A foundation-building analysis of the AI data-center buildout in mid-2026, explaining how the electric grid, capital markets, local government finance, and US regulatory structure each work, and what each shows about the cost of supplying AI infrastructure at the pace of the announcements.
Safe AI AcademyMay 30, 202640 min read3 views
In the four years following the November 2022 public release of ChatGPT, the world's largest technology companies have collectively committed more capital to building the physical infrastructure required to train and operate artificial intelligence systems than was committed during the same period to new single-family housing construction in the United States. That infrastructure consists, in physical terms, of large purpose-built buildings full of high-performance computer chips, the electrical systems required to power those chips, the cooling systems required to keep them from overheating, and the land, water rights, and transmission lines required to support them. The buildout is referred to in industry shorthand as "AI infrastructure," and four years in it has begun to produce measurable effects on the systems that surround it.
Three terms are used throughout this article and are useful to define at the start:
AI data center. A purpose-built facility containing rows of computer servers fitted with specialized AI chips, the networking equipment that connects them, the electrical infrastructure that powers them, and the liquid or air cooling systems that prevent them from overheating. A typical 2026 AI campus draws between 100 and 800 megawatts of electricity continuously, which is comparable to the consumption of a small city.
GPU (graphics processing unit). A specialized computer chip originally designed for video-game graphics, now repurposed for AI because the matrix arithmetic that GPUs are built to perform is the same kind of arithmetic that AI training and inference require. The current high-end model used in AI training, the Nvidia H100, draws approximately 700 watts at full load. A single frontier-model training run uses tens of thousands of GPUs in coordinated operation.
Hyperscaler. Industry shorthand for the small number of technology companies that operate cloud-computing infrastructure at very large scale and that are, accordingly, the principal builders and customers of AI data centers. The most frequently cited US hyperscalers are Alphabet (Google), Microsoft, Amazon, Meta, and Oracle.
The bull case for the buildout, advanced by the companies funding it and the investors holding their securities, is that AI capability will continue to compound, that demand for AI services will continue to grow, and that the data centers being built will, over time, generate operating cash flow commensurate with the capital being deployed to build them. The bull-case data points are still in force as of late May 2026. Combined fiscal-year-2026 capital expenditure guidance from Alphabet, Microsoft, and Meta totals approximately $459 billion, with Amazon signaling an additional $200 billion. Nvidia, the dominant supplier of the GPUs the data centers contain, reached a $5 trillion market capitalization in October 2025 and reported approximately $81.6 billion of revenue for the single fiscal quarter ending in late April 2026.
Stay Updated
Get notified when we publish new articles and course announcements.
The subject of this analysis is what the buildout has cost the systems that surround it during the same period. Those costs are documented across four domains, each of which has its own working mechanism that the buildout has placed under pressure: the electric grid that powers the data centers, the capital markets that fund the data-center buildout itself, the local-government tax base that hosts the campuses, and the federal and state regulatory structure that authorizes them. Each section below first explains how the relevant system normally works, then identifies what is specific about the way AI data centers interact with that system, then presents what the evidence shows about the resulting cost, and then sets out what the convergence of evidence implies for what comes next.
How the electric grid works, and why AI data centers have stressed it
Electricity, unlike most goods, cannot be stored at the scale at which it is consumed. The amount of electricity being generated must equal the amount being used at every instant, or the system fails. This basic physical requirement determines how the electric grid in North America is organized, operated, and paid for.
Bulk electric system. The interconnected network of high-voltage transmission lines, power plants, and substations that delivers electricity from where it is produced to where it is consumed. The contiguous United States is divided into three large synchronous regions (the Eastern Interconnection, the Western Interconnection, and the Texas Interconnection) that operate independently from one another.
To match generation to consumption at every moment, the bulk electric system is overseen by regional transmission organizations and independent system operators (collectively, RTOs/ISOs), which are nonprofit entities responsible for coordinating the dispatch of power plants across multi-state regions. The largest such organization in the United States is the Pennsylvania-New Jersey-Maryland Interconnection, known as PJM.
Regional transmission organization (RTO). A nonprofit entity that operates the wholesale electricity market for a multi-state region and dispatches power plants in real time to balance supply with demand. RTOs do not own power plants or transmission lines themselves; they coordinate the participants who do.
PJM Interconnection. The largest RTO in the United States, serving approximately 65 million people across thirteen states (Pennsylvania, New Jersey, Maryland, Delaware, Ohio, Virginia, West Virginia, North Carolina, Indiana, Kentucky, Michigan, Tennessee, Illinois) plus the District of Columbia. PJM operates the wholesale energy market, the capacity market, and the transmission-system planning processes for its region.
PJM and other RTOs operate two distinct wholesale electricity markets. The energy market pays generators for the electricity they actually produce, settled at locational marginal prices in dollars per megawatt-hour. The capacity market pays generators in advance for a binding commitment to be available during peak demand, regardless of whether they are dispatched in that period. The two markets are economically distinct, and the capacity market is the one that has become directly relevant to the AI buildout.
Capacity market. The wholesale market in which generators are paid in advance for the binding commitment to be available during peak demand events. PJM's capacity market is called the Base Residual Auction. Capacity prices clear in dollars per megawatt-day, where one megawatt of guaranteed availability for a single day defines the unit. The capacity market exists because most generators would not be economic to maintain on energy-market revenue alone (peak-demand periods are rare but reliability-critical), and the capacity payment is what makes maintaining the peak-period reserve financially sustainable.
The wholesale prices that clear in those markets reach residential electricity bills through state rate-case proceedings. Investor-owned utilities (the regulated companies that deliver electricity to most US residential customers) recover their wholesale costs through retail electricity rates approved by state public utility commissions. The allocation of those wholesale costs across residential, commercial, and industrial customer classes is determined by the formulas approved during each rate case.
Rate case. A formal proceeding in which a regulated investor-owned utility requests approval from its state public utility commission to charge particular retail electricity rates, supported by evidence of the utility's wholesale costs, capital investments, and demand forecasts. Rate cases determine how new wholesale-cost burdens are allocated among the residential, commercial, and industrial customer classes.
With this foundation in place, the AI-specific story becomes legible. AI data centers are the largest new electrical load class to appear on the US grid in approximately fifty years. A single hyperscale AI campus can consume between 100 and 800 megawatts of electricity continuously, which is comparable to the consumption of a small city and several orders of magnitude larger than the largest residential customer or even most traditional industrial customers. The arithmetic of AI training is what makes this consumption pattern unusual: tens of thousands of GPUs operating in coordinated synchrony, each drawing roughly 700 watts, with very little of the variability that traditional industrial loads exhibit between shifts.
The visible signal of how this load is interacting with the grid is the cleared capacity-market price. The PJM Base Residual Auction for the 2026 to 2027 delivery year cleared at $329.17 per megawatt-day, more than ten times the $28.92 per megawatt-day at which the same auction cleared for the 2024 to 2025 delivery year. A price cap negotiated with the office of Pennsylvania Governor Josh Shapiro prevented the auction from clearing higher, and that cap is set to expire before the next auction. According to PJM's 2025 Load Report, peak demand on the PJM grid is projected to grow by approximately thirty-two gigawatts between 2024 and 2030, with all but approximately two of those gigawatts attributable to data-center load. New generation is being added to the PJM system at approximately two to three gigawatts per year, which does not close the projected gap. PJM has publicly stated that, as of summer 2026, its grid has "just enough" capacity to meet its reliability requirements.
The downstream effect on residential customers operates through the rate-case mechanism. According to reporting by the Oregon Capital Chronicle, residential electricity rates at Portland General Electric and Pacific Power (the two largest investor-owned utilities in Oregon) are more than 50 percent higher in 2026 than they were in 2020, after six consecutive years of approved rate increases driven principally by industrial-class data-center load growth. Over the same six-year period, residential electricity demand in Oregon grew approximately ten percent while industrial-class demand grew approximately seventy percent. The arithmetic of cost allocation explains the rest. The new wholesale-cost burden, driven principally by data-center demand growth, has been allocated to residential customers in proportions larger than residential customers' contribution to the demand growth.
The reliability dimension of the AI load has produced a separate signal, distinct from the price signal. On the night of July 10, 2024, a lightning arrestor failed on a 230 kilovolt transmission line in northern Virginia's data-center corridor. The fault itself was electrically routine, the kind that bulk-system operators handle multiple times per year without consequence. The response from the AI data centers in the corridor was not. Approximately sixty data-center facilities automatically transferred to backup power within seconds of the disturbance, producing a near-instantaneous loss of approximately 1,500 megawatts of load from the bulk electric system. According to the incident review conducted by the North American Electric Reliability Corporation, this was the largest simultaneous load loss from a single fault that NERC had ever recorded.
NERC (North American Electric Reliability Corporation). The nonprofit organization, with federal regulatory authority delegated by the Federal Energy Regulatory Commission, responsible for setting and enforcing reliability standards for the bulk electric system across the United States, Canada, and a portion of Mexico. NERC's alerts (graduated Level 1 through Level 3) communicate increasing degrees of urgency about emerging reliability risks.
The technical reason this event matters for AI infrastructure specifically is a mismatch between how the bulk electric system was designed and how AI data centers behave. The grid is engineered around the assumption that disturbances primarily originate on the generation side. When a generator fails, system frequency falls below the 60 Hz standard, and grid operators have a century of practice managing the recovery. When load drops at comparable magnitude (as it does when AI data centers detect a disturbance and self-trip to backup power as a class), frequency rises above 60 Hz, and the operational protocols for managing that condition are less developed. The historical reason the protocols are underdeveloped is that no individual customer class was previously large enough to trigger this condition. AI data-center campuses, with their electrically sensitive power-conditioning equipment, now are.
NERC's posture has shifted in stages in response. NERC issued a Level 2 alert on the topic during 2025 and subsequently concluded that registered transmission entities had, in NERC's own language, "generally lacked adequate processes" to respond to it. On May 4, 2026, NERC escalated to a Level 3 "Essential Actions" Alert, the highest urgency category that NERC issues. The Level 3 alert directs transmission planners to gather granular electrical performance data from large data-center customers, to install dynamic fault recorders to characterize how data-center facilities actually behave during system disturbances, and to require large loads to "ride through" short voltage and frequency dips rather than self-tripping en masse. Acknowledgements from registered entities were due by May 11, 2026; substantive responses are due by August 3, 2026.
A typical 2026 AI campus draws power from two sources and exports its costs in two directions: combustion emissions to fence-line residents adjacent to on-site generation, and capacity-price pass-through to grid-connected households.
Capacity-market clearing prices at ten times their 2024 level, residential rate trajectories that have outpaced inflation by a factor of two for six consecutive years, and NERC's escalation to its highest-tier alert describe the same underlying condition. The bulk electric system is being asked to serve a load class for which it was not engineered, and the rate at which new generation is being added is not closing the gap. The next two delivery years of PJM capacity-market clearing prices, the August 2026 industry response window on the Level 3 alert, and the outcomes of pending state rate cases on cost allocation between residential and industrial customer classes are likely to determine whether the constraint is absorbed through tariff and rate-class reform or whether it becomes physically binding and forces interconnection-queue gating for new data-center customers.
How the AI buildout is funded, and why announced capacity has not become built capacity
The large public companies funding the AI data-center buildout are doing so through capital expenditure rather than operating expense. The distinction matters because capital expenditure is accounted for differently than year-to-year spending, and because the scale of the capital expenditure is the most direct measure of how much physical capacity is being built.
Capital expenditure (capex). Spending by a company on long-lived physical assets that are expected to produce revenue over multiple future years. Capex is recorded on the balance sheet as an asset rather than as a current-year expense, and is then converted to expense gradually over the asset's useful life through depreciation. Examples include constructing buildings, purchasing manufacturing equipment, and, in the AI context, building data centers and buying GPUs.
Depreciation. The accounting process by which the cost of a long-lived asset is allocated as expense across the asset's expected useful life, rather than recognized in the year the asset was purchased. The useful-life assumption directly determines the annual depreciation expense and, accordingly, reported earnings: a shorter useful life produces a larger annual expense and lower reported earnings, all else equal.
For the AI buildout, hyperscaler capex covers the chips, the real estate, the electrical infrastructure, the cooling equipment, the networking equipment, and the construction labor required to bring new data-center capacity online. Combined fiscal-year-2026 capex guidance from Alphabet, Microsoft, and Meta totals approximately $459 billion. Amazon has separately signaled approximately $200 billion of additional 2026 capex. The aggregate is within the same order of magnitude as the gross domestic product of South Africa.
A separate question is whether and how that capex is being recovered through operating cash flow. As of late May 2026, approximately four years after ChatGPT's consumer release, no major frontier AI laboratory has demonstrated durable profitability on the underlying AI technology itself. The companies that build and operate AI-facing systems, including OpenAI, Anthropic, and the AI subsidiaries of the hyperscalers themselves, are still consuming capital faster than they generate it. The companies that have made reliable profits from the buildout are not the AI labs but their hardware suppliers, principally Nvidia.
Picks-and-shovels model. A pattern, frequently observed in commodity booms, in which the suppliers of the equipment used by the boom's participants capture more reliable returns than the participants themselves. The term originates from the California Gold Rush, in which most prospectors did not strike gold but the merchants supplying their picks and shovels did. In the AI buildout, the picks-and-shovels layer is the chip and electrical-infrastructure supply chain (principally Nvidia, plus the companies that build transformers, switchgear, and cooling systems). The "prospecting" layer is the frontier AI laboratories themselves.
This is the structural context against which to read the announced-versus-built capacity question. The companies funding the buildout have committed publicly to building specific amounts of data-center capacity by specific dates. Whether they are actually doing so is independently measurable, and the measurements have produced a gap.
Sightline Climate, a market-intelligence firm specializing in energy-infrastructure tracking, finds that of approximately sixteen gigawatts of US data-center capacity announced for 2026 delivery, only approximately thirty to fifty percent is actually under construction. For 2027 delivery, approximately 6.3 gigawatts is under active construction against approximately 21.5 gigawatts announced. The Sightline definition of "under construction" includes any project from a poured foundation through final fit-out, which is generous to the announcement side. Independent journalism by Ed Zitron surfaced the Sightline data set in early 2026. Heatmap Pro, which tracks announced US data-center projects, reports that 99 of approximately 770 tracked projects are currently contested by local opposition, that approximately twenty-five projects were canceled outright in 2025, and that an additional approximately twenty to twenty-five projects were canceled during the first quarter of 2026, representing approximately $41.7 billion in shelved investment in that quarter alone. In March 2026, Bloomberg reported that Oracle and OpenAI had dropped a 600 megawatt expansion at the Stargate data-center campus in Abilene, Texas, the single largest AI-infrastructure project announced to date. Microsoft subsequently agreed to rent the dropped capacity, which softens the headline at the system level but does not change the underlying signal that the project as originally announced is not being built as announced.
Announced 2027 US AI data-center capacity versus capacity actually under construction as of early 2026 (Sightline Climate).
The supply-chain dynamics that explain the announced-versus-built gap are also visible at the upstream end of the chain. Nvidia's chief executive Jensen Huang stated in September 2025 that the company had shipped approximately ten gigawatts of GPUs over the course of 2025, which he equated to four to five million chips. Goldman Sachs Research estimated at approximately the same time that the total operational AI data-center capacity worldwide was approximately 7.7 gigawatts. The discrepancy is interpretable through a pattern that operations-management literature calls the bullwhip effect.
Bullwhip effect. A pattern in supply chains in which buyers downstream of a supply constraint place orders well in excess of their immediate need, because the marginal cost of stockpiling is lower than the marginal cost of being unable to fulfill demand later. The buildup of inventory across the chain produces order signals to upstream suppliers that overstate actual end demand. The pattern is named for the way small movements at the handle of a bullwhip produce large movements at the tip.
Symptoms consistent with the bullwhip effect are visible across the AI buildout's supply chain. The International Energy Agency reports that wait times for grid transformers and cables have doubled in the past three years. Two fully constructed data-center facilities near Nvidia's Santa Clara headquarters reportedly remain offline pending utility interconnection. Nvidia's own inventory balance roughly doubled year-over-year to approximately $21 billion at the end of fiscal year 2026.
A related question concerns the assumptions under which the GPUs already deployed in hyperscaler data centers are being depreciated for accounting purposes. Because depreciation expense depends on the useful-life assumption (defined above), the useful-life assumption that hyperscalers apply to their GPUs directly affects their reported earnings. In November 2025, investor Michael Burry (known for the pre-2008 short positions documented in The Big Short) argued publicly that the hyperscalers are understating GPU depreciation by booking useful lives of five to six years when the economic useful life of a GPU in the current AI market is closer to two to three years. Burry's analysis estimated approximately $176 billion of understated depreciation across hyperscaler accounts between 2026 and 2028, with Oracle's earnings overstated by approximately twenty-seven percent and Meta's by approximately twenty-one percent by the end of that window. Burry characterized the situation as parallel to Cisco's late-1990s overbuild rather than as parallel to Enron's fraud. Nvidia issued a public rebuttal defending the longevity of its own chips, which Burry characterized as a category error on the grounds that the depreciation argument concerns economic obsolescence rather than physical operation: the question is not whether an older chip can still run, but whether it remains cost-competitive against newer generations that perform the same work at roughly one-third the power draw.
The combination of record upstream profits, expanding announced capacity, contracting actually-built capacity, doubling supplier inventory, and contested customer-side depreciation assumptions describes an investment cycle in which the supply-side response has visibly outrun the conditions required to validate it at the cash-flow level. The next two earnings cycles, specifically the hyperscaler disclosures around GPU useful-life assumptions and the trajectory of Nvidia's inventory balance against customer payment terms, are likely to determine whether the cycle resolves into a credible profitability path for the downstream operators or forces a partial writedown of currently-capitalized hardware.
How local government finances data centers, and what the fiscal record shows
When a hyperscaler decides to site a new AI data center, the location decision is influenced not only by the availability of land, water, and electricity but also by the incentive package offered by the host state and municipality. These incentive packages combine three principal instruments that draw on the way US local government is financed. To make the next portion of this analysis legible, those instruments need to be defined.
Property tax. A tax assessed annually on the value of real property (land and buildings) and, in some jurisdictions, on tangible personal property such as equipment. Property tax is the single largest source of own-source revenue for US local governments and is the principal revenue stream for school districts in most states.
Property-tax abatement. A negotiated reduction or elimination of the property-tax liability of a specific facility for a defined period (frequently ten to thirty years). Abatements are typically granted in exchange for projected employment, capital investment, or other economic activity that the host jurisdiction expects the facility to generate.
Sales-tax exemption. A provision exempting a specific category of equipment purchases (frequently including data-center servers and electrical equipment) from state and local sales tax. For multibillion-dollar equipment purchases, the value of the exemption can be substantial.
Enterprise zone. A designated geographic area within which businesses meeting defined criteria receive a package of tax benefits, typically including both property-tax abatements and sales-tax exemptions. Enterprise zones were introduced in many states as economic-development tools in the 1980s and 1990s, with eligibility criteria that have since been adjusted in many states to capture data-center investment.
In the AI-buildout context, these instruments operate as transfers from the local tax base to the operating company in exchange for projected employment and economic activity. Whether the projected employment and activity actually materialize at the levels used to justify the transfers has been the subject of multiple empirical analyses in 2025 and 2026.
Texas is a particularly testable case because it offers both property-tax abatements and sales-tax exemptions, publishes its incentive data with relative transparency, and contains a sufficient number of approved and rejected data-center projects to support a difference-in-differences analysis. Michael Hicks, director of the Center for Business and Economic Research at Ball State University, published such an analysis in late 2025 examining employment outcomes across Texas counties that approved data-center projects against matched counties that did not. His conclusion, stated directly, was that "no data center had a measurable effect on employment." The new positions documented during data-center construction and operation in his sample appeared to be predominantly transfers from other construction and information-sector activity that would have taken place regardless, rather than incremental employment. On that basis, Hicks concluded that the incentive packages cannot be justified on the employment-creation grounds typically cited in their approval.
The Hicks finding is contested at the local level. A Taylor County, Texas commissioner who approved an 80 percent property-tax abatement covering Lancium, Oracle, and OpenAI projected four hundred to twelve hundred jobs at completion of the Stargate campus. The realized employment counts are not yet final. The track record of comparable projections across other states is what frames the academic economic literature's skepticism toward the Stargate projection as a default position rather than as an unusual one.
The fiscal costs are more straightforward to measure than the employment benefits, because foregone tax revenue is recorded directly. A July 2025 Good Jobs First analysis found that Oregon's public schools lost approximately $275 million in 2024 to corporate property-tax abatements, with approximately three-quarters of the enterprise-zone savings allocated to data-center operators and the Hillsboro School District absorbing the largest single share of the loss. Regional reporting documents a parallel set of secondary effects in Hillsboro: farmland converted to data-center sites at a faster rate than farmers can purchase replacement land, locally-anchored manufacturer Intel conducting layoffs totaling approximately 2,400 positions in the same window, and concurrent housing-cost pressure on residents whose wages have not increased in step. In Abilene, Texas, where the Stargate campus is under construction, the published economic data for the most recent quarter shows hotel revenue up sharply year-over-year, short-term-rental occupancy up approximately forty percent, near-full apartment occupancy, and residential rents up approximately twenty percent year-over-year, with a substantial share of the construction workforce traveling in from out of state.
A single AI data-center campus produces five distinct downstream effects in its host community, four of which are costs.
The environmental dimension of the buildout becomes specific at the facility level, where AI data-center operators have increasingly chosen to power their campuses with on-site generation rather than rely entirely on the public grid. This pattern reflects the long interconnection-queue waits described in the previous section. To make the next portion of the analysis legible, the relevant term needs to be defined.
Behind-the-meter generation. Electrical generation owned and operated by the customer, located on the customer's own property, downstream of the utility's electric meter. Because the electricity produced behind the meter does not pass through the utility's distribution system, it does not require interconnection-queue clearance and is not subject to wholesale market rules. It is, however, subject to federal Clean Air Act standards as administered by state environmental agencies.
The Texas air permit for the Stargate Abilene AI data-center campus, TCEQ permit 177263, authorizes ten gas turbines (a mix of Solar Titan 350 and General Electric LM2500 units) producing approximately 360 megawatts of on-site generation, supplemented by 62 emergency backup generators producing approximately 170 megawatts of additional capacity. The permit allows the turbines to burn diesel fuel for up to 200 hours per twelve-month period.
The xAI Colossus data-center cluster near Memphis, Tennessee, presented a parallel case under a different regulatory interpretation. xAI initially operated twenty-seven trailer-mounted natural-gas turbines at its Southaven, Mississippi site without Clean Air Act stationary-source permits, on the theory that turbines designated as "temporary" or "portable" were exempt from those permit requirements as "nonroad engines." In January 2026, the Environmental Protection Agency finalized New Source Performance Standards under 40 CFR Part 60 subpart KKKKa that explicitly state stationary combustion turbines, including those described as temporary or portable, require Clean Air Act permits.
New Source Performance Standard (NSPS). A category of federal air-quality regulation issued by the EPA that establishes emissions limits for newly built or modified stationary sources of air pollution. NSPS rules are codified at 40 CFR Part 60 and are enforceable through state environmental agencies. The January 2026 NSPS finalization for stationary combustion turbines was the action that removed xAI's legal basis for treating its Southaven turbines as exempt.
The EPA's fact sheet for the final rule noted that the units had never actually been exempt as "nonroad engines," characterizing the prior practice as an interpretation the federal regulator was declining to continue. Following the January 2026 finalization, xAI applied to permanently install forty-one gas turbines at the Southaven site, representing approximately 1.2 gigawatts of generation capacity. The Mississippi Environmental Quality Permit Board approved that application on March 10, 2026. The NAACP, represented by the Southern Environmental Law Center, filed suit in federal court in April 2026 over the twenty-seven turbines that had operated without permits prior to the rule finalization. The Southern Environmental Law Center separately filed an appeal of the permit-board approval of the forty-one permanent turbines. Fence-line residents in Southaven, including Krystal Polk, whose multigenerational family home is adjacent to the site, reported in interviews with Mississippi Free Press that they did not learn of the plant's planned operation until after the turbines were running on the property.
Across the cases documented, the distribution of costs and benefits from AI data-center development is structurally consistent. The property-tax abatement is local and immediate. The foregone school-district revenue is local and persistent. The projected employment lift remains projected rather than realized in matched-county empirical analyses. The environmental costs are concentrated in the specific neighborhoods that host the facilities, frequently outside the public review processes that would ordinarily attend a generating facility of comparable scale. Municipalities considering further data-center incentive packages during the next budget cycle will be evaluating them against a measurable empirical record from similar packages in similar towns, which changes the political and fiscal calculus under which the next round of approvals takes place and is likely to produce both increased scrutiny in jurisdictions still negotiating and an increased rate of organized opposition in jurisdictions where construction is already underway.
How the US regulatory response is layered, and what 2026 has produced
US regulation of large infrastructure projects is divided across three layers of government, each operating under its own statutory authority. Federal agencies set baseline standards under federal statutes such as the Clean Air Act and the Federal Power Act. State legislatures pass laws governing siting, taxation, and utility regulation within their borders. State public utility commissions implement state utility law through rate-case proceedings and tariff orders. Local governments (counties, municipalities, school districts) control zoning, property-tax assessment, and incentive packages. The AI data-center buildout is regulated by all three layers simultaneously, and through the first five months of 2026 each layer has produced its own response.
State public utility commission (PUC). A state-level regulatory body responsible for overseeing investor-owned utilities operating within the state. PUCs approve retail electricity rates through rate-case proceedings, authorize specific tariff structures, and order utilities to procure particular categories of generation. State PUCs are the principal venue through which state utility law is implemented.
The state legislative response has materialized at scale during the first five months of 2026. State legislative tracking by the consulting group MultiState documents that more than three hundred state-level bills addressing data-center development were filed across more than thirty states during the first six weeks of 2026, a volume that exceeds the entire prior decade in several substantive categories. The substance of those bills divides into four broad categories. The first category contains moratoriums on new data-center construction above a defined capacity threshold. The second contains special "large-load" rate classes designed to allocate the costs of new grid interconnections to the data-center customers that drive them, rather than to residential ratepayers. The third contains rollbacks of existing property-tax abatement and enterprise-zone benefits for data-center facilities. The fourth contains reporting and environmental-review requirements that the industry has historically been exempt from.
At least eleven states filed moratorium bills covering data-center facilities above twenty megawatts: Maryland, Vermont, Virginia, New York, Oklahoma, South Dakota, Wisconsin, Michigan, Minnesota, South Carolina, and Georgia. Maine came closest to enacting a statewide moratorium. The Maine legislature passed LD 307, which would have established a moratorium on data centers with capacity exceeding twenty megawatts. Governor Janet Mills vetoed the bill on April 24, 2026. The Maine legislature attempted to override the veto on April 29, 2026 and fell short of the required margin. At least eighteen additional states filed special rate-class legislation of the second category. Beyond the formal legislative record, citizen-organization tracking indicates that local opposition groups have blocked or delayed approximately $60 billion in announced US data-center projects over the preceding twelve months.
The 2026 sequence of federal regulatory, state legislative, and litigation milestones in the response to the AI data-center buildout.
The federal regulatory response has progressed in parallel with the state legislative response. The January 2026 EPA NSPS finalization, described in the preceding section, removed the legal basis for the "temporary turbine" classification that xAI and other operators had relied on to defer Clean Air Act permitting on behind-the-meter generation. The May 4, 2026 NERC Level 3 alert, described in the first section, reframed AI data-center load behavior from an emerging reliability risk into a formal planning obligation across all NERC-registered transmission entities. On May 12, 2026, the Oregon Public Utility Commission, implementing a state large-load law passed by the Oregon legislature in 2025, ordered Portland General Electric and Pacific Power to establish a tariff structure under which data-center customers pay rates that reflect the costs they impose on the system, to require renewable-energy procurement for new data-center loads before those loads come online, and to impose exit fees on data-center projects that are abandoned mid-construction. Several other state public utility commissions are studying the Oregon order as a template.
The litigation track is also active. The NAACP federal-court suit against xAI's Southaven operations and the Southern Environmental Law Center's appeal of the Mississippi permit-board approval of the forty-one permanent turbines are both in active litigation as of late May 2026. Cancellation tracking by Heatmap Pro indicates approximately twenty-five US data-center project cancellations in 2025 and approximately twenty to twenty-five additional cancellations in the first quarter of 2026, which represents approximately a fourfold increase in cancellation rate against the prior multi-year average. Water consumption has emerged as the single most-cited concern in cancellation cases, displacing the air-quality and noise concerns that had been most prominent in 2024 and 2025. Land-mortgage pricing near announced campuses has begun to incorporate cancellation risk as a non-trivial probability rather than as a tail-risk exception.
The institutional response now visible across federal regulators, state legislatures, state public utility commissions, federal courts, and credit markets describes a convergence in which independent actors operating under different statutory mandates have arrived at compatible assessments of the AI buildout's external costs within a compressed twelve-to-fifteen-month window. The convergence reduces the probability that the cost picture is being misread by any single class of actor, and increases the probability that the next phase of the AI buildout will be conducted under a different regulatory cost structure than the one the original announcements were drafted to assume.
Where the costs and returns now sit
The conditions described in the preceding sections do not amount to a prediction that the AI buildout will contract sharply. The bull-case data points remain in force. Nvidia reported revenue of approximately $81.6 billion for its first fiscal quarter of 2027, the quarter ending in late April 2026, a year-over-year increase of approximately 85 percent, with data-center revenue of approximately $75.2 billion. Goldman Sachs Research continues to forecast global AI data-center demand growing from approximately 31 gigawatts in 2025 to approximately 66 gigawatts by 2027. A substantial body of analyst commentary interprets the announced-versus-built capacity gap as a supply-chain bottleneck rather than as absent demand. Michael Burry's depreciation analysis was framed as overbuild rather than fraud. An accurate reading of the mid-2026 picture holds the cost evidence and the bull-case evidence in the same analytic frame.
What 2026 has produced is the documentation of the cost side of the buildout in sufficient detail that the cost side has become legible to the same audiences that follow the announcement side. Capacity-market clearing prices at ten times their 2024 level are not interpretable by PJM as a market anomaly. Twenty-seven gas turbines operating without Clean Air Act permits are not interpretable by the NAACP as a regulatory technicality. Two hundred and seventy-five million dollars of foregone Oregon school funding from data-center abatements in a single year is not interpretable by Oregon school districts as a routine pass-through. Nvidia inventory doubling against extending customer payment terms is not interpretable by audit-minded investors as a sign of unambiguous downstream demand. The AI buildout's measurable economy of consequences has expanded to roughly the scale of its measurable economy of upstream profits, while its projected economy of downstream returns at the laboratory and applications layer remains substantially projected.
The question of whether the next phase of the AI buildout closes the gap between cost and return through profit realization, through cost-allocation reform, through supply-side capital-market discipline, or through some combination of these mechanisms, is the question that the next eighteen to twenty-four months of data will resolve. The data that will determine the resolution are already being generated, in PJM capacity-market auctions, in hyperscaler quarterly disclosures, in Federal Energy Regulatory Commission and EPA dockets, in state legislative calendars, and in the active case docket of the United States District Court for the Northern District of Mississippi.
Comments
0 commentsBe the first to leave a comment.
Leave a comment
Posted a comment before?