantial existing substation — the Olive 345 kV station — allowing interconnection at scale without first waiting on entirely new long-haul lines. Amazon’s investment was made expressly contingent on securing long-term energy-service agreements. AEP is building two new 345 kV delivery substations, Larrison Drive and New Prairie, plus fault-duty upgrades at Olive, a transmission package of approximately $185 million targeting an initial roughly 1.1 gigawatts of firm service.[22]
What makes the Indiana case especially instructive is the sheer magnitude of the power requirement relative to the state. A filing with Indiana utility regulators cited an estimate of approximately 2,250 megawatts of service capability at full development — enough electricity for roughly 1.5 million households, or up to half the households in the state.[21] Ben Inskeep of the Citizens Action Coalition put the comparison in terms that local officials could not easily dismiss.
The potential electricity usage from data centers coming to Indiana, such as the Amazon data center in New Carlisle, is staggering and hard to comprehend. … I&M’s new data centers will use more electricity by 2030 than all residential customers in the state of Indiana used in 2023.
— Ben Inskeep, Program Director, Citizens Action Coalition of Indiana [21]
At that scale, a data center ceases to be another commercial customer and becomes a regional energy-planning event. Generation planning, transmission investment, residential rate design, and industrial-development policy begin converging around a single campus. The utility’s own analysis holds that the arithmetic can favor existing customers: by Indiana Michigan Power’s numbers, a one-gigawatt addition lowers residential and commercial rates by roughly seven percent each while nudging industrial rates up about four percent, and Amazon’s expert, extending that analysis to roughly four gigawatts of hyperscaler load, put generation savings to existing customers at approximately $122 million per year.[22]
The counter-argument is equally serious and equally quantitative. The typical I&M residential bill rose approximately 47 percent between 2016 and 2025, so the baseline against which savings are measured is itself steeply rising; critics warn of stranded costs if a small number of concentrated customers scale back; and regional reliability is tightening, with PJM reserve margins projected to fall toward eight percent by 2028–2029.[22] Both propositions can be true simultaneously — large loads can lower average unit costs while increasing systemic risk — and the resolution is not rhetorical but institutional, residing in the specific structure of the electric service agreement.
Indiana has, notably, begun to build that structure. In a filing concerning a separate Alphabet-affiliated campus in Monrovia, AES Indiana sought approval of an electric services agreement that includes financial assurances, minimum demand commitments, and exit provisions expressly framed as protections for existing customers, with rates set to reimburse the utility for the incremental annual revenue requirement.[21] That is the template. Whether it becomes the norm is the policy question.
4.3 Ohio: where compute begins commissioning its own electricity
Ohio demonstrates the phase transition — the point at which AI capital stops selecting locations where electricity already exists and begins creating electricity geography around itself. The state contains, simultaneously, the most contested small-scale example and the most spectacular large-scale one.
The small case: Bowling Green and the Apollo Generating Station
The Wood County project described in this paper’s introduction is the compressed version of the phenomenon. An eight-hundred-acre Meta data center rises on what was farmland; a 350-megawatt behind-the-meter gas plant with approximately 120 megawatts of battery storage is approved in under three months without public hearings under an expedited process intended for single-customer generation; the applicant appears in filings as a shell entity; the draft air permit becomes public after construction has begun.[4,5] Ohio law can allow such approvals in as little as forty-five days without hearings. Reuters found more than a dozen comparable projects nationally that won approval in under a year with little or no notice to residents, two of which are already operating.[4]
This is the Powered Periphery at its least attractive: fast, opaque, and technically legal. It is also, importantly, the version most likely to generate the political backlash examined in Section 6, because the residents affected did not merely disagree with the decision — they were not told a decision was being made.
The large case: PORTS-Pike
The Pike County development operates at an entirely different order of magnitude and, notably, with an entirely different disclosure posture. On 17 August 2026, Nvidia announced that it had secured land, power, and shell capacity through a partnership with SB Energy at the PORTS-Pike Technology Campus — the site of the former Portsmouth Gaseous Diffusion Plant, a Cold War-era uranium-enrichment facility spanning private land and federal property leased from the Department of Energy.[23]
The structure is worth setting out precisely, because it is one of the most consequential financings in the history of digital infrastructure. SB Energy will build, own, and operate the campus under a twenty-year lease to OpenAI, which is the customer for eight IT-gigawatts of capacity. Nvidia is the exclusive AI compute infrastructure provider and will invest $1.5 billion in SB Energy, joining SoftBank Group and OpenAI as investors. Under residual-value guaranties disclosed in a Form 8-K filed with the Securities and Exchange Commission on 17 August 2026, Nvidia provides credit support covering leases for approximately 4.25 gigawatts of IT load in aggregate, with an option to extend support to roughly 3.75 additional gigawatts at its sole discretion, and a cumulative payment obligation capped at $105 billion.[23,24,28] The first 800 megawatts are expected online in 2028.[25]
The energy program is proportionate. SB Energy and SoftBank will build at least ten gigawatts of new generation to yield eight IT-gigawatts of usable AI factory capacity, and will invest at least $4.2 billion in new regional grid infrastructure through a partnership with AEP Ohio that the parties describe as designed to protect ratepayers. A Department of Energy fact sheet from March 2026 indicated plans including at least 9.2 gigawatts of natural gas.[23,28] OpenAI has committed an incremental $40 million to build on SB Energy’s originally announced $40 million community benefits fund, for an initial $80 million anchor, with combined community investment, Codex credits, and existing funds reported at more than $160 million.[23]
Jensen Huang framed the transaction in terms that reveal the underlying siting logic with unusual clarity — not as the acquisition of a building, but as the acquisition of a durable position in the electrical system.
We are securing long-lived infrastructure for Nvidia compute so OpenAI can deploy the most productive AI factories that can be upgraded repeatedly with each new generation delivering more intelligence and better economics.
— Jensen Huang, Founder and Chief Executive Officer, Nvidia [26]
Read that sentence closely. The asset being secured is not the chips, which will be replaced several times over the lease term. It is the land, power, and shell — the capacity to host successive generations of silicon on the same energized footprint for twenty years. This is the Marginal Gigawatt Principle expressed as a corporate finance structure: when the scarce input is energized land rather than silicon, the rational strategy is to secure the land and let the silicon rotate through it.
The transaction has attracted the obvious criticism, namely that a chipmaker guaranteeing its customer’s lease obligations for a campus that will exclusively deploy its own chips constitutes circular financing with systemic risk implications. Huang publicly rejected the characterization.[27] The dispute is genuine and unresolved, and this paper takes no position on it beyond noting that the structure exists precisely because powered land at this scale cannot be financed conventionally — which is itself evidence for the scarcity the paper describes.
There is a final dimension to PORTS-Pike that deserves emphasis. The site is a decommissioned federal enrichment facility in a distressed Appalachian county that waited four decades for redevelopment. Its advantages — an existing federal land position, legacy heavy-industrial interconnection, a community with industrial memory, and a political consensus in favor of reuse — are exactly the advantages that brownfield periphery possesses and greenfield periphery does not. Both routes into the Powered Periphery exist. They produce very different local politics.
4.4 Pennsylvania: from energy region to compute region
Pennsylvania adds a fourth configuration, and in some ways the most economically radical one, because it collapses the supply chain between fuel and compute almost entirely.
On 11 August 2026, Alpha Compute Corp. announced a binding term sheet granting an exclusive option to acquire mineral, surface, and pore-space assets in northern Pennsylvania for a base purchase price of $55 million, with a $3 million deposit payable on execution of a definitive property purchase agreement. The assets include unleased Marcellus gas rights across approximately 1,800 oil-and-gas mineral acres carrying a 100 percent net revenue interest, subject to title confirmation, together with surface and pore-space property. The plan is a greenfield 200-megawatt behind-the-meter data-center campus fueled by on-site Marcellus gas, with potential expansion toward one gigawatt; development plans envision twelve wells drilled from two pads, leveraging nearby gas transmission and 115 kV electric infrastructure. A third-party evaluation cited by the company estimated that gas from the property could support 200 megawatts of continuous generation for ten years at an all-in power cost of approximately $0.0585 per kilowatt-hour, below referenced PJM commercial and industrial rates.[29,30]
Chief Executive Brittany Kaiser described the motivation to Reuters in terms that align precisely with the European evidence: surging demand for AI computing is driving a scramble for power across the United States, prompting developers to secure their own energy sources rather than wait for grid connections.[29]
The structural significance is worth stating plainly. The traditional model involved an energy producer selling fuel to a power producer, which sold electricity through a regulated grid to a data center, which sold compute to a customer. Four counterparties, four margins, four regulatory regimes, and one very long interconnection queue. The emerging model compresses those stages geographically and contractually:
The compressed value chain
Fuel → Generation → Data center → Compute — occurring on or near the same property, under common ownership, outside the interconnection queue, and largely outside the retail tariff. This is what pushes the Powered Periphery deep into the Five-Layer AI Economy, because the Energy Layer and the Datacenter Layer cease to be separate businesses.
It should be said clearly that Alpha Compute’s project is at term-sheet stage and remains subject to due diligence, financing, permits, and definitive agreements; no power or data-center capacity currently operates at the site.[30] The case is included here not as an accomplished fact but as an illustration of a transaction structure that would have been unthinkable five years ago and that a public company now considers worth $55 million to option. It is also, as Section 6 shows, a transaction structure that collided within a week with a very significant change in Pennsylvania’s regulatory posture.
4.5 The American pattern: four models, one mechanism
Louisiana, Indiana, Ohio, and Pennsylvania are not identical, and that is precisely the point. The Powered Periphery is not a single development template that can be exported. It is a mechanism — energized land attracts frontier compute — that expresses itself differently depending on what a region has.
Table 4 — Four American configurations of the Powered Periphery
| Case | Core endowment | Scale and structure | Power strategy | Distinctive risk |
| Richland Parish, LA (Meta Hyperion) | Very large greenfield land, cooperative state, capable utility | ≈5 GW IT; >$50bn; ≈3,200–4,000 acres; JV with Blue Owl | Utility-built: Entergy >7 GW gas, +2.5 GW renewables funded by Meta, ≈240 mi transmission | Concentration risk for a single state’s grid; ~1,000 permanent jobs against $50bn+ |
| New Carlisle, IN (AWS) | Existing EHV transmission and the Olive 345 kV substation | ≈$11bn announced; ≈2,250 MW at full development | Regulated utility expansion; ≈$185m AEP transmission package; large-load service agreements | Stranded-cost exposure; PJM reserve margins toward ~8% by 2028–29 |
| Pike County, OH (SB Energy / OpenAI / Nvidia) | Federal brownfield with legacy industrial interconnection | 8 IT-GW; 20-year lease; up to $105bn Nvidia credit support | Developer-built: ≥10 GW new generation incl. up to 9.2 GW gas; $4.2bn AEP Ohio grid partnership | Circular-financing critique; counterparty concentration |
| Northern PA (Alpha Compute) | Owned Marcellus gas rights and pore space | 200 MW planned, option to 1 GW; $55m asset option | Behind-the-meter, on-site wells; ≈$0.0585/kWh all-in target | Early stage; new state permitting regime imposed 18 Aug 2026 |
Table 4. Compiled by the author from company disclosures, SEC filings, regulatory filings, and press reporting through 20 August 2026.[15,18,20,21,22,23,24,29,30]
Louisiana offers enormous greenfield scale and a utility willing to add a third to the state’s generating capacity. Indiana combines existing industrial transmission geography with regulated utility expansion and an emerging template for large-load service agreements. Ohio illustrates both the fastest and the most opaque version of the phenomenon and, at PORTS-Pike, the version where generation is commissioned around compute at national scale. Pennsylvania demonstrates the gravitational pull of owned fuel.
Other states will develop their own variants according to their comparative advantages — Texas around gas and interconnection speed, the Southeast around nuclear uprates and cheap land, the Mountain West around renewables and cooling, Appalachia around brownfields and industrial memory. The emerging American AI map will therefore not be determined by the identification of one ‘next Silicon Valley.’ It will consist of many specialized power territories, each combining land, generation, transmission, fiber, political support, and industrial history in a different proportion.
The Powered Periphery, in other words, is plural. And because it is plural, it is competitive — which brings the argument to the layered framework in which these dynamics operate, and then to the politics they produce.
Section 5 — How the Powered Periphery Reshapes the Five-Layer AI Economy
The Five-Layer AI Economy — Energy, Chips, Datacenters, Models, and Applications and Agents — is usually presented as a conceptual stack, a way of organizing the value chain from electrons to interfaces. What the evidence assembled in this paper demonstrates is that the stack has acquired a physical geography, and that the geography does not distribute evenly across the layers. The lower layers are becoming intensely place-dependent at precisely the moment when the upper layers are becoming placeless. That divergence is the structural signature of the Powered Periphery.
5.1 Layer One: Energy becomes the geographic anchor
The Five-Layer AI Economy begins with Energy because every layer above it ultimately depends upon it, and the Powered Periphery reinforces that hierarchy in a way that earlier phases of computing did not.
When electricity was a manageable input — when a large data center drew thirty or fifty megawatts and the utility could accommodate it within ordinary planning margins — technology companies were free to optimize location around labor, customers, tax policy, and connectivity. Power was a cost line, not a constraint. When electricity requirements reach gigawatt scale, Energy stops being an input to the location decision and becomes the location decision. It begins choosing the geography for everything above it.
Every element of the electricity system therefore becomes a siting variable. Generation resources matter, and their fuel type determines the political character of the project. Transmission matters, and its available headroom determines the timetable. Natural-gas infrastructure matters, because pipeline proximity determines whether behind-the-meter generation is even possible. Nuclear facilities matter, both as existing baseload and as candidates for uprates. Renewable availability matters, both economically and for corporate procurement commitments. Substations matter. Transformer lead times matter, and have become one of the least-discussed hard constraints in the entire industry. Interconnection queue position matters, sometimes more than anything else.
The global aggregate makes the point at scale. The International Energy Agency’s Energy and AI report projected that global data-center electricity consumption would more than double by 2030 to approximately 945 terawatt-hours — slightly more than the entire electricity consumption of Japan today — with electricity demand from AI-optimized data centers projected to more than quadruple.[8,10] Fatih Birol framed the finding in a way that has since become the organizing sentence of the entire policy debate.
AI is one of the biggest stories in the energy world today — but until now, policy makers and markets lacked the tools to fully understand the wide-ranging impacts. Global electricity demand from data centers is set to more than double over the next five years, consuming as much electricity by 2030 as the whole of Japan does today. The effects will be particularly strong in some countries.
— Dr Fatih Birol, Executive Director, International Energy Agency [8]
That last sentence — the effects will be particularly strong in some countries — is the international-scale version of this paper’s argument. The IEA found that in the United States, data centers are on course to account for almost half of the growth in electricity demand; in Japan, more than half; in Malaysia, as much as one-fifth.[8] Concentration is the rule, not the exception, and the same concentration occurs one level down, within countries, at the level of parishes and counties.
The IEA’s 2026 update sharpened the picture further. Data-center electricity demand grew by seventeen percent in 2025, while electricity use at AI-focused data centers surged by fifty percent; the agency now projects that consumption at AI-focused facilities will triple between 2025 and 2030.[9] Birol also introduced a reframing that matters for how peripheral regions should think about their position.
The IEA was early in recognizing that there is no AI without energy — and that countries that provide secure, affordable and rapid access to electricity will be one step ahead. Now, we see that while AI is still an energy taker, it is also becoming an energy maker — driving forward innovative solutions like next-generation nuclear reactors, flexible data centers and long-duration energy storage.
— Dr Fatih Birol, Executive Director, International Energy Agency [9]
The domestic American picture is equally stark. Lawrence Berkeley National Laboratory’s 2024 report, prepared for Congress under the Energy Act of 2020, found that total United States data-center electricity use climbed from 58 terawatt-hours in 2014 to 176 terawatt-hours in 2023 — approximately 4.4 percent of total United States electricity — and projected an increase to between 325 and 580 terawatt-hours by 2028, or roughly 6.7 to 12 percent of national consumption.[6] The laboratory’s 2025 update revised the trajectory upward, estimating that data centers could account for approximately 11.8 percent of total United States electricity by 2030, within a scenario range of 9.5 to 15.3 percent.[7]
Figure 4. United States data-center electricity consumption, historical and projected. The divergence after 2023 is the AI training build-out. Sources: Lawrence Berkeley National Laboratory, 2024 report and 2025 update.[6,7]
Layer One, in short, no longer sits underneath the stack as a passive utility. It has become the layer that determines where the other four can physically exist.
5.2 Layer Two: Chips become locationally dependent assets
An advanced accelerator has enormous theoretical value. An advanced accelerator without electricity, cooling, and networking is stranded capital — a depreciating asset generating no return while its replacement is already being taped out.
As hyperscalers deploy hundreds of thousands and eventually millions of accelerators, silicon economics therefore become inseparable from site economics. A rural AI campus is, in the most literal financial sense, a machine for converting electrical capacity into a place where semiconductor capital can earn a return. The scarcity sequence has evolved through four distinct stages in roughly four years:
- Can we obtain the GPU? — the shortage phase of 2022–2023, when allocation was the binding constraint.
- Can we obtain enough GPUs? — the scaling phase, when supply improved and the constraint moved to capital and manufacturing capacity.
- Can we obtain enough electricity to operate them? — the present phase, in which energized capacity rather than silicon determines deployment velocity.
- Where can we continuously operate and replace generations of those accelerators for twenty years? — the emerging phase, in which the asset being acquired is the durable site rather than the perishable hardware.
The Nvidia–OpenAI–SB Energy arrangement in Ohio is the clearest available expression of stage four. The residual-value guaranties cover land, power, and shell — not chips. The twenty-year lease term exceeds the useful life of any accelerator generation by an order of magnitude. Nvidia’s own framing of the deal describes securing long-lived infrastructure that can be upgraded repeatedly with each new generation.[23,26] The company is not financing a data center; it is financing a permanent socket for its own product line.
This is Layer Two being reorganized by Layer One, and it has a corollary that peripheral regions should understand clearly: a region that hosts durable energized capacity captures a claim on many successive generations of silicon investment, not merely on the one being installed today. That is the strongest economic argument available to a host community, and it is rarely the one that developers lead with.
5.3 Layer Three: the datacenter becomes an industrial campus
Layer Three experiences the most visible transformation, and it is the layer where the vocabulary has most conspicuously failed to keep pace with the object.
The traditional image of the data center — an anonymous, low-slung, windowless warehouse in a business park, distinguished chiefly by its security fencing — has become inadequate to describe what is being built. Gigawatt AI facilities resemble industrial complexes: dedicated electrical systems, cooling plants of substantial physical scale, on-site substations, transmission infrastructure, water treatment, and increasingly their own generation. The design language is closer to petrochemical or metals processing than to office computing.
That distinction transforms local regulation, and much of the political conflict documented in Section 6 originates in the mismatch. A warehouse can be handled through zoning. A multi-gigawatt AI campus requires energy planning. A cloud facility can be discussed as technology infrastructure by an economic-development office. An AI factory simultaneously becomes water policy, industrial policy, tax policy, environmental policy, air-quality policy, and grid policy — and it typically arrives at a county planning commission that has jurisdiction over exactly one of those things.
The physical footprint numbers make the mismatch concrete. Globally, data centers now occupy on the order of 1.2 million acres, and the supply-chain infrastructure surrounding them generates further development.[60] A single campus in Richland Parish spans thousands of acres. Eight hundred acres in Wood County, Ohio, converted from farmland to a construction site within roughly a year.[4] These are land-use events of a magnitude that American local government has historically encountered only in the context of airports, military installations, and interstate highways — all of which came with decades of federal process attached.
5.4 Layer Four: models become geographically portable
Layer Four explains why the Powered Periphery can exist at all, and it is the hinge on which the entire argument turns.
AI models are unusually portable economic outputs. The intelligence produced by an enormous training run is, once complete, a set of weights that can be replicated and distributed globally at negligible marginal cost. This weakens — nearly severs — the traditional relationship between production location and customer location that has organized industrial geography since the railway.
A frontier model trained in rural Louisiana can serve users in Los Angeles. A model trained in northern Sweden can power agents in Paris. A model optimized in Ohio can be deployed through cloud infrastructure on six continents. The distance between where intelligence is produced and where it is consumed can be enormous without any economic penalty whatsoever, because the good being transported weighs nothing.
It is worth appreciating how unusual this is. Most heavy industries that migrated toward stranded energy did so despite a transport penalty on the output — aluminum ingots must still be shipped. AI training has the energy intensity of heavy industry with the transport economics of software. That combination has no close historical precedent, and it is what permits the physical production layer to migrate outward without dragging the market with it.
There is one important qualification, and it is regulatory rather than physical. Data-residency and sovereignty rules can impose an artificial transport penalty on weights and on the data used to produce them. The European Union’s preference for in-territory infrastructure, and the sovereignty framing of the AI gigafactory program, are precisely attempts to reimpose geography on a layer that would otherwise have none.[63,64] Sovereignty, in this sense, is the political system’s response to Layer Four’s placelessness.
5.5 Layer Five: applications and agents reconnect the periphery to the world
Layer Five closes the loop. Applications, copilots, robots, and autonomous agents convert centrally or remotely manufactured intelligence into useful economic activity, and they do so at the point of consumption.
Most end users will never know where the underlying computation occurs, and the abstraction is deliberate. The user sees an interface. The corporation sees an application programming interface. The agent sees tools. Beneath those abstractions may sit a multimillion-square-foot campus beside farmland, a nuclear station, a gas pipeline, a hydroelectric resource, or a rural substation several hundred miles away — and a behind-the-meter turbine across the street from a home daycare.
The Five-Layer AI Economy therefore produces a remarkable geographic paradox, and it is the paradox this paper exists to describe: the intelligence becomes increasingly ubiquitous while the infrastructure producing it becomes increasingly place-dependent. The cloud made location appear irrelevant. AI makes location strategically decisive again — but only at the bottom of the stack, where almost no user ever looks.
Table 5 — The Five-Layer AI Economy and its geographic anchoring
| Layer | Economic function | Geographic behavior | Who holds the leverage |
| 1 — Energy | Supplies the physical input that makes all other layers possible | Immobile. Determined by fuel basins, hydrology, transmission history, and siting law | Utilities, state commissions, grid operators, landowners |
| 2 — Chips | Converts capital into computational capability | Mobile as hardware, immobile as deployed capital; follows energized sites | Foundries, designers, and now the owners of land-power-shell |
| 3 — Datacenters | Converts electricity and silicon into usable compute | Increasingly peripheral for training, metro-proximate for inference | Developers, counties, permitting authorities, host communities |
| 4 — Models | Converts compute into transferable machine intelligence | Essentially placeless; constrained only by sovereignty and data-residency rules | Frontier labs, national regulators |
| 5 — Applications & Agents | Converts intelligence into economic activity at the point of use | Follows population and enterprise demand — back to the metropolis | Enterprises, platforms, end users |
Table 5. Author’s framework. Note the shape: the stack is anchored at the bottom and free at the top, which is why compute production migrates outward while consumption does not.
Section 6 — The Politics of the Powered Periphery
Everything described so far is an economic mechanism. What follows is what happens when that mechanism meets a democracy. The single most important development of 2026 in this domain is not a technical one — it is the discovery, by rural and small-metropolitan communities across the United States and Europe, that they possess more leverage over the AI build-out than anyone, including themselves, had assumed. The Powered Periphery creates political power precisely because the periphery contains voters.
6.1 From economic development to infrastructure governance
Governors and county officials understandably perceive enormous AI projects as economic-development opportunities, and in a narrow sense they are. Billions of dollars of investment reshape tax bases, construction employment, and a region’s perception of its own relevance. The announcement is a genuine political asset.
But projects measured in gigawatts cannot be governed as though they were conventional corporate relocations, and the attempt to do so is the source of most of the conflict now visible. A gigawatt campus does not simply occupy a site; it draws on infrastructure systems shared with everyone else. Electricity becomes political because bills are visible monthly. Water becomes political because it is hyperlocal and finite. Tax incentives become political because they are quantifiable and comparable. Transmission becomes political because lines cross other people’s property. Land use becomes political because conversion is permanent. And project secrecy becomes political because secrecy in a public process reads, correctly, as an assertion that the public is not a party to the decision.
The scale of the shift in public sentiment can be dated with unusual precision. Five years ago, data centers were ribbon-cutting events. By mid-2026, more than 100 local communities had enacted moratoriums, more than 300 state data-center bills had been filed in the first six weeks of 2026 alone, and several states that once competed to offer the largest incentives — Virginia, Georgia, and Oklahoma — were reconsidering those programs entirely.[42] On 25 March 2026, Senator Bernie Sanders and Representative Alexandria Ocasio-Cortez introduced the AI Data Center Moratorium Act, which would pause new large-scale AI data-center construction until Congress passes legislation addressing AI safety, worker protections, and environmental standards.[70]
Public opinion is more nuanced than the moratorium count suggests, and the nuance is important for developers who assume the opposition is purely about bills. A Harvard/MIT poll overseen by Harvard researcher Stephen Ansolabehere found that approximately 40 percent of respondents supported data centers in their areas against approximately 32 percent opposed — less support than auto factories and e-commerce warehouses attract, but more than petrochemical facilities, which drew 23 percent support and 52 percent opposition. Roughly two-thirds expected a data center to raise local electricity prices ‘a lot’ or ‘somewhat.’ The poll’s central finding, however, was that resistance is driven less by electricity prices than by how the projects might alter communities.[46] That is a finding about identity and consent, not about arithmetic, and it is not addressable through a rate case.
6.2 Pennsylvania, 18 August 2026: the regulatory turn
The clearest single demonstration of the political turn occurred two days before this paper’s data cut-off, and it occurred in a state whose governor had been among the industry’s most prominent champions.
On 18 August 2026, Governor Josh Shapiro signed Executive Order 2026-05, directing the Pennsylvania Department of Environmental Protection to review permit applications only where developers have made a legally binding commitment to meet the Governor’s Responsible Infrastructure Development (GRID) Requirements and have obtained local approval. The order removes all AI data-center proposals from the state’s Permit Fast Track Program and makes future data-center projects ineligible for it; prohibits the use of nondisclosure agreements for data-center projects; imposes new public transparency requirements; disqualifies non-compliant projects from the existing sales-tax exemption on data-center equipment; and directs the DEP to publish a publicly accessible permitting map of all proposed projects.[31,32,34]
The GRID Requirements themselves are organized around four principles: energy affordability, transparency and community engagement, workforce and economic development, and environmental protection. In practice they require developers to bring their own power and pay all associated electricity costs, hire locally and provide community benefits, be transparent with local government, and meet stringent air and water standards, with an increasing share of generation coming from renewables, advanced nuclear, or battery storage.[31,34]
The scale of the speculative problem the order responds to is documented in the state’s own numbers. The DEP had become aware of more than 100 data-center proposals in publicly sourced databases; 58 projects had engaged with the department at some level of formality; 15 had applied for at least one permit; and only 5 had received all permits required for a first phase of development.[32] That distribution — a hundred proposals producing five permitted projects — is the statistical portrait of a speculative land rush, and it explains why the state’s stated objective was to weed out speculative proposals rather than to stop development.
Shapiro’s own framing was unusually direct for an executive-action announcement, and it is worth recording precisely because it comes from a governor who had previously secured a multi-billion-dollar Amazon commitment for the state.
I’ve heard loud and clear from the good people of Pennsylvania as I’ve traveled our commonwealth, and I’m here to say today that we will not be bullied by these developers.
— Governor Josh Shapiro, Commonwealth of Pennsylvania [33]
This executive order represents the most stringent requirements in the entire country. And my word to developers is you better follow these requirements. Otherwise you shouldn’t plan to do business here.
— Governor Josh Shapiro, Commonwealth of Pennsylvania [31]
On the tax exemption specifically, Shapiro’s remarks placed the argument on distributional rather than environmental grounds, which is a notable rhetorical relocation of the issue.
These are the wealthiest companies in the world — they shouldn’t get a tax break for gouging our communities.
— Governor Josh Shapiro, Commonwealth of Pennsylvania [31]
The reaction demonstrates why this issue does not resolve along conventional partisan lines. Bloomberg Law characterized the order as imposing the nation’s strictest data-center rules.[35] The Chamber of Progress, a technology-industry-aligned group, welcomed it, with its director of economic analysis arguing that transparency and community protection prior to construction represent the best path forward. Environmental advocates found it insufficient: Megan McDonough, Pennsylvania state director of Food & Water Watch, argued that the measures do very little to protect the people who will live with the consequences.[34] Republican gubernatorial nominee Stacy Garrity attacked it from the opposite direction, accusing the governor of reversing a position he had held for thirteen months.[36] A single executive order was simultaneously too strict, too weak, and too late — which is the characteristic signature of a genuinely contested infrastructure question rather than a partisan one.
The operational mechanism deserves attention because it is more consequential than the rhetoric. DEP Secretary Jessica Shirley clarified that developers who do not follow the standards will simply not have their permit applications reviewed until local land-use approval is obtained.[34] This converts local consent from a soft constraint into a hard sequencing requirement. In the language of Section 3, Pennsylvania has legislated energization probability into a function of community consent.
6.3 The federal counter-current
It is essential to note that state and federal policy are currently pulling in opposite directions, and that the resulting tension is one of the defining features of the American Powered Periphery.
On 23 July 2025, alongside America’s AI Action Plan, the White House issued an executive order on Accelerating Federal Permitting of Data Center Infrastructure, explicitly framed around facilitating the rapid and efficient build-out of this infrastructure by easing federal regulatory burdens. The order directs agencies to provide financial support for qualifying projects, streamline environmental and permitting review, expand FAST-41 coverage, and identify suitable federal land for siting new data centers; it also revoked the prior administration’s Executive Order 14141.[52] Qualifying-project status under the associated Commerce initiative requires a minimum capital commitment of $500 million.[54] In November 2025, Executive Order 14363 launched the Genesis Mission, a Department of Energy-led initiative to harness AI for scientific discovery, explicitly modeled on the Manhattan Project and Apollo program framing.[53]
The federal posture is therefore acceleration; the emerging state posture in several jurisdictions is conditionality. Federal land is being offered for siting while state permits are being conditioned on local consent. This is not a contradiction that will resolve itself quietly, and it produces a genuinely novel jurisdictional landscape in which the same project may be simultaneously fast-tracked federally and stalled locally.
The federal government has also intervened on the ratepayer question directly. In March 2026 the White House issued a Ratepayer Protection Pledge Proclamation, signed by seven major technology companies.[40,54] The pledge’s practical significance is contested — Harvard’s Ari Peskoe has argued that it does not change the underlying arithmetic of who pays for grid expansion[39] — but its existence is itself evidence that the cost-allocation question has become politically unavoidable at the highest level.
6.4 Community consent as an infrastructure input
One of the largest strategic errors an AI developer can make is assuming that rural implies politically passive. The evidence of 2026 suggests the opposite may be closer to the truth.
Smaller communities frequently possess a closer relationship with land, water, schools, and local government than residents of large metropolitan regions. Local government is proximate and personally known. Land ownership is often multi-generational. A giant industrial facility in such a setting is extraordinarily visible in a way it would not be in an outer suburb. Residents know which farmland disappeared and whose family owned it. They see the construction traffic. They hear the generators. They watch the transmission towers arrive. They notice the water infrastructure. And they compare utility bills, in public, at length.
Community acceptance must consequently be treated as an infrastructure input with the same seriousness as transformer availability. The site-selection equation introduced in Section 1 should therefore be restated in its full operational form:
The durable-capacity equation
Power + Land + Fiber + Water + Community Consent = Durable Compute Capacity. A project possessing the first four without the fifth is not merely unpopular. It is financially fragile, because in an increasing number of jurisdictions the fifth is now a legal precondition for the first four.
The Brookings research on community benefits agreements makes the institutional case for formalizing this. Nicol Turner Lee and Darrell West have argued that such agreements are necessary rather than optional for data centers, and Anthony Pipa and Adam Aley, examining rural implications specifically, conclude that rural communities are attractive sites but must weigh the economic-development trade-offs given the limited number of permanent jobs, complex fiscal arrangements, and environmental risks — underscoring the need for transparency, a focus on community benefits, and stronger local input into siting and permitting.[44,45]
There is also a positive-sum version of this argument that deserves more attention than it receives. Daniel Goetzel, Mark Muro, and Shriya Methkupally argue that the standard development model — speedy dealmaking and opaque negotiations — delivers short-term construction jobs and revenue but little durable local upside, while AI-era scaling and competition for mega-sites, grid access, and permits are giving regions new leverage. Their conclusion is that negotiated co-investments can anchor regional technology ecosystems, and that regions should ask for this.[43] The leverage is real. Most communities simply have not yet learned to price it.
6.5 Rural competition between states, and the race-to-the-bottom risk
The next stage of this dynamic will involve competition not merely between technology hubs but between power-rich states and regions — and the terms of that competition will determine how the gains are distributed.
Governors can compete through tax incentives. Utilities can compete through special large-load tariffs. Counties can pre-zone land and pre-permit substations. States can streamline environmental review. Economic-development agencies can prepare shovel-ready sites with completed interconnection studies. Regions can market nuclear assets, gas resources, renewable potential, and transmission headroom as explicit products.
But aggressive competition creates a familiar risk. If every state attempts to attract hyperscalers by promising the cheapest electricity, the fastest permitting, and the largest incentives, states may socialize the infrastructure costs while hyperscalers capture much of the economic upside. The Virginia data-center sales-tax exemption alone cost an estimated $1.6 billion in fiscal year 2025.[42] Whether that figure represents a giveaway or a bargain depends entirely on what was purchased with it — and, as the following section shows, the rigorous evidence on that question is considerably more surprising than either side of the debate typically admits.
The strongest state strategy is therefore not to attract the most gigawatts. It is to attract the highest-quality gigawatts: projects whose infrastructure commitments, employment, fiscal contributions, environmental standards, and community obligations create durable regional benefit. That is a harder thing to announce at a press conference and a much better thing to own in ten years.
6.6 AI federalism and the midterm dimension
The Powered Periphery is producing what might reasonably be called AI federalism: an increasingly complicated division of authority among federal agencies, governors, state utility commissions, regional grid operators, counties, and municipalities, in which no single level of government possesses complete control over any large project.
The federal government controls certain permits, certain federal land, export policy, and a national acceleration agenda. State utility commissions control tariffs, cost allocation, and service agreements — arguably the most consequential levers of all. Governors control state permitting posture and executive discretion, as Pennsylvania has demonstrated. Grid operators control interconnection and capacity markets. Counties control zoning and, increasingly through state action, an effective veto. Landowners control the land. Any of these actors can slow a multi-billion-dollar project; none can guarantee one.
By the November 2026 midterm elections, AI infrastructure can no longer be treated as a technology-policy subject confined to specialist committees. Candidates in affected states will increasingly confront a set of questions that divide conventional political categories:
- Should residential customers subsidize grid expansion driven by large industrial loads, or should those loads pay their own way through separate tariff classes?
- Should hyperscalers be required to finance dedicated generation, and should that generation be required to be firm, clean, or both?
- Should productive agricultural land be converted to industrial use, and should states use zoning overlays to steer development toward brownfields?
- Should communities receive guaranteed minimum tax revenue that survives incentive renegotiation?
- Should water-intensive projects face regional limits, and should peak rather than average water demand be the disclosed metric?
- Should utilities be required to disclose large-load contracts, or may they remain confidential as commercially sensitive?
- Should companies be permitted to negotiate projects with public bodies behind nondisclosure agreements?
- Should states prioritize nuclear, natural gas, renewables, or storage in serving AI load, and who bears the stranded-cost risk of that choice?
- Should local governments possess explicit veto authority, and at what level of government should that authority be established?
- Should an AI campus be treated as an ordinary large customer or as strategic industrial infrastructure with a distinct regulatory regime?
These questions divide conventional coalitions because there are legitimate arguments on several sides of each. A pro-growth governor may favor investment while insisting that hyperscalers bear infrastructure costs. A rural county may welcome tax revenue while opposing water withdrawals. A utility may welcome enormous new customers while worrying about reliability and stranded assets. Environmental organizations may support clean-energy-powered AI while opposing new gas generation. Ratepayer advocates may support economic development while demanding safeguards against cost shifting. Organized labor may support construction employment while questioning the permanent job count.
Brookings analysis of the political geography of AI exposure found that sixty-two of the hundred most AI-exposed United States counties voted Democratic in the 2024 presidential election, a correlation the authors attribute to occupational sorting rather than ideology.[43] The Powered Periphery adds a second and largely orthogonal political map: the counties that host the infrastructure are frequently not the counties whose labor markets are most exposed to the models it produces. Two different AI politics are therefore forming in two different sets of places, and they are not natural allies.
6.7 From rural extraction to rural participation
The most important policy objective, and the one against which this entire phenomenon should ultimately be judged, is whether the Powered Periphery participates in the value it helps create.
A rural county should not merely host accelerators. It can build technical education around them: community colleges training electricians, cooling technicians, network specialists, water-treatment operators, and power engineers — the Louisiana Delta Community College scholarship program funded by Meta is a small but real instance of the model.[16] Universities can establish energy-and-compute research programs. Local manufacturers can enter data-center supply chains, which are deep and include everything from switchgear to structural steel to specialized piping. Utilities can modernize infrastructure that serves everyone. Municipalities can negotiate long-lived community-benefit agreements with enforceable terms. Agricultural regions can protect their most productive land through zoning while steering development toward marginal parcels and brownfields. Former industrial regions can reuse contaminated land that has no other economic future — the PORTS-Pike precedent. Waste heat can support district heating or industrial processes where technically and economically practical, as the Narvik project in Norway is attempting.[62]
The difference between extraction and development is a single test, and it is one that can be applied honestly ten years after the fact: is the region economically stronger after the construction crews leave? Not richer during the build. Stronger afterwards — in skills, infrastructure, institutions, fiscal capacity, and optionality. That should become the central political test of the Powered Periphery, and it is a test that very few of the projects reviewed for this paper are currently structured to pass.
Section 7 — What the Evidence Actually Shows: The Contested Ledger
A paper that argued only its own case would not be worth writing. The empirical literature on the local effects of data centers expanded dramatically between 2024 and 2026, and it does not point in a single direction. On the two questions that matter most to host communities — what happens to electricity bills, and what happens to jobs — the best available research is genuinely contested, and in one case the most rigorous studies point in opposite directions. This section sets those findings against one another rather than selecting the convenient ones, because the credibility of everything argued above depends on doing so.
7.1 The electricity-bill question: three serious answers that disagree
Begin with the fact that public belief is close to unanimous. Roughly two-thirds of Americans surveyed expect a nearby data center to raise local electricity prices, and roughly 70 percent oppose data centers being sited near them.[40,46] Average residential electricity rates rose approximately 32 percent nationally between July 2020 and July 2025, and approximately 39 percent over five years by another measure.[42] The correlation is visible to everyone with a mailbox.
The causal question is harder, and three serious research programs have reached materially different conclusions.
The cost-shifting argument
Eliza Martin and Ari Peskoe of Harvard Law School’s Environmental and Energy Law Program reviewed nearly fifty regulatory proceedings concerning utility rates for data centers and argued that the conventional ratemaking model is being used to socialize costs that benefit a narrow set of very large customers. Their core mechanism is straightforward: government-regulated utility rates socialize a utility’s costs of providing service on the premise that society benefits from growing electricity use; data centers upend that premise because the infrastructure being built serves a handful of corporations rather than the public generally.[37]
The subjectivity and complexity of ratemaking conceals utility attempts to funnel revenue to their competitive lines of business by overcharging captive ratepayers. … The public faces significant risks that utilities will profit from new data centers by making major investments and then shifting costs to their captive ratepayers.
— Eliza Martin and Ari Peskoe, Harvard Law School Environmental and Energy Law Program [37]
Peskoe has extended the argument to the wholesale market, and the magnitudes he cites are large.
Data centers are causing tens of billions of dollars of price increases in wholesale power markets and driving utilities to spend tens of billions of dollars on delivery infrastructure. In general, these cost increases are spread to all ratepayers by the utility.
— Ari Peskoe, Director, Electricity Law Initiative, Harvard Law School [38]
A structural feature of his critique is that the evidence is deliberately hard to obtain: utilities routinely request confidential treatment of their contracts with data centers, which limits scrutiny of proposed deals and narrows regulators’ options.[37] Confidentiality is not merely a transparency problem; it is an evidentiary one, and it is the reason Pennsylvania’s prohibition on nondisclosure agreements is more analytically significant than it first appears.
The wholesale-market data supports the direction of Peskoe’s claim even where it does not settle the retail question. PJM capacity prices cleared at $28.92 per megawatt-day for the 2024/2025 delivery year, $269.92 for 2025/2026, and $329.17 for 2026/2027 — an increase of roughly a factor of ten in two auctions, with the final figure constrained by a price cap arising from a settlement between PJM and Governor Shapiro.[56,57] PJM’s independent market monitor, Monitoring Analytics, found that data-center load accounted for approximately $6.5 billion, or 40 percent, of the $16.4 billion in costs from one recent capacity auction, and that data-center forecasts made up 45 percent of $47.2 billion in capacity costs across the last three auctions. Its conclusion was unambiguous.[55]
Data center load growth is the primary reason for recent and expected capacity market conditions, including total forecast load growth, the tight supply and demand balance, the significant shortfall in cleared capacity, and high prices.
— Monitoring Analytics, Independent Market Monitor for PJM Interconnection [55]
Researchers at Carnegie Mellon University have projected that data-center growth could increase electricity bills by approximately 8 percent nationally and by as much as 25 percent in some regional markets.[71] One analysis published in mid-2026 put cumulative price increases attributable to data centers at roughly $23 billion.[72]
The counter-finding
Against this, a 2026 study by Asa Watten, John Bistline, and Geoffrey Blanford at the Electric Power Research Institute used an instrumental-variables approach and reached the opposite conclusion: that data centers caused average retail electricity rates in the United States to fall modestly between 2015 and 2024. Their estimate implies that a doubling of data-center capacity causes residential retail prices to fall by approximately 3.5 percent for a fixed level of residential demand.[40]
The mechanism they propose is not exotic. Electricity distribution is a natural monopoly with very high fixed costs; when a large customer arrives and pays for its share of those fixed costs, the fixed-cost burden per unit of delivered energy falls for everyone. The authors explicitly address the natural objection — that industrial and commercial customers pay lower rates and might therefore be subsidized by residential ones — and find that the effect did not materialize on average over their sample period.[40] They also point out that the historical record is consistent with their result: through the second half of the twentieth century, United States electricity demand rose while real prices fell, and inflation-adjusted rates were roughly flat from 2021 to 2024 even as rates rose nominally.[40]
Indiana Michigan Power’s own filings point the same way at the utility level, estimating that a one-gigawatt addition lowers residential and commercial rates by roughly 7 percent each, with roughly $122 million in annual generation savings to existing customers at approximately four gigawatts of hyperscaler load.[22]
The middle finding
A third study by Zhenxuan Wang, linking facility-level data on data-center entry to utility-level data on retail prices, infrastructure investment, and rate-case proceedings, found that data-center entry between 2010 and 2024 increased average retail electricity prices by 2.7 percent — 2.1 percent for residential customers, 2.8 percent for commercial, and 4.2 percent for industrial.[41] The paper’s framing is the most useful of the three for policy purposes, because it locates the answer not in the physics of load but in institutions: who pays depends on the rules that finance, regulate, and allocate the necessary infrastructure investment.
Table 6 — Three research findings on data centers and retail electricity prices
| Study | Method and period | Headline finding | Interpretive caution |
| Martin & Peskoe, Harvard ELI (2025) | Legal and institutional review of ~50 rate proceedings | Ratemaking structures and confidential contracts can shift Big Tech’s costs to captive ratepayers | Identifies mechanisms and risk; not an econometric estimate of realized average effect |
| Watten, Bistline & Blanford, EPRI (2026) | Instrumental-variables estimation, U.S., 2015–2024 | Data centers modestly reduced average retail rates; doubling capacity → ≈3.5% lower residential prices | Sample predates the largest AI build-out; historical averages may not survive gigawatt-scale concentration |
| Zhenxuan Wang (2026) | Facility-level entry linked to utility prices and rate cases, 2010–2024 | Entry raised average retail prices 2.7% (residential 2.1%, commercial 2.8%, industrial 4.2%) | Effect size depends heavily on state regulatory institutions; wide variation across jurisdictions |
Table 6. The three programs differ in method, period, and unit of analysis, which explains much of the divergence. All three predate the 2026 wave of multi-gigawatt campuses.[37,40,41]
How should a reader hold three findings that disagree? The honest synthesis is this. The average historical effect of data centers on retail rates is small and its sign is genuinely uncertain. The wholesale capacity-market effect in constrained regions such as PJM is large, recent, and not seriously disputed. The distributional effect depends almost entirely on state regulatory institutions — on whether large-load tariffs exist, whether minimum demand commitments and exit provisions protect existing customers, and whether contracts are visible to regulators and the public. And every one of these studies uses data that predates the gigawatt era. Extrapolating from 2015–2024 averages to an eight-gigawatt campus is an extrapolation well beyond the support of the data, in both directions.
The practical implication for a host community is therefore not ‘bills will rise’ or ‘bills will fall.’ It is that the outcome is a policy variable, not a physical constant — which is a considerably more useful thing to know, because policy variables can be negotiated.
7.2 The employment question: fewer than promised, more than claimed
The employment debate has, until recently, been conducted almost entirely without rigorous evidence. Industry-sponsored studies compared data-center counties to other counties and reported large gains; critics compared announced permanent headcount to announced capital investment and reported near-zero returns. Neither approach was methodologically adequate, because data-center sites are selected precisely for characteristics — cheap land, available power, fiber, recent economic growth — that predict growth independently.
Dany Bahar and Greg Wright addressed this directly by assembling a dataset of approximately 1,500 United States data-center facilities together with 52 announced-but-canceled projects, linked to county-level Bureau of Labor Statistics employment and wage data from 2003 to 2024. Comparing labor markets with built facilities to labor markets with announced-but-canceled facilities controls for the site-selection problem, since the canceled counties were selected on the same criteria.[42]
Their findings, updated in August 2026, are nuanced in a way that should discomfort both camps. Labor markets receiving their first large data center see employment in the data-processing sector rise by 56 percent over the first decade of operations, and telecommunications employment rise by 43 percent. At a typical treated county, these estimates imply roughly 100 to 200 jobs depending on facility type. Wages were unchanged. Home prices rose modestly, by 2 to 5 percent.[42]
Three further findings from the same work are, in my judgment, more important for the Powered Periphery than the headline numbers.
- Facility type is decisive. Hyperscale counties — those hosting facilities built by cloud and AI companies to run their own workloads — see telecommunications-sector gains that colocation counties do not, because a hyperscale campus creates demand for high-capacity fiber and network operations, while a colocation facility leases space to remote tenants who may have no local operational presence at all.[42]
- Incentives are poorly targeted. In hyperscale counties, state incentives represent approximately 2 percent of total construction investment, because location decisions are driven by power availability, land, and fiber rather than by tax breaks. In colocation counties, incentives represent approximately 62 percent of total investment — meaning subsidies matter most for precisely the facilities that generate the smallest employment benefits.[42]
- Workers see little direct gain. Wages are unaffected while home prices rise 2 to 5 percent, which is a mildly regressive combination for existing residents who do not own property.[42]
The second finding is the most policy-relevant sentence in the entire empirical literature reviewed for this paper, and it deserves to be stated in its implication: if hyperscale siting is driven by power, land, and fiber rather than by tax policy, then hyperscale tax incentives are largely paying for behavior that would have occurred anyway — while colocation incentives, which do change behavior, buy the weaker outcome. That is close to an exact inversion of how most state incentive programs are designed.
It also, incidentally, validates this paper’s central mechanism from an unexpected direction. Bahar and Wright did not set out to test the Powered Periphery thesis; they set out to measure employment. Their finding that hyperscale location decisions are driven by power, land, and fiber rather than tax breaks is an independent econometric confirmation of Assad Noori’s observation that data centers are being brought to where the power is.
7.3 Water: the constraint that is hyperlocal and therefore underestimated
Water has been the most distorted element of the public debate, in both directions, and it is worth being precise about what the research actually establishes.
The widely circulated figure that a single hundred-word AI response consumes roughly a water bottle originated in work by Pengfei Li, Jianyi Yang, Mohammad Islam, and Shaolei Ren, and has been substantially misreported. Ren himself now puts the figure closer to 15 milliliters for a comparable prompt, including roughly five milliliters for on-site cooling, and describes the original estimate as outdated for today’s systems.[49] The per-query framing was never the useful one.
The useful finding is entirely different and considerably more consequential for the Powered Periphery. Research by Ren with Yuelin Han, Pengfei Li, and Adam Wierman of Caltech shifts the analysis from annual averages to peak demand, and from national aggregates to municipal water systems. Peak demand for data centers using evaporative cooling can run six to thirty times the annual average. The study estimates that United States water systems may require between $10 billion and $58 billion in new infrastructure by 2030 to accommodate data-center growth, requiring up to 1.45 billion gallons of additional peak daily capacity.[47]
People recognize power as a constraint for data center growth, but most of them haven’t realized water is a hidden and even more binding constraint in many communities. … Even if you have money, the water source is another challenge. In many cases, the water is naturally replenished by snowpack and reservoirs. But reservoirs and snowpack are limited. You may have money to build treatment plants and pipes, but money can’t buy more snowpack.
— Shaolei Ren, Associate Professor, Bourns College of Engineering, University of California, Riverside [47]
Ren’s framing of why this is a peripheral-geography problem specifically is the key analytical contribution.
Water is a hyperlocal resource. Typically, this infrastructure is sized based on population growth, and if you have a large data center, their demand can easily exceed that of the local municipal population for future growth.
— Shaolei Ren, University of California, Riverside [48]
That sentence identifies the precise mismatch. Municipal water systems in peripheral counties are engineered against demographic projections. A five-gigawatt campus is not a demographic event. It is an industrial one, arriving in a system with no industrial design margin — which is exactly the condition that obtains across most of the Powered Periphery.
The research team’s recommendations translate directly into the Rural AI Bargain framework in Table 3: developers should report peak water use rather than annual averages; should partner with local communities to fund infrastructure upgrades with verifiable outcomes; should add enough capacity to offset their own use and preserve supplies for future community growth; and should work with utilities to shift between water-based and dry cooling depending on whether the water system or the power grid is more stressed at a given moment.[47] None of these is technically difficult. All of them require the transparency that nondisclosure agreements preclude.
7.4 The macroeconomic frame, and the bubble question
Two further considerations belong on the ledger, because a paper about the geography of an investment boom that ignored the possibility that it is a boom would be incomplete.
The first is that the AI infrastructure build-out has become macroeconomically significant in the United States in a way that complicates the local cost-benefit analysis. Harvard economist Jason Furman has calculated that information-processing investment drove approximately 92 percent of United States GDP growth in the first half of 2025, and the AI-infrastructure build-out contributed roughly 0.8 percent of GDP by early 2026.[69] A Virginia legislative analysis credited that state’s data centers with approximately $9.1 billion in GDP and 74,000 jobs annually, while also projecting that a typical Dominion Energy residential customer could see costs climb $14 to $37 per month by 2040 — driven, the analysis argued, not by data centers as such but by the difficulty of building infrastructure and the risk that fixed costs impose on other ratepayers.[69] Both halves of that finding are real.
The second is concentration risk, and it applies with particular force to the periphery. A county whose fiscal base, construction sector, and utility rate base all become dependent on a single campus with a single corporate counterparty has acquired an economic structure closer to a mining town than to a diversified regional economy. The circular-financing critique of the Nvidia–OpenAI–SB Energy structure is, from the perspective of a peripheral county, not an abstract question about capital markets. It is a question about whether the guarantor of a twenty-year lease will still exist in year twelve, and about who is responsible for a half-built ten-gigawatt generation program if it does not.
This is precisely why the reversibility and decommissioning items in Table 3 are not boilerplate. Pennsylvania’s own numbers — 100-plus proposals producing five fully permitted first phases — indicate that a substantial majority of announced projects will never be built.[32] A community that has rezoned land, committed water, and approved transmission for a project that evaporates has borne real costs for a benefit that never arrived.
7.5 The global dimension: who is left out of the Powered Periphery
A final entry on the ledger concerns the two-thirds of the world for which this entire debate is a distant one, and it complicates any triumphal reading of AI’s geographic democratization.
The World Bank’s World Development Report 2026: The Promise of Artificial Intelligence, released on 4 August 2026, is the institution’s first comprehensive assessment of AI’s implications for low- and middle-income countries. Its conclusions are notably at odds with the framing that dominates in wealthy economies. The report advises developing economies not to compete on large models or hyperscale data centers, and instead to adapt small, low-cost AI tools to local conditions while investing in the prerequisites: electricity generation and distribution, connectivity, local data, skills, and institutions.[50]
AI has thrown developing economies a lifeline, and they should seize it. … They do not need large models or big data centers to reap its benefits. But they must hurry: AI is spreading faster and is more context-specific than earlier general-purpose technologies like electricity and the internet.
— Indermit Gill, Senior Vice President and Chief Economist, World Bank Group [50]
The window to get this right is narrow. AI presents a once-in-a-lifetime opportunity to solve problems that have resisted solutions for generations.
— Gaurav Nayyar, Director, World Development Report 2026, World Bank Group [51]
The report’s underlying numbers explain the advice. Approximately 14.2 percent of jobs in high-income countries face potential automation risk from generative AI against 4.5 percent in low- and middle-income countries, while productivity could be improved in more than 16 percent of existing jobs in developing economies against more than 18 percent in advanced ones. In Sub-Saharan Africa, nearly one-third of rural schools still lack reliable electricity and more than two-thirds lack dependable internet.[50,51]
The implication for this paper is sobering and should be stated plainly. The Powered Periphery is a phenomenon of rich-country peripheries — of American parishes and counties, and of European regions inside the single market with functioning grids, enforceable contracts, and transmission systems built for twentieth-century heavy industry. Richland Parish is peripheral relative to Silicon Valley. It is not peripheral relative to the global distribution of electrical and institutional capacity. Africa accounts for less than one percent of global data-center capacity despite hosting 18 percent of the world’s population. The migration described in this paper redistributes AI infrastructure within advanced economies while leaving the far larger global asymmetry substantially intact — and in some respects reinforcing it, since every gigawatt of turbine, transformer, and accelerator capacity absorbed by a rural American campus is a gigawatt not available elsewhere.
Section 8 — What Have We Learned? Nine Pillars
The argument can be distilled into nine propositions. Each is stated as a claim, followed by the evidence that supports it and, where relevant, the limits of that support. Together they constitute what I believe can responsibly be asserted about the Powered Periphery on the evidence available as of 20 August 2026.
Pillar 1 — Power is becoming more important than proximity
The first lesson is the most fundamental. Artificial intelligence does not eliminate geography; it changes which geography matters.
The cloud era rewarded proximity to network hubs and to customers, because latency, interconnection, and enterprise relationships were the binding constraints. The frontier-AI era increasingly rewards proximity to scalable electricity, because power availability and interconnection timing are the binding constraints. Europe’s movement from an average of 29 miles to approximately 109 miles from major hubs is a physical manifestation of that change in the constraint set, and it is corroborated independently by DC Byte’s finding that eight of nine European gigawatt-plus projects are sited away from major cities and by JLL’s finding that greenfield development has risen from 8 to 39 percent of the pipeline.[1,3]
The global city remains important. But the power-rich periphery becomes strategically important alongside it. For AI training in particular, electrons increasingly outweigh commuting distance — and the industry’s own practitioners now say so on the record.[1]
Pillar 2 — Rural land is becoming an energy-compute asset, priced in megawatts
The second lesson is that artificial intelligence rewrites rural land valuation at the level of the unit of account.
Acreage becomes valuable not because of its size but because of what can be attached to it: transmission, generation, fiber, cooling, water, substations, expandable campus footprint, and permitting. A previously ordinary tract can acquire strategic value purely through its position within the energy system. JLL’s decision to price European land in euros per megawatt of IT load rather than per hectare is not a presentational convenience; it is the market’s own acknowledgment that the underlying asset has changed.[1] American appraisers report parcels transacting between $250,000 and $1 million per acre depending on the infrastructure available with the land.[61]
This introduces a genuinely new proposition into real-estate economics: the productive value of land increasingly equals the quantity of durable compute that can be energized upon it. The Powered Periphery is therefore not simply a rural-development story. It is the emergence of a new asset class at the intersection of property, electricity, and computational infrastructure — one for which the appraisal profession, the lending market, and the tax code do not yet have adequate instruments.
Pillar 3 — The Five-Layer AI Economy has acquired a physical geography
The third lesson is that the Five-Layer AI Economy cannot be understood purely as a conceptual stack.
Energy determines where Chips can operate. Chips determine the economics of Datacenters. Datacenters manufacture and serve Models. Models power Applications and Agents. Every layer may appear digital from the user’s perspective, but the lower layers are profoundly physical, and the physicality is increasing rather than diminishing with scale.
When billions of people interact with AI agents, they will experience intelligence as something weightless and instantaneous. Yet the machinery behind those interactions occupies enormous tracts of land, consumes electricity measured in gigawatts, depends on transformers weighing hundreds of tons with multi-year lead times, and is tied permanently to specific geographies. The paradox is exact: AI becomes more virtual at the top of the stack precisely because it becomes more industrial at the bottom.
Pillar 4 — At sufficient scale, compute stops following power and starts commissioning it
The fourth lesson is one that the 2026 evidence establishes more clearly than the 2024 evidence could.
The early Powered Periphery selected locations where electricity already existed — New Carlisle beside the Olive 345 kV substation being the archetype.[22] At sufficient scale, AI capital begins creating electricity geography around itself. Entergy is building more than seven gigawatts of gas generation for a single Louisiana campus, an increase of more than thirty percent in the state’s entire grid capacity.[18] SB Energy and SoftBank will build at least ten gigawatts of new generation to yield eight IT-gigawatts at PORTS-Pike.[23] Alpha Compute proposes to drill its own wells and generate behind the meter.[30] Cleanview data shows at least 57 off-grid plants proposed or under construction to serve individual data centers, totaling roughly 73,000 megawatts.[4]
Both forces now operate simultaneously, and the transition between them is the single most important structural development of the past eighteen months. It means that the supply of powered land is no longer fixed by the historical accident of where the grid was built. It also means that the environmental, air-quality, and cost-allocation consequences of AI growth are increasingly determined outside the regulated planning processes that were designed to govern them.
Pillar 5 — Rural communities have become AI stakeholders with real leverage
The fifth lesson is political, and it is the one most consistently underestimated by developers.
Communities hosting AI infrastructure have ceased to be passive locations on corporate site-selection maps. A county commission can affect the timing of billions of dollars of infrastructure. A state utility regulator can determine the economics of a hyperscale campus through a single tariff decision. A governor can accelerate or restrict an entire pipeline, as Pennsylvania demonstrated on 18 August 2026.[31,32] Residents can alter permitting outcomes. Landowners can refuse — and, as the Raudabaugh, Huddlestone, Bare, and Grosser cases show, some of them do refuse eight-figure offers.[59,60] Ratepayer groups can challenge utility investments. Environmental organizations can contest generation and water plans.
The AI race between the United States and China is discussed in Washington as a geopolitical competition. But many of the physical decisions determining American compute capacity are made in county commissions and state capitals far from Washington. The global AI race therefore possesses a profoundly local political foundation — and the actors at that foundation are only now discovering what they are holding.
Pillar 6 — The empirical record is contested, and the outcome is a policy variable
The sixth lesson is epistemic, and it is a corrective to the confidence with which both sides of this debate typically speak.
On electricity prices, three serious research programs disagree: Harvard’s legal-institutional analysis identifies systematic cost-shifting risk; EPRI’s instrumental-variables study finds a modest historical price decrease; Zhenxuan Wang’s facility-level study finds a 2.7 percent increase.[37,40,41] On employment, the best-identified study finds real but modest gains — 56 percent growth in data-processing employment and 43 percent in telecommunications over a decade, amounting to roughly 100 to 200 jobs in a typical county — with wages unchanged and home prices up 2 to 5 percent.[42]
The correct conclusion is not that the evidence is useless. It is that the outcome is institutionally determined rather than technologically determined. Whether a host community gains or loses depends on the structure of the electric service agreement, the existence of a large-load tariff, the presence of minimum demand commitments and exit provisions, the transparency of contracts, the design of incentives, and the enforceability of community benefits. Every one of these is a choice. None is a law of nature.
Bahar and Wright’s finding that state incentives constitute roughly 2 percent of investment in hyperscale counties but roughly 62 percent in colocation counties should reorganize state incentive policy entirely.[42] Subsidizing the facilities that were coming anyway, while the subsidy that actually changes behavior buys the weaker outcome, is not a defensible allocation of public money once it has been measured.
Pillar 7 — Water, not power, may be the binding constraint in many peripheral communities
The seventh lesson is the one most likely to be underestimated over the next five years.
Power constraints are visible, quantified, contested in public dockets, and increasingly addressed through behind-the-meter generation. Water constraints are hyperlocal, sized against demographic rather than industrial projections, and frequently invisible until a peak-demand event. Peak water demand for evaporatively cooled facilities can run six to thirty times the annual average; United States water systems may require $10 to $58 billion in new infrastructure by 2030 to accommodate data-center growth; and, as Ren observes, money cannot buy more snowpack.[47]
The policy prescription follows directly and is remarkably cheap to implement: require disclosure of peak rather than average water demand, name the source, require developer contribution to capacity exceeding the facility’s own use, and permit cooling-mode flexibility so that facilities can shift load between the water system and the power grid depending on which is stressed. A jurisdiction that adopts these four requirements has addressed most of the water risk at essentially no cost to itself.[47]
Pillar 8 — Community consent has become a financial variable, not a public-relations one
The eighth lesson follows from the seventh and the fifth, and it represents the clearest change in the operating environment between 2024 and 2026.
When Pennsylvania conditioned permit review on prior local land-use approval, it converted community consent from a soft constraint into a hard sequencing requirement with direct balance-sheet consequences.[34] When more than 100 communities enacted moratoriums and more than 300 state bills were filed in six weeks, the aggregate effect was to introduce a new category of project risk that did not exist in the underwriting models of 2023.[42] When the Harvard/MIT polling found that opposition is driven more by concerns about community change than by electricity prices, it established that the risk cannot be retired through rate design alone.[46]
For a developer, the implication is that consent is now an input to be secured early and expensively, in the same way transformers and turbines are secured early and expensively. For a community, the implication is that the leverage exists but is perishable: it is strongest before a project is announced and weakest after land has been optioned and expectations set.
Pillar 9 — The winning periphery will be powered, connected, and politically durable
The ninth and final lesson is that cheap land alone will not produce AI prosperity, and neither will cheap power.
The regions that succeed will combine several conditions simultaneously. They will possess sufficient and expandable electricity. They will possess assemblable land. They will possess high-capacity fiber. They will manage water responsibly and transparently. They will maintain reliable supporting infrastructure. They will train the technicians, electricians, and operators the facilities require rather than importing them. And they will establish political agreements strong enough to survive changing administrations, elections, community opposition, cost overruns, and economic cycles.
That last characteristic deserves particular emphasis, because it is the one least amenable to being purchased. A data center expected to operate for twenty years requires more than a building permit. It requires political durability — a settlement that a successor governor, a newly elected county commission, and a mobilized group of residents can all live with. The most successful Powered Periphery will therefore not necessarily be the region offering the fastest approval. It will be the region capable of constructing the strongest long-term compact among developer, utility, landowner, government, and community.
Speed is what developers ask for. Durability is what they actually need. The regions that understand the difference will capture disproportionate value from the next two decades of AI infrastructure; the regions that offer speed alone will discover that they have sold their leverage at the bottom of the market.
Conclusion: The Center of Artificial Intelligence May Be Moving to the Edge
For most of the digital era, economic geography seemed to reward concentration without limit. Software companies clustered in Silicon Valley. Financial technology concentrated around New York and London. Cloud infrastructure accumulated around Northern Virginia, Frankfurt, Amsterdam, and other network-rich metropolitan regions. Talent attracted capital; capital attracted companies; companies attracted infrastructure; infrastructure reinforced the city. The feedback loop appeared self-sustaining, and for three decades it was.
Artificial intelligence does not destroy that system. But it introduces a competing geographic force powerful enough to redraw substantial parts of it. That force is electricity.
Reuters’ 19 August 2026 report on Europe makes the transition unusually visible, and it is worth restating the finding once more in its full form because it is the evidentiary spine of this paper. Hyperscale facilities scheduled for 2026 through 2028 are being developed an average of approximately 109 miles from major hubs, against approximately 29 miles during the previous delivery period. Greenfield projects have risen from about 8 percent to about 39 percent of the pipeline. Inner-city projects are expected to fall from about 13 percent to about 5 percent. Eight of nine proposed gigawatt-plus developments are sited away from major cities, distributed from rural Spain to northern Sweden. And the price of powered land ranges from roughly €2.7 million per megawatt in Amsterdam to as little as €200,000 per megawatt around Bordeaux.[1,3]
Europe is documenting quantitatively what the United States has already begun experiencing physically. In Louisiana, Meta is transforming Richland Parish into a five-gigawatt training complex supported by an Entergy generation program that will expand the state’s grid capacity by more than thirty percent.[15,18] In Indiana, Amazon’s New Carlisle development is forcing policymakers and utilities to contemplate power requirements at a scale comparable to major generating facilities.[21] In Ohio, Meta’s Wood County project and the Nvidia-backed OpenAI campus at PORTS-Pike reveal how farmland, brownfields, generation, and compute can become parts of a single industrial system.[4,23] In Pennsylvania, developers are assembling sites where natural gas, generation, and data-center infrastructure sit on the same deed — even as Governor Shapiro’s executive order demonstrates that local consent and infrastructure protections are becoming binding constraints rather than courtesies.[29,31]
Different regulatory systems and energy markets produce different development patterns on the two continents. The underlying economic mechanism is nonetheless remarkably similar: AI infrastructure moves toward places where power can be secured at scale, at speed, and with sufficient room to expand.
This is why I chose the title Powered Periphery. The periphery consists of locations once considered secondary to the principal economic centers of the digital age. The qualifier powered explains why those locations are becoming central to the next phase — their advantage derives from electricity, land, water, infrastructure, expandability, and increasingly the political ability to host industrial-scale computation.
The term also resolves the apparent contradiction between the European and American evidence. Europe is now showing measurable outward movement away from its traditional metropolitan data-center hubs, documented in a dataset. The United States has already produced enormous examples in Louisiana, Indiana, Ohio, Pennsylvania, and elsewhere, documented in filings and construction sites. One record is statistical and one is physical, but they are records of the same phenomenon.
None of this means that London, Frankfurt, Amsterdam, Northern Virginia, or Silicon Valley become irrelevant. Inference, connectivity, corporate leadership, research, finance, and human talent will continue to concentrate in metropolitan economies. Live capacity in the FLAP-D markets has more than doubled since 2019 and continues to grow.[2] What is emerging is not replacement but a geographic division of labor: the metropolis remains a center of invention, capital, applications, and consumption; the periphery increasingly becomes a center of computational production.
That division may become one of the defining geographic characteristics of the Five-Layer AI Economy. The Energy Layer is pulling the Chip and Datacenter Layers toward new territories. Those facilities produce Models whose intelligence returns, through the Application and Agent layers, to billions of users located almost anywhere. What begins as electricity in rural Louisiana, northern Sweden, Indiana, or Ohio can appear milliseconds later as intelligence on a screen in Manhattan, London, Los Angeles, Singapore, or Tokyo.
That is the paradox at the center of this paper: artificial intelligence is becoming globally accessible because the infrastructure manufacturing it is becoming intensely local.
But the paradox carries an obligation, and it is where this argument should end rather than with the pleasing symmetry. The localness is not abstract. It is Breanne Kidd’s window in Middleton Township, and the turbine across the street that arrived without notice. It is Mervin Raudabaugh choosing $2 million and preservation over $15 million and conversion. It is a municipal water system in a rural parish sized for population growth and asked to supply an industrial peak. It is a county commission of five people holding, for a few weeks, decisive authority over an investment larger than the county’s entire history of capital formation. Whether the Powered Periphery becomes rural revitalization or rural extraction will be settled in those rooms, on those terms, by people whose names will not appear in any earnings call.
The next frontier of artificial intelligence may therefore not be found only in another laboratory, another model architecture, or another semiconductor. It may be found beside a transmission line. Behind a substation. Near a nuclear reactor. Above a natural-gas field. Across former industrial land. Or on acreage that only a few years earlier seemed impossibly distant from the center of the technology economy.
In the cloud era, distance from the city usually meant distance from digital power. In the AI era, that relationship is beginning to reverse.
The periphery is becoming powerful because the periphery is becoming powered.
Endnotes and Sources:
[1] Simon Jessop and Iain Withers, Reuters, “Europe AI Data Centres Seek Cheaper, Quicker Energy and Land,” 19 August 2026. Contains the JLL distance, greenfield, and powered-land figures and the quoted remarks of Assad Noori, Martin Jensen, and Rupert Duckworth. https://kelo.com/2026/08/19/europe-ai-data-centres-seek-cheaper-quicker-energy-and-land/
[2] Data Centre Magazine, “JLL Report: How AI Is Redrawing Europe’s Data Centre Map,” summarizing JLL’s EMEA Mid-Year Data Centre Report 2026, August 2026. https://datacentremagazine.com/news/jll-report-how-ai-is-redrawing-europes-data-centre-map
[3] Techzine Global, “AI Data Centers Seek Space Outside European Hotspots,” August 2026. Source for the inner-city share falling from 13% to 5% and the DC Byte gigawatt-project distribution. https://www.techzine.eu/news/infrastructure/143726/ai-data-centers-seek-space-outside-european-hotspots/
[4] Reuters, “Fast-Tracked Power Plants Fuel AI Boom, with Little Public Scrutiny,” 16 June 2026. Source for Breanne Kidd, the Apollo Generating Station, and the Cleanview count of 57 off-grid plants totaling 73,000 MW. https://www.yahoo.com/news/us/articles/fast-tracked-power-plants-fuel-090337075.html
[5] BG Independent News, “Power Plant for Meta Data Center in Wood County Fast Tracked — Approved by State with No Public Hearings,” 3 February 2026. https://bgindependentmedia.org/power-plant-for-meta-data-center-in-wood-county-fast-tracked-approved-with-no-public-hearings/
[6] Arman Shehabi, Sarah J. Smith, Alex Hubbard, Alex Newkirk, Nuoa Lei, Md Abu Bakar Siddik, Billie Holecek, Jonathan Koomey, Eric Masanet and Dale Sartor, 2024 United States Data Center Energy Usage Report, Lawrence Berkeley National Laboratory, LBNL-2001637, December 2024. https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report_1.pdf
[7] Lawrence Berkeley National Laboratory, United States Data Center Energy Usage Report: 2025 Update, U.S. Department of Energy, OSTI, 2026. https://www.osti.gov/biblio/3374245
[8] Fatih Birol, International Energy Agency, “AI Is Set to Drive Surging Electricity Demand from Data Centers While Offering the Potential to Transform How the Energy Sector Works,” accompanying the special report Energy and AI, April 2025. https://www.iea.org/news/ai-is-set-to-drive-surging-electricity-demand-from-data-centres-while-offering-the-potential-to-transform-how-the-energy-sector-works
[9] International Energy Agency, “Data Centre Electricity Use Surged in 2025, Even with Tightening Bottlenecks Driving a Scramble for Solutions,” April 2026. Contains Fatih Birol’s “energy taker … energy maker” remarks. https://www.iea.org/news/data-centre-electricity-use-surged-in-2025-even-with-tightening-bottlenecks-driving-a-scramble-for-solutions
[10] International Energy Agency, “Energy Demand from AI,” chapter of Energy and AI, 2025. https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai
[11] Stanford Institute for Human-Centered Artificial Intelligence, The 2026 AI Index Report, April 2026. https://hai.stanford.edu/ai-index/2026-ai-index-report
[12] United Nations University Campus Computing Center, “What the 2026 Stanford AI Index Report Tells Us About the State of AI,” 2026. Source for 5,400 U.S. data centers and 29.6 GW of AI data-center power capacity. https://c3.unu.edu/blog/2026-stanford-ai-index-report-takeaways
[13] Statista, “Big Tech’s AI Spending to Reach $760 Billion in 2026,” compiled from Q2 2026 earnings disclosures of Microsoft, Alphabet, Meta and Amazon. https://www.statista.com/chart/35046/capital-expenditure-of-meta-alphabet-amazon-and-microsoft/
[14] Tom’s Hardware, reporting Financial Times analysis, “Google, Microsoft, Meta, and Amazon Capex Spending to Hit $725 Billion in 2026,” April 2026. https://www.tomshardware.com/tech-industry/big-tech/big-techs-ai-spending-plans-reach-725-billion
[15] Jonathan Vanian, CNBC, “Meta’s Louisiana Data Center Investment to Reach $50 Billion, Aided by Generous Tax Incentives,” 13 July 2026. https://www.cnbc.com/2026/07/13/meta-louisiana-data-center-investment-reaches-50-billion-amid-ai-push.html
[16] Cris Tolomia, Quartz, “Meta Expands Louisiana Hyperion Data Center to 5 Gigawatts,” 13 July 2026. Source for the $1bn local infrastructure commitment, $1.6bn in local contracts, and the Louisiana Delta Community College scholarship program. https://qz.com/meta-louisiana-hyperion-data-center-expansion-5-gigawatts-071326
[17] ConstructConnect, “Meta Expands Louisiana Data Center to 5GW, Lifts Richland Parish Investment Above $50 Billion,” 14 July 2026. Includes U.S. data-center construction spending of $58.1bn year-to-date through May 2026. https://news.constructconnect.com/meta-expands-louisiana-data-center-to-5gw-lifts-richland-parish-investment-above-50-billion
[18] Global Data Center Hub, “Meta Scales Hyperion to 5GW and Over $50 Billion in Louisiana,” 20 July 2026. Source for the Entergy generation program, 240 miles of transmission, and the Blue Owl joint-venture structure. https://www.globaldatacenterhub.com/p/meta-scales-hyperion-to-5gw-and-over
[19] TechCrunch, “Gas Power Plants Approved for Meta’s $10B Data Center, and Not Everyone Is Happy,” 21 August 2025. https://techcrunch.com/2025/08/21/gas-power-plants-approved-for-metas-10b-data-center-and-not-everyone-is-happy
[20] Associated Press, “Amazon Cloud Computing Unit Plans to Invest $11 Billion to Build Data Center in Northern Indiana,” including remarks by Governor Eric Holcomb. https://www.barchart.com/story/news/25736721/amazon-cloud-computing-unit-plans-to-invest-11-billion-to-build-data-center-in-northern-indiana
[21] Government Technology, “Indiana Data Center Could Bring Significant Power Use,” April 2026. Source for the 2,250 MW estimate and the quoted remarks of Ben Inskeep of the Citizens Action Coalition. https://www.govtech.com/artificial-intelligence/indiana-data-center-could-bring-significant-power-use
[22] Measured AI, “AWS New Carlisle Data Center Campus,” July 2026. Source for the Olive 345 kV interconnection, the ≈$185m AEP transmission package, and I&M’s rate-impact analysis. https://measuredai.substack.com/p/aws-new-carlisle-data-center-campus
[23] NVIDIA Corporation, “NVIDIA Guarantees SB Energy’s PORTS-Pike Technology Campus in Ohio to Exclusively Host NVIDIA AI Compute,” press release, 17 August 2026. https://nvidianews.nvidia.com/news/nvidia-guarantees-sb-energy-s-ports-pike-technology-campus-in-ohio-to-exclusively-host-nvidia-ai-compute
[24] NVIDIA Corporation, Form 8-K, U.S. Securities and Exchange Commission, 17 August 2026. https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000069/sbeoainvidia-portsrelease.htm
[25] Reuters, via Nikkei Asia, “Nvidia to Provide Up to $105bn Guarantee for OpenAI’s Ohio Data Center,” 17 August 2026. https://asia.nikkei.com/business/technology/artificial-intelligence/nvidia-to-provide-up-to-105bn-guarantee-for-openai-s-ohio-data-center
[26] CDO Magazine, “Nvidia Backs OpenAI Ohio Data Center Lease,” August 2026. Contains the quoted remarks of Jensen Huang and the construction and operating employment figures. https://www.cdomagazine.tech/aiml/nvidia-backs-openai-ohio-data-center-lease
[27] Axios, “OpenAI Announces Massive Data Center in Ohio with Nvidia Guarantee,” 17 August 2026. https://www.axios.com/2026/08/17/openai-nvidia-ohio-data-center-sb-energy
[28] Unite.AI, “NVIDIA Guarantees Up to $105B for 8-GW Ohio AI Campus Leased by OpenAI,” August 2026. Includes the residual-value guaranty structure and the DOE March 2026 fact-sheet generation figures. https://www.unite.ai/nvidia-guarantees-up-to-105b-for-8-gw-ohio-ai-campus-leased-by-openai/
[29] Reuters, “Exclusive: Alpha Compute to Buy Pennsylvania Land, Gas Rights for $55 Million Data Center Campus, Company Says,” 11 August 2026, including remarks by chief executive Brittany Kaiser. https://whbl.com/2026/08/11/exclusive-alpha-compute-to-buy-pennsylvania-land-gas-rights-for-55-million-data-center-campus-company-says/
[30] Alpha Compute Corp., Form 6-K, Exhibit 99.1, U.S. Securities and Exchange Commission, 11 August 2026. https://www.sec.gov/Archives/edgar/data/0001095435/000117184326005422/exh_991.htm
[31] Governor Josh Shapiro, Commonwealth of Pennsylvania, “Remarks at Signing Ceremony for an Executive Order on Data Centers,” 18 August 2026. https://www.pa.gov/governor/newsroom/2026-press-releases/gov–shapiro-s-remarks-at-signing-ceremony-for-an-exec–order-on
[32] Commonwealth of Pennsylvania, Office of the Governor, “Governor Shapiro Signs Executive Order on Data Center Development in PA,” 18 August 2026. Contains the GRID Requirements and the Department of Environmental Protection project counts. https://www.pa.gov/governor/newsroom/2026-press-releases/governor-shapiro-signs-executive-order-on-data-center-developmen
[33] NBC News, “Gov. Josh Shapiro Issues a New Executive Order on Data Centers in Pennsylvania,” 18 August 2026. https://www.nbcnews.com/politics/2028-election/gov-josh-shapiro-executive-order-data-centers-pennsylvania-rcna593177
[34] Whitney Downard, Pennsylvania Capital-Star, “Gov. Shapiro Signs Data Center Executive Order. Critics Say It Falls Short,” 19 August 2026. Contains remarks by DEP Secretary Jessica Shirley and by Megan McDonough of Food & Water Watch. https://penncapital-star.com/technology-information/gov-shapiro-signs-data-center-executive-order-critics-say-it-falls-short/
[35] Bloomberg Law, “Shapiro Signs Order With Nation’s ‘Strictest’ Data Center Rules,” 18 August 2026. https://news.bloomberglaw.com/environment-and-energy/shapiro-signs-order-with-nations-strictest-data-center-rules
[36] The Philadelphia Inquirer, “Gov. Josh Shapiro Signs Executive Order Restricting Data Center Development in Pennsylvania,” 18 August 2026. https://www.inquirer.com/politics/pennsylvania/josh-shapiro-data-center-order-20260818.html
[37] Eliza Martin and Ari Peskoe, Extracting Profits from the Public: How Utility Ratepayers Are Paying for Big Tech’s Power, Harvard Law School Environmental and Energy Law Program, March 2025. https://eelp.law.harvard.edu/extracting-profits-from-the-public-how-utility-ratepayers-are-paying-for-big-techs-power/
[38] Ari Peskoe, quoted in PolitiFact, “How Much Have Data Centers Increased Electricity Prices?” 12 June 2026. https://politifact.com/factchecks/2026/jun/12/elizabeth-warren/data-centers-rising-electricity-costs/
[39] Ari Peskoe, Salata Institute for Climate and Sustainability, Harvard University, “The Data Center Boom Is Colliding with the Grid’s Hardest Problems,” 17 March 2026. https://salatainstitute.harvard.edu/data-centers-ai-artificial-intelligence-grid-permitting-transmission-electricity-energy
[40] Asa Watten, John Bistline and Geoffrey Blanford, “Have Data Centers Raised Your Electric Bill? Causal Evidence from the United States,” arXiv:2606.19777, June 2026. https://arxiv.org/abs/2606.19777
[41] Zhenxuan Wang, “Who Pays for Growth? Evidence from Data Centers and the Grid,” June 2026. https://www.zhenxuanwang.org/publications/data-center-elec-price/
[42] Dany Bahar and Greg Wright, “New Evidence on Data Center Employment Effects,” The Brookings Institution, updated 10 August 2026, summarizing “Data Centers and Local Labor Markets.” https://www.brookings.edu/articles/new-evidence-on-data-center-employment-effects/
[43] Daniel Goetzel, Mark Muro and Shriya Methkupally, “Turning the Data Center Boom into Long-Term, Local Prosperity,” The Brookings Institution, 5 February 2026. https://www.brookings.edu/articles/turning-the-data-center-boom-into-long-term-local-prosperity/
[44] Anthony F. Pipa and Adam Aley, “The Local Implications of Data Centers for Rural Communities in the US,” The Brookings Institution, 2 March 2026. https://www.brookings.edu/articles/local-implications-data-centers-rural-communities-us/
[45] Nicol Turner Lee and Darrell M. West, “Why Community Benefit Agreements Are Necessary for Data Centers,” The Brookings Institution, 29 January 2026. https://www.brookings.edu/articles/why-community-benefit-agreements-are-necessary-for-data-centers/
[46] Stephen Ansolabehere, Harvard University, reported by Axios, “Exclusive Poll: Here Are the Concerns About Data Centers,” Harvard/MIT survey, 3 April 2026. https://www.axios.com/2026/04/03/data-centers-concerns-ai-electricity-harvard-mit
[47] Shaolei Ren, Yuelin Han, Pengfei Li and Adam Wierman, “Small Bottle, Big Pipe: Quantifying and Addressing the Impact of Data Centers on Public Water Systems,” reported by University of California, Riverside, March 2026. https://news.ucr.edu/articles/2026/03/09/data-center-water-spikes-could-cost-billions
[48] Shaolei Ren, quoted in E&E News by POLITICO, “Thirsty Data Centers Fuel Local Angst over Water Infrastructure,” April 2026. https://www.eenews.net/articles/thirsty-data-centers-fuel-local-angst-over-water-infrastructure/
[49] Pengfei Li, Jianyi Yang, Mohammad A. Islam and Shaolei Ren, “Making AI Less ‘Thirsty’: Uncovering and Addressing the Secret Water Footprint of AI Models,” arXiv:2304.03271. https://arxiv.org/abs/2304.03271
[50] World Bank Group, World Development Report 2026: The Promise of Artificial Intelligence, press release including remarks by Chief Economist Indermit Gill, 4 August 2026. https://www.worldbank.org/en/news/press-release/2026/08/04/ai-offers-lifeline-to-developing-economies-in-an-era-of-weak-growth
[51] Gaurav Nayyar, World Bank Group, reported by Agence France-Presse / France 24, “World Bank Warns Developing Countries to Embrace AI or Be Left Behind,” 4 August 2026. https://www.france24.com/en/live-news/20260804-world-bank-warns-developing-countries-to-embrace-ai-or-be-left-behind
[52] The White House, Executive Order, “Accelerating Federal Permitting of Data Center Infrastructure,” 23 July 2025. https://www.whitehouse.gov/presidential-actions/2025/07/accelerating-federal-permitting-of-data-center-infrastructure/
[53] The White House, “Fact Sheet: President Donald J. Trump Unveils the Genesis Mission to Accelerate AI for Scientific Discovery,” 24 November 2025. https://www.whitehouse.gov/fact-sheets/2025/11/fact-sheet-president-donald-j-trump-unveils-the-genesis-missionto-accelerate-ai-for-scientific-discovery/
[54] Bipartisan Policy Center, “Strategic Federal Actions Aim to Strengthen AI and Energy Infrastructure,” updated 15 May 2026, including the Ratepayer Protection Pledge and Executive Order 14318 thresholds. https://bipartisanpolicy.org/explainer/strategic-federal-actions-aim-to-strengthen-ai-and-energy-infrastructure/
[55] Monitoring Analytics, Independent Market Monitor for PJM, reported by Utility Dive, “Data Centers Were 40% of PJM Capacity Costs in Last Auction: Market Monitor,” 7 January 2026. https://www.utilitydive.com/news/data-centers-pjm-capacity-auction/808951/
[56] Institute for Energy Economics and Financial Analysis, “Projected Data Center Growth Spurs PJM Capacity Prices by Factor of 10.” https://ieefa.org/resources/projected-data-center-growth-spurs-pjm-capacity-prices-factor-10
[57] Enel North America, “PJM 2026/2027 Capacity Auction Results,” July 2025. https://www.enelnorthamerica.com/insights/blogs/pjm-2026-2027-capacity-auction-results
[58] American Farm Bureau Federation, “Balancing Data Center Growth with American Agriculture,” Market Intel, 2026. Source for the section 1031 exchange mechanism and USDA farm electricity expenditure forecasts. https://www.fb.org/market-intel/balancing-data-center-growth-with-american-agriculture
[59] Chris Bennett, AgWeb, “Land Rush: Data Center Stampede Puts Farmers in Crosshairs of Controversy,” July 2026. Source for the Raudabaugh, Huddlestone and Bare cases. https://www.agweb.com/news/business/farmland/land-rush-data-center-stampede-puts-farmers-crosshairs-controversy
[60] Andy Castillo, Farm Progress, “Data Center Payouts Are Huge. Saving Farmland? Priceless,” 20 January 2026. Source for the Tim Grosser case and the global data-center land footprint. https://www.farmprogress.com/conservation-and-sustainability/data-center-payouts-are-huge-saving-farmland-priceless
[61] Daniel Boring, American Society of Farm Managers and Rural Appraisers, reported by RFD-TV, “Data Centers Reshape Farmland Values Across the Southeast,” August 2026. https://www.rfdtv.com/data-centers-reshape-farmland-values-across-the-southeast
[62] STL Partners, “The EU’s AI Gigafactory Initiative: What It Means for Digital Infrastructure,” May 2026. Source for Stargate Norway and the Start Campus at Sines. https://stlpartners.com/articles/data-centres/eu-ai-gigafactory-initiative/
[63] European Commission, “AI Factories,” Directorate-General for Communications Networks, Content and Technology, 2026. https://digital-strategy.ec.europa.eu/en/policies/ai-factories
[64] European Economics, “InvestAI Initiative and the AI Gigafactories Call,” June 2026, including Council Regulation (EU) 2026/150 and the public-private funding ratio. https://www.europeaneconomics.com/en/investai-initiative-ai-gigafactories/
[65] Avanza Energy, “The $176 Billion Detour: How Europe’s Data Centers Built Three Regulatory Workarounds When the Grid Said No,” June 2026. Source for the AWS EMEA remarks, Ember secondary-market projections, and the ENTSO-E warning of 30 April 2026. https://avanzaenergy.substack.com/p/the-176-billion-detour-how-europes
[66] Datacentre Review, “AI Boom Pushes Data Centres Away from Europe’s Major Hubs,” August 2026, including remarks by Eva Sóley Guðbjörnsdóttir of atNorth. https://datacentrereview.com/2026/08/ai-boom-pushes-hyperscale-data-centres-away-from-europes-major-hubs/
[67] Deeptech.build, “The €37 Billion Race: EU AI Gigafactories Europe and the Sovereignty Bet,” April 2026, including ACER data on the EU–U.S. industrial electricity price gap. https://www.deeptech.build/content/gigafactories-deep-tech
[68] Jesse Jenkins, Princeton University, quoted by Associated Press, “Meta Signs Three Nuclear Power Deals to Help Support Its AI Data Centers,” 2026. https://finance.yahoo.com/news/meta-signs-three-nuclear-power-135710396.html
[69] American Enterprise Institute, “The Data Center Backlash Is Really About Abundance,” August 2026. Source for the Jason Furman GDP-growth calculation and the Virginia JLARC findings. https://www.aei.org/commentary/the-data-center-backlash-is-really-about-abundance/
[70] Senator Bernie Sanders and Representative Alexandria Ocasio-Cortez, “Sanders, Ocasio-Cortez Announce AI Data Center Moratorium Act,” 25 March 2026. https://www.sanders.senate.gov/press-releases/news-sanders-ocasio-cortez-announce-ai-data-center-moratorium-act/
[71] Michael Blackhurst and colleagues, Carnegie Mellon University, “Data Center Growth Could Increase Electricity Bills 8% Nationally and as Much as 25% in Some Regional Markets,” 16 July 2025. https://www.cmu.edu/work-that-matters/energy-innovation/data-center-growth-could-increase-electricity-bills
[72] Fortune, “Data Centers Have Already Hiked Electricity Prices on the Public by $23 Billion. Good Luck Clawing That Back,” 14 July 2026. https://fortune.com/2026/07/14/data-centers-23-billion-electricity-bills/
[73] Marketplace, “Data Centers Lowered Electric Bills in Some Places — For Now,” 10 July 2026. https://www.marketplace.org/story/2026/07/10/data-centers-lowered-electric-bills-in-some-places-for-now
A note on method. This paper draws on primary corporate disclosure (SEC Forms 8-K and 6-K, company press releases, and earnings guidance through the second-quarter 2026 reporting season), primary government sources (executive orders, state utility commission filings, gubernatorial orders and remarks, and Department of Energy laboratory reports), peer-reviewed and working-paper economics (Harvard Law School, the Electric Power Research Institute, Brookings, Carnegie Mellon, the University of California, Riverside, and the California Institute of Technology), institutional analysis (the International Energy Agency, the World Bank Group, and the European Commission), and contemporaneous reporting from Reuters, the Associated Press, CNBC, Bloomberg Law, Axios, and specialist trade and agricultural press. Where the empirical literature is contested — most notably on retail electricity price effects — competing findings are presented alongside one another in Section 7 rather than reconciled to a single conclusion, and the limits of each research design are stated.
A note on figures. All charts were prepared by the author from the figures reported in the cited sources. Figure 1 and Figure 2 use JLL data as reported by Reuters on 19 August 2026. Figure 3 uses company capital-expenditure disclosures compiled by Statista and the Financial Times. Figure 4 uses Lawrence Berkeley National Laboratory’s 2024 report and 2025 update. Table 2’s final column is an illustrative arithmetic extension by the author and is not a market quotation; very large sites are not priced by linear extrapolation. Tables 1, 3, 4 and 5 are the author’s own frameworks, synthesized from the sources indicated in their captions.
