Introduction: The Day the Datacenter Became Measurable
On September 21, 2026, two of America’s largest technology economies sent strikingly different messages about the same emerging problem, and in doing so they marked, perhaps more clearly than any single event before them, the moment at which the artificial-intelligence datacenter ceased to be a private industrial object and became a public economic quantity that governments intend to measure before they permit it to grow. The coincidence of timing was remarkable, but the convergence of substance was more remarkable still, because the two states arrived at the same underlying anxiety through political cultures, regulatory institutions, and energy systems that could hardly be more different from one another.
In California, Governor Gavin Newsom signed seven bills establishing what his office described as the most comprehensive datacenter oversight framework in the nation, a package addressing electricity consumption, water use, land use, utility-rate structures, grid-upgrade costs, environmental review, and the information available to local communities considering large projects [1]. The seven measures reveal how much the policy conversation has changed in only a few years. AB 1577, authored by Assemblymember Rebecca Bauer-Kahan, establishes new datacenter reporting requirements. AB 2383, by Assemblymember Rick Chavez Zbur, addresses electricity service for large energy-use facilities. AB 2469 and AB 2619, both by Assemblymember Diane Papan, require water-use disclosures before local governments may approve new or expanded facilities and mandate reporting of estimated or actual water sources and consumption. SB 886, the California Technology Innovation and Ratepayer Protection Act by Senators Steve Padilla and Jerry McNerney, together with SB 1168 on datacenter rate structures, directs the California Public Utilities Commission to create new power rates for datacenters that cover the cost of connecting the facilities to the grid and supplying them with electricity, so that those costs are not silently transferred to households. SB 887 removes datacenters from blanket environmental-review exemptions under the California Environmental Quality Act while offering streamlined review to projects that meet state standards for water and energy conservation [1, 2].
The political framing was unambiguous. Senator Padilla, who wrote SB 886 and SB 887, characterized the package in terms that placed cost allocation and community voice at the center of the state’s approach:
“These are some of the nation’s strongest data center ratepayer protections.”
— Senator Steve Padilla (D-San Diego), author of SB 886 and SB 887 [3]
Nor did the legislation emerge in a vacuum. A July 2026 poll by the Public Policy Institute of California found that nearly three-quarters of Californians opposed the construction of new datacenters in their communities, a level of public resistance that helps explain why a governor who had vetoed similar measures in prior sessions chose, in September 2026, to sign all seven [4]. California’s package therefore does not treat a datacenter merely as a building requiring a permit. It treats the datacenter simultaneously as an electricity load, a water customer, a land user, an infrastructure participant, an employer, and an economic-development project whose consequences must be visible before the project proceeds.
Several hours later and roughly fifteen hundred miles to the east, Texas provided a very different illustration of the same underlying transition. Governor Greg Abbott directed the Texas Commission on Environmental Quality to halt all air and water permits sought by datacenter projects until the Electric Reliability Council of Texas and the Texas Water Development Board complete comprehensive audits of the sector’s electricity consumption, water use, tax incentives, community impacts, and ownership structures, an escalation of the interconnection moratorium the governor had first ordered in early August [5, 6]. The governor’s own words compressed the state’s verification philosophy into a single sentence:
“Data centers must pay their own way, protect our grid and water.”
— Governor Greg Abbott of Texas, directive to the TCEQ, September 21, 2026 [5]
The numbers behind the Texas decision explain its urgency. By August 2026, interconnection-queue requests before ERCOT totaled approximately 474 gigawatts of proposed large loads, a figure the governor himself noted was more than five times the record peak electricity demand ever recorded on the ERCOT system, with roughly ninety percent of the new power requests coming from datacenters [7]. A queue of requested capacity is emphatically not the same thing as capacity that ultimately will be built, and much of that 474 gigawatts represents duplicated, speculative, or exploratory applications. But that is precisely the point: a grid operator cannot plan billions of dollars of transmission and generation around demand it cannot verify, and ERCOT’s own leadership had already conceded the inadequacy of its legacy procedures. As Kristi Hobbs, ERCOT’s Vice President of System Planning and Weatherization, told the grid operator’s board in December 2025:
“We have outgrown the process that was established for reviewing these large loads.”
— Kristi Hobbs, Vice President of System Planning and Weatherization, ERCOT [8]
California and Texas therefore arrived at the same question through different institutional routes. California’s approach emphasizes disclosure, utility cost allocation, environmental review, and local information. Texas’s approach emphasizes verification, audits, resource impacts, and temporary restrictions while regulators determine which proposed loads are credible and what demands they may impose on the system. Neither approach by itself defines a national model. Together, however, they illuminate an emerging question that may become central to the next phase of the Five-Layer AI Economy: how should society measure an artificial-intelligence datacenter before approving the infrastructure necessary to serve it?
Until recently, the answer seemed relatively straightforward. A datacenter could be described by acreage, construction cost, server capacity, jobs, tax revenue, and megawatts of electrical demand. Hyperscalers measured GPU capacity. Investors measured capital expenditure. Utilities measured peak load. Developers measured available land. Governors and economic-development agencies measured investment commitments and employment. Artificial intelligence is making those measurements insufficient, because the scale of the current buildout has grown to the point where the datacenter’s claims on the surrounding economy dwarf the building itself. A 500-megawatt AI campus is not simply a 500-megawatt building. Behind those 500 megawatts may sit new substations, transmission lines, transformers, generation contracts, backup generators, water infrastructure, roads, emergency services, workforce requirements, tax incentives, and years of utility planning. Its economic footprint can extend hundreds of miles from the physical campus and decades beyond the construction schedule. The megawatt visible at the meter may therefore be only the final unit in a much larger resource chain.
That distinction becomes increasingly important as the scale of AI infrastructure grows. Lawrence Berkeley National Laboratory’s 2025 Update, released in June 2026 under lead author Arman Shehabi, estimates that United States datacenters could consume approximately 649 terawatt-hours of electricity by 2030, representing 11.8 percent of forecasted total U.S. electricity use, with scenario bounds ranging from roughly 9.5 percent to 15.3 percent depending on hardware efficiency, utilization, and adoption assumptions [9]. The laboratory’s researchers were careful to explain why efficiency alone will not arrest this growth, observing in the report that:
“the scale and growth of computational demand more than offsets”
— Lawrence Berkeley National Laboratory, United States Data Center Energy Usage Report: 2025 Update, on why hardware efficiency gains do not reduce total consumption [10]
The International Energy Agency reaches a parallel conclusion from the global vantage point. Its landmark report Energy and AI projects that worldwide datacenter electricity demand will more than double by 2030 to roughly 945 terawatt-hours, slightly more than the entire present electricity consumption of Japan, and that in the United States datacenters are on course to account for almost half of all growth in electricity demand between now and 2030, by which point the American economy is expected to consume more electricity for processing data than for manufacturing aluminum, steel, cement, and chemicals combined [12]. IEA Executive Director Fatih Birol framed the stakes plainly:
“AI is one of the biggest stories in the energy world today.”
— Dr. Fatih Birol, Executive Director, International Energy Agency [12]
Yet electricity is only the beginning. Electricity requires generation, and generation may require fuel, renewable capacity, nuclear output, batteries, or other resources. Electricity delivery requires substations and transmission. Cooling can require water or alternative thermal systems. Large campuses require land. Construction requires workers, equipment, and materials. Operations require backup generation, telecommunications, and emergency-response capacity. Tax incentives create fiscal commitments. Roads, public safety, and municipal infrastructure generate local obligations. If any of these resources is scarce, its opportunity cost matters as much as its nominal price. The datacenter therefore increasingly resembles not merely a piece of digital infrastructure but a bundle of resource claims, and this is where the concept of Datacenter Measures becomes useful.
The next phase of AI infrastructure policy may not be defined by a single federal datacenter law, a single environmental standard, or a single electricity tariff. Instead, it may emerge from dozens of measurements gradually being attached to each project: contracted megawatts, actual megawatts, peak megawatts, water withdrawal, consumptive water use, land footprint, new generation, transmission upgrades, substations, backup generation, tax incentives, construction employment, permanent employment, emergency-service requirements, and infrastructure costs. These measurements could eventually become as important to an AI campus as financial statements are to a corporation. A company cannot adequately be understood merely by knowing its revenue; investors examine liabilities, assets, cash flows, capital commitments, and risks. Likewise, a datacenter cannot necessarily be understood merely by knowing its announced megawatts. The meaningful question may become: what resources must exist behind those megawatts, who provides them, who finances them, who bears their risks, and what happens if projected demand does not materialize?
This question already reaches well beyond California and Texas. Virginia, home to the largest datacenter concentration on earth, unveiled its Data Center Accountability Framework on September 18, 2026, when Governor Abigail Spanberger signed Executive Order 22 implementing sweeping transparency, environmental, and energy-cost directives and creating a rapid-response task force on artificial intelligence [19, 20]. Pennsylvania, through Governor Josh Shapiro’s Executive Order 2026-05 of August 18, 2026, removed all datacenter projects from the state’s Fast Track permitting program, prohibited nondisclosure agreements between developers and public agencies, and conditioned environmental permitting on legally binding commitments to affordability, transparency, workforce, and environmental standards [22, 23]. At the federal level, the U.S. House of Representatives passed the Ratepayer Protection Act, H.R. 9340, on September 16, 2026, by the extraordinary margin of 417 to 3, directing state utility commissions to consider standards under which datacenters drawing more than 100 megawatts would pay the full incremental cost of serving their loads; the bill’s ultimate disposition in the Senate remains a separate legislative question [13, 14]. And international policy is moving in the same informational direction: on the very same September 21, the European Commission proposed disclosure rules requiring datacenters above 500 kilowatts of capacity to report their energy and water efficiency through a common EU rating and labelling scheme, including how their water use relates to local water stress, illustrating that the measurement of digital infrastructure is becoming a broad policy movement rather than an exclusively American debate [17, 18].
The central proposition of this paper is therefore not that every state should adopt California’s approach, Texas’s approach, or any single regulatory formula. It is that the AI infrastructure boom is entering an era in which measurement precedes allocation. Before utilities allocate billions of dollars to transmission and generation, regulators increasingly want credible demand forecasts. Before communities allocate land and water, officials increasingly seek information about resource intensity. Before governments award incentives, policymakers increasingly examine public costs alongside anticipated revenues. Before lenders finance infrastructure, they increasingly ask how firmly power has been contracted and who bears stranded-asset risk. And before the next gigawatt of AI capacity is built, the fundamental policy question increasingly may be not simply whether a datacenter can be constructed, but rather what that datacenter requires from the surrounding economic system. That is the emerging domain of Datacenter Measures.
Why I Chose the Title “Datacenter Measures”
I chose Datacenter Measures because the word measures deliberately carries two meanings that together capture the transition now occurring around AI infrastructure, and because no other single word so precisely joins the institutional and the quantitative dimensions of the story this paper tells. The first meaning is institutional: government measures, which is to say laws, regulations, audits, rate structures, disclosure rules, permitting requirements, and planning procedures. California’s seven-law package and Texas’s audit-and-permit approach demonstrate two different forms of these governmental measures. Virginia’s executive framework, Pennsylvania’s two-track permitting order, Ohio’s precedent-setting datacenter tariff, and the European Commission’s labelling proposal are still others. Rather than implying that one model will become universally dominant, Datacenter Measures provides an umbrella broad enough to examine how different jurisdictions are responding to the same physical challenge through different legal and regulatory institutions.
The second meaning is quantitative. AI infrastructure increasingly must be understood through measures of electricity, water, land, transmission capacity, generation, tax incentives, workforce, public infrastructure, and community effects. A project described simply as “500 MW” conceals most of the economic system required to make those megawatts usable, in the same way that a corporation described simply by its revenue conceals its balance sheet. Datacenter Measures therefore describes a possible evolution from counting datacenters and announced megawatts toward measuring their complete resource footprint. The title fits the paper because its subject is not regulation alone and not resource consumption alone; it is the convergence of the two. Governments are creating new measures precisely because the AI economy itself increasingly needs to be measured, and the industry, in turn, is discovering that verified measurement may be the price of continued permission to build.
Section 1: The Invisible Resource Content of One Megawatt
1.1 The Megawatt Is No Longer a Sufficient Unit of AI Infrastructure
The first task of any serious accounting framework is to dismantle the apparently simple phrase “one megawatt of datacenter capacity,” because the phrase has become the industry’s dominant unit of self-description at exactly the moment when it conceals more than it reveals. A megawatt describes electrical power, but it does not reveal what physical and financial infrastructure made that power deliverable, nor what claims its delivery placed upon the surrounding region. Two datacenters drawing identical loads can impose very different requirements on their host systems depending on geography, existing grid capacity, water availability, generation mix, redundancy standards, transmission constraints, and operating profile. A megawatt served from surplus capacity on an uncongested network in a water-rich region is an almost trivially different economic object from a megawatt that requires a new combined-cycle gas plant, a new 765-kilovolt line, a new substation, and evaporative cooling drawn from a drought-stressed aquifer, even though both appear identical on a press release.
This paper therefore introduces the concept of the AI Megawatt Resource Envelope: the full set of physical, financial, and public commitments that must exist for one megawatt of information-technology load to operate reliably. The purpose of the envelope is not to claim that every megawatt requires the same quantities; it is precisely the opposite. The relevant measurements vary enormously by location and architecture, and it is exactly that variance which a single headline number suppresses.
| Component of the Envelope | What Must Exist Behind 1 MW of IT Load | Why It Varies by Project |
| Generation capacity | Firm and intermittent supply, typically 1.1–1.5 MW of system capacity per MW of IT load once cooling, conversion losses, and reserve margins are included | Depends on regional reserve margins, generation mix, and whether the load is flexible or must run continuously |
| Transmission capacity | Lines, switchgear, and network upgrades able to deliver the power to the site | Depends on congestion, distance from generation, and voltage class required |
| Substation and transformer capacity | Step-down capacity dedicated or shared at the point of interconnection | Long transformer lead times make this a binding constraint in many regions |
| Cooling requirement | Water withdrawal and consumption, or the electricity penalty of dry cooling | Evaporative cooling in arid regions versus air or liquid cooling in temperate ones |
| Backup capacity | Diesel or gas generators, batteries, and fuel storage sized to carry the facility through outages | Redundancy standard (N+1, 2N) and local air-permit constraints |
| Land requirement | Building footprint plus substations, generator yards, water infrastructure, corridors, and setbacks | Campus design, security zones, and on-site generation choices |
| Public infrastructure | Roads, fire protection, hazardous-materials response, and emergency planning | Local government capacity and facility scale |
| Financial commitments | Utility capital, developer capital, tax incentives, and cost-allocation mechanisms | Tariff design, contract structure, and stranded-asset protections |
1.2 From IT Load to System Load
Datacenter announcements typically emphasize information-technology capacity, which is to say the power delivered to the servers themselves, but the surrounding facility also consumes electricity for cooling, pumps, fans, networking, power conversion, and supporting infrastructure. The industry’s conventional efficiency metric, Power Usage Effectiveness, captures the ratio between total facility power and IT power, and modern hyperscale facilities have driven PUE impressively close to its theoretical floor. Yet for the purposes of Datacenter Measures, PUE is only one layer of measurement, because the more consequential question sits upstream of the facility entirely: how much energy-system capacity must ultimately be planned, financed, and constructed in order to deliver one unit of productive AI compute? The answer includes reserve margins the grid operator must hold against the new load, losses across the transmission and distribution network, and the generation headroom required to serve a facility whose load factor may approach ninety percent around the clock. The distinction becomes economically decisive when hundreds of megawatts are aggregated into a single campus and multiple campuses cluster on the same utility system, because at that point the difference between IT load and system load is no longer a rounding error but a multi-billion-dollar planning parameter.
The energy intensity of the workloads themselves compounds the issue. Researchers at the Massachusetts Institute of Technology have documented that generative AI is not simply more computing of the familiar kind but a categorically heavier class of demand. As Dr. Noman Bashir, Computing and Climate Impact Fellow at the MIT Climate and Sustainability Consortium, has explained:
“a generative AI training cluster might consume seven or eight times more energy”
— Dr. Noman Bashir, MIT Climate and Sustainability Consortium and CSAIL, comparing generative AI training clusters to typical computing workloads [33]
1.3 The Generation Behind the Meter
Electricity demand cannot be analyzed independently from electricity supply, and one of the defining characteristics of the present buildout is the extraordinary diversity of supply arrangements now being assembled behind AI campuses. A facility may be supported by some combination of existing grid generation, new natural-gas plants, nuclear generation, renewable power-purchase agreements, batteries, fuel cells, onsite generation, geothermal energy, or emerging small modular reactor technologies, and each combination carries a distinct profile of cost, emissions, reliability, and timing risk. The relevant measurement is therefore not simply annual electricity consumption. It is also incremental generating capacity required; dispatchability; capacity factor; the timing of production relative to datacenter demand; the balance between firm and intermittent supply; backup requirements; fuel infrastructure; and the expected life of the generation asset relative to the expected life of the computing demand it serves. A thirty-year gas plant financed against a datacenter whose chips will be obsolete in five years embodies a temporal mismatch that deserves explicit measurement, because someone must bear the risk of that mismatch, and history suggests it will not always be the developer. Datacenter Measures thus converts “electricity use,” a consumption statistic, into a generation obligation, a planning and financing commitment extending decades into the future.
1.4 The Transmission and Transformer Multiplier
Generation located somewhere on the system does not automatically mean electricity can reach a datacenter, and it is in the wires, transformers, and substations that the hidden capital intensity of the AI boom becomes most visible. Transmission lines, switchgear, transformers, and distribution infrastructure must physically connect generation with load, and these assets are precisely the ones with the longest lead times, the most contested siting processes, and the most complicated cost-allocation questions. As Dr. Deepjyoti Deka, research scientist at the MIT Energy Initiative and program manager of its Data Center Power Forum, has observed of the fundamental geographic problem:
“the wires may not have sufficient capacity to carry the electricity”
— Dr. Deepjyoti Deka, Research Scientist, MIT Energy Initiative, on why adequate generation elsewhere on a grid does not guarantee deliverability [32]
The scale of the resulting utility capital programs is without modern precedent. American Electric Power, one of the nation’s largest transmission owners, raised its five-year capital plan to 78 billion dollars in May 2026, up six billion dollars in a single quarter, driven by newly approved transmission investments in PJM and SPP and new gas-fired generation in Indiana, with transmission alone accounting for roughly 33 billion dollars, or forty-two percent, of the plan [25]. By mid-2026 the company reported approximately 63 gigawatts of expected incremental contracted load by 2030, nearly ninety percent of it tied to datacenters, sitting atop an active interconnection queue of roughly 190 gigawatts that includes a ten-gigawatt SB Energy campus in Piketon, Ohio, and a multibillion-dollar Google development in Putnam County, West Virginia [26]. AEP’s chairman and chief executive Bill Fehrman described the moment with corporate understatement:
“substantial demand growth across our footprint, particularly from data centers”
— Bill Fehrman, Chairman, President and CEO, American Electric Power, Q1 2026 earnings release [25]
Under Datacenter Measures, grid infrastructure therefore becomes measurable on a project basis through a simple causal chain: datacenter megawatts imply substation megawatts, which imply transmission upgrades, which imply generation requirements, which imply capital investment, and each link of that chain deserves its own line in the project’s resource statement, because each link has its own cost, its own schedule, and its own answer to the question of who pays.
1.5 Water as a Computational Input
Water enters the AI infrastructure equation through multiple channels simultaneously: through evaporative cooling systems at the facility itself, through the cooling requirements of the thermal power plants that generate the facility’s electricity, and through the water embedded in the production of that electricity across the supply chain. The crucial measurement discipline is therefore to distinguish between water withdrawal, which is the volume taken from a source, and water consumption, which is the volume evaporated or otherwise removed from the local hydrological cycle; to distinguish between cooling technologies such as evaporative versus air cooling; to capture seasonal and peak-day requirements rather than annual averages alone; and to situate every gallon within the water stress of the surrounding watershed, because a gallon consumed in a water-rich jurisdiction is not economically identical to a gallon consumed in a drought-stressed region.
The academic literature has moved decisively on this question over the period from 2023 to 2026, led substantially by Professor Shaolei Ren’s group at the University of California, Riverside. Ren and his co-authors Pengfei Li, Jianyi Yang, and Mohammad A. Islam estimated in their influential study, published in final form in Communications of the ACM, that global AI demand could account for 4.2 to 6.6 billion cubic meters of water withdrawal annually by 2027, a volume on the order of half the total annual freshwater withdrawal of the United Kingdom [31]. In March 2026, Ren’s group, working with Caltech professor Adam Wierman, published a further study estimating that the water infrastructure required to serve American datacenter growth could cost between 10 billion and 58 billion dollars, and recommending that developers report peak water use rather than annual averages, that companies fund community water-infrastructure upgrades with verifiable outcomes, and that facilities coordinate cooling methods with grid conditions [30]. Ren’s summary of the field’s blind spot deserves quotation, because it identifies precisely the measurement gap this paper seeks to close:
“water is a hidden and even more binding constraint in many communities”
— Professor Shaolei Ren, Bourns College of Engineering, University of California, Riverside [30]
California’s new framework explicitly elevates water information into datacenter planning, with AB 2469 preventing local approval of new or expanded facilities until developers disclose projected water use and AB 2619 requiring operators to report estimated or actual water sources and consumption, while the ratepayer measures require developers to bear specified infrastructure-upgrade costs [1, 2]. The European Commission’s September 2026 proposal reaches the same conclusion from a different direction, requiring qualifying facilities to disclose water-use efficiency and its relationship to local water-stress levels [17].
1.6 Land Is Part of the Compute Stack
Land requirements extend well beyond the building footprint, and any accounting that measures only the acreage under the server halls will systematically understate the industrial geography of AI. A mature hyperscale campus may require land for the datacenter buildings themselves; for substations; for generator yards and their fuel storage; for battery installations; for cooling equipment; for water treatment and storage infrastructure; for transmission corridors approaching the site; for roads and truck access; for warehouses and logistics; for security zones and setbacks; and, increasingly, for dedicated on-site generation whose own footprint can rival that of the computing facility. The physical footprint of AI is therefore considerably larger than the server hall, and it accretes across a region rather than sitting neatly inside a fence line. The Five-Layer AI Economy begins conceptually with energy and chips, but at hyperscale it increasingly becomes an exercise in industrial geography, in which the siting of compute is constrained less by fiber and latency than by the availability of contiguous land parcels adjacent to high-voltage infrastructure and adequate water.
1.7 The Workforce Hidden Inside the Megawatt
Datacenters are famously capital-intensive and employment-light in steady-state operation, and critics of public incentives frequently point to the modest permanent headcounts of facilities representing billions of dollars of investment. Yet the infrastructure supporting them requires substantial skilled labor during construction, and this labor is drawn from exactly the trades already stretched by the simultaneous national buildout of transmission, semiconductor fabrication plants, housing, and energy projects. Electricians, engineers, welders, linemen, pipefitters, equipment operators, HVAC specialists, and project managers are demanded simultaneously by datacenters and by every other component of the industrial expansion, which means the workforce claim of a campus is a claim against a genuinely scarce resource whose price and availability affect every other project in the region. Datacenter Measures should therefore always include both construction workforce intensity and permanent operating employment, and the two figures should remain permanently separate in any disclosure, because they describe very different economic effects: the first is a large, temporary, regional labor-market draw, while the second is a small, durable, local employment base. Conflating them, as promotional announcements routinely do, weakens every downstream analysis built upon them.
1.8 From Megawatt to Resource Statement
The analytical destination of this section is the paper’s foundational tool: the Datacenter Resource Statement. For every major project, the statement would identify, at minimum, the project’s electricity profile in contracted, expected, and peak megawatts together with load factor; its generation arrangements, distinguishing existing from incremental and firm from intermittent supply; its grid requirements, including substations, transmission additions, transformer requirements, and interconnection costs; its water profile, covering withdrawal, consumption, source, cooling method, and local water stress; its land footprint, including associated infrastructure beyond the campus boundary; its capital structure, separating private investment from utility investment attributable to the project; its public support, including tax abatements, grants, and infrastructure assistance; its employment, separated into construction and permanent categories; and its community-infrastructure requirements, from roads to fire protection to emergency services. The Datacenter Resource Statement would not determine whether a project should be approved, and it is essential to insist on that limitation, because measurement and judgment are different functions. What the statement would create is a common factual basis from which corporations, regulators, lenders, and communities could evaluate the project on comparable terms, and Section 4 of this paper develops the statement’s full architecture.
Section 2: California Disclosure Versus Texas Verification
2.1 September 21, 2026: Two States, Two Measurement Philosophies
California and Texas provide an unusually clean natural experiment in datacenter governance precisely because both acted decisively on the same day yet chose fundamentally different policy instruments, and because the two states together represent such a large share of American economic output, electricity consumption, and datacenter development that their choices will shape the strategies of every national developer regardless of what any other jurisdiction does. California moved toward information disclosure and structured cost responsibility, embedding the datacenter within its existing environmental-review, water-planning, and utility-regulation institutions. Texas moved toward audit, verification, and temporary permitting restraint, effectively declaring that the state would not extend further regulatory approvals until it could distinguish real demand from speculative demand. The underlying concern is nonetheless similar in both capitals: large projected AI loads increasingly have consequences that extend far beyond the boundaries of individual corporate campuses, and the institutions that approve those campuses can no longer act on the information the industry has traditionally volunteered [1, 5, 7].
2.2 California: Measure Before Costs Become Socialized
California’s seven-bill package approaches datacenters through multiple systems rather than through a single permitting rule, and the structure of the package is itself instructive, because it demonstrates that the state’s legislature has concluded that no single agency and no single statute can capture the full resource footprint of an AI campus. Electricity matters, and so SB 886, SB 1168, and AB 2383 direct the Public Utilities Commission toward rate structures under which datacenters bear the cost of connecting and serving their own load. Water matters, and so AB 2469 and AB 2619 make water disclosure a precondition of local approval and an ongoing reporting obligation. Land-use and environmental review matter, and so SB 887 closes the exemption pathways while rewarding conservation-compliant projects with expedited treatment. Reporting matters in its own right, and so AB 1577 establishes the informational backbone on which the other measures depend. Community visibility matters, and the entire package is framed around ensuring, in the words of the governor’s office, that communities have more control over water, electricity, and land use [1, 2]. This provides an important conceptual precedent for Datacenter Measures: California’s policymakers are treating AI infrastructure as a multi-resource project rather than as merely an electricity customer, which is exactly the analytical shift this paper argues the entire field must make.
2.3 The California Data Model
California’s framework can be interpreted analytically as five interlocking categories of measurement, each answering a distinct question that a community, a regulator, or a ratepayer advocate would want answered before approval. Measure A concerns electricity: how much power does the facility require, and what new generation or procurement obligations follow from it? Measure B concerns the grid: what upgrades are required, and, critically, who bears their cost? Measure C concerns water: how much will be required, from which source, and under what drought assumptions? Measure D concerns land and environment: what consequences accompany the project, and what review will they receive? Measure E concerns community: what workforce and local information should be available before development proceeds? These five categories will become useful again in Section 5, when the paper constructs a possible national reporting template, because they demonstrate that a comprehensive disclosure regime can be assembled from measures that each remain administratively tractable on their own.
2.4 Texas: Measure Whether the Demand Is Real
Texas presents a different problem, and in some respects a harder one, because the question confronting Texas regulators is not primarily how to allocate the costs of real demand but how to identify which demand is real in the first place. The state accumulated an enormous pipeline of requests from datacenters and other large loads seeking grid connections: from 63 gigawatts at the end of 2024, the ERCOT large-load queue nearly quadrupled to roughly 226 to 233 gigawatts by late 2025, with more than seventy percent attributable to datacenters, and then continued climbing to approximately 474 gigawatts by August 2026, more than five times the highest peak demand the ERCOT system has ever served [7, 8, 36]. No serious analyst believes 474 gigawatts will be built. The figure reflects duplicated applications filed in multiple jurisdictions, speculative site banking, exploratory positioning by developers who may never secure chips or capital, and genuine projects whose scale remains uncertain. But that is precisely why verification has become a first-order policy function. For grid planners, speculative projects create their own costs even when they never break ground: transmission cannot efficiently be planned around fictional demand, generation cannot rationally be financed against duplicated applications, and long-lead equipment such as transformers cannot be allocated accurately if developers reserve capacity in several states simultaneously. The industry has even acquired a vocabulary for the problem, with utilities and grid operators now openly describing the flood of speculative or “phantom” load requests that may never materialize [36].
Texas therefore introduces a second fundamental Datacenter Measure, distinct from anything in the California package: demand credibility. The question is not merely how many megawatts are requested, but how many megawatts are sufficiently financed, permitted, contracted, and technologically committed to have a realistic probability of becoming operational, and the August and September 2026 directives make the production of exactly that information a condition of further state approvals, with ERCOT’s comprehensive audit expected to conclude in December 2026 and the TCEQ ordered to report its compliance by October 19, 2026 [5, 6].
2.5 The Difference Between Announced MW and Bankable MW
The Texas experience motivates a taxonomy that may prove to be one of the most durable contributions of this entire policy period, because it transforms the megawatt from a publicity number into a maturity ladder along which every project can be located and tracked. The ladder runs from announcement to utilization, and each rung represents a genuine increase in the probability that the demand will materialize and therefore in the confidence with which infrastructure may be planned around it.
| Maturity Stage | Definition | Planning Confidence |
| Announced MW | Capacity publicly discussed in press releases and corporate announcements | Minimal; promotional in nature |
| Requested MW | Capacity formally sought through interconnection or utility processes | Low; queues contain heavy duplication |
| Reserved MW | Capacity provisionally allocated by a utility or grid operator | Low to moderate; still revocable |
| Contracted MW | Capacity supported by executed electric-service agreements with minimum-take obligations | Moderate to high; financial commitment exists |
| Financed MW | Projects with committed equity and debt behind them | High; capital is at risk |
| Construction MW | Capacity attached to facilities physically being built | High; site control and equipment committed |
| Energized MW | Capacity actually connected to the grid | Realized as capability |
| Utilized MW | Electricity actually consumed in operation | Realized as demand |
The distance between the top and bottom of this ladder is the space in which forecasting errors, stranded assets, and misallocated public resources are born. A state that plans against announced megawatts will overbuild; a state that plans only against energized megawatts will underbuild and strand its economy; the entire art of the coming decade lies in planning against the correct intermediate rungs, which is impossible unless projects are measured and reported by rung.
2.6 California Measures Externalities; Texas Measures Credibility
The analytical distinction between the two states can therefore be stated with unusual compactness. California asks: what will this datacenter require from the surrounding system, and how do we prevent those requirements from being silently transferred to households? Texas asks: is this datacenter’s demand sufficiently credible for the surrounding system to plan around it at all, and how do we prevent speculative demand from distorting billions of dollars of infrastructure investment? Both are Datacenter Measures in the full sense this paper intends. They simply measure different risks, and a mature national framework will need both, because externality measurement without credibility screening invites planning around phantoms, while credibility screening without externality measurement invites cost-shifting around real projects.
2.7 The Emerging State Laboratory
The United States may therefore develop datacenter policy in the classic pattern of American federalism, from the bottom up, with states functioning as the laboratories Justice Brandeis famously described. California is testing disclosure and cost allocation. Texas is testing verification and audit. Virginia is testing executive-driven transparency and accountability structures in the world’s most mature datacenter market. Ohio is testing tariff-based large-load contracting through the precedent set by its Public Utilities Commission. Pennsylvania is testing binding community and permitting requirements layered atop energy abundance. Indiana is testing the interaction between industrial incentives, new generation, and large-load expansion within AEP’s and other utilities’ rapidly growing service territories. Arizona is illuminating the collision between electricity-intensive economic development and water-constrained geography in the shadow of Colorado River scarcity. Rather than assuming these states will converge on identical rules, the remainder of this paper examines them as policy laboratories producing different measurements of the same industrial transformation, and Section 4 assembles their approaches into a comparative framework.
Section 3: Measuring the Costs Beyond the Server Hall
3.1 The Electricity Bill Is Only the First Bill
A datacenter pays for electricity, and the industry is understandably quick to note that hyperscale operators are among the largest and most reliable utility customers in the country. But paying an electricity bill does not automatically mean that the bill captures every system cost associated with serving the load, and the gap between what is billed and what is caused has become one of the most consequential accounting questions in American utility regulation. If accommodating a campus requires a new transmission line, a new substation, a new gas plant, or a major distribution upgrade, the decisive questions become how those long-lived assets are financed, over what period, under what tariff, and with what allocation of risk if future load differs from forecasts. Datacenter Measures therefore insists on separating commodity electricity cost, which is the price of energy actually consumed, from infrastructure cost, which is the capital burden of the physical system built to make that consumption possible, because the two follow entirely different economic logics and can be shifted onto entirely different parties.
The most rigorous scholarly treatment of this gap comes from Harvard Law School, where Eliza Martin and Ari Peskoe of the Electricity Law Initiative published their March 2025 study, Extracting Profits from the Public: How Utility Ratepayers Are Paying for Big Tech’s Power. Martin and Peskoe documented at least three avenues through which datacenter energy costs migrate into the bills of ordinary customers: special contracts negotiated between utilities and datacenter owners and reviewed through opaque, often confidential regulatory processes; rate structures that socialize infrastructure built predominantly for large loads; and colocation arrangements between datacenters and existing power plants that can raise wholesale prices and distort delivery rates for everyone else [28]. Their conclusion about prevailing industry practice was blunt:
“luring data centers with discounted contracts or lopsided tariffs is unsustainable”
— Eliza Martin and Ari Peskoe, Harvard Law School Electricity Law Initiative, Extracting Profits from the Public (2025) [28]
Peskoe’s own summary of the stakes, delivered as the report circulated through statehouses and utility commissions during 2025 and 2026, has become one of the most quoted sentences of the entire debate:
“We’re all paying for the energy costs of the world’s wealthiest corporations.”
— Ari Peskoe, Director, Electricity Law Initiative, Harvard Law School [29]
3.2 Transmission as Embedded AI Capex
When a hyperscaler announces a ten-billion-dollar campus, the public headline focuses on corporate investment, and the figure is genuinely enormous. Yet utilities may simultaneously invest billions of dollars in transmission and generation across the surrounding region, and those expenditures, financed through regulated rate base and recovered over decades, are as much a part of the project’s true capital cost as the servers themselves. This observation motivates a broader, two-part definition of AI capital expenditure. Direct AI Capex comprises GPUs, servers, buildings, cooling systems, and private onsite infrastructure, the spending that appears in corporate earnings reports. Enabling AI Capex comprises transmission, substations, generation, water systems, and infrastructure required outside the campus, the spending that appears in utility capital plans and municipal budgets. The second category is frequently invisible in public discussion even though it determines whether the first can operate at all.
Both categories are now growing at rates that strain historical comparison. On the direct side, the second-quarter 2026 earnings season revealed that Microsoft, Alphabet, Amazon, and Meta collectively expect capital expenditures approaching 760 billion dollars for calendar 2026, up from roughly 413 billion dollars in 2025, with Amazon guiding toward approximately 220 billion dollars, Alphabet toward 195 to 205 billion dollars, Meta toward 130 to 145 billion dollars, and Microsoft toward roughly 175 billion dollars on a calendar basis [34]. Goldman Sachs, revising its models after the first-quarter reports, projected a combined 5.3 trillion dollars of capital expenditure for the four largest hyperscalers between fiscal 2025 and fiscal 2030, with a baseline aggregate estimate of 7.6 trillion dollars between 2026 and 2031 across compute, datacenters, and power [35]. On the enabling side, AEP’s 78-billion-dollar five-year plan, of which transmission alone constitutes 33 billion dollars and of which nearly ninety percent of incremental contracted load is datacenter-driven, is only the most visible example of a pattern repeating across essentially every large American utility [25, 26]. The essential point for Datacenter Measures is that these two capital streams belong in the same analytical frame, because they are two halves of a single industrial system, and measuring one without the other misstates both the scale and the risk distribution of the AI buildout.
3.3 Generation Commitments and the Attribution Problem
A large campus can reshape long-term generation planning across an entire region, and utilities are responding through new gas generation, renewables, storage, nuclear uprates, and life extensions of plants once scheduled for retirement. For cost-allocation purposes, however, the raw quantity of new generation is less important than its attribution, and the measurement discipline must therefore distinguish capacity built specifically because of large-load demand from capacity that would have been constructed anyway in the ordinary course of load growth and fleet turnover. That distinction sounds technical, but it is where billions of dollars change hands: if a gas plant is attributed to general system needs, its cost is socialized across all customers; if it is attributed to a datacenter cluster, cost-causation principles argue that the cluster should bear it. The counterfactual is genuinely difficult to establish, which is exactly why it must be established through transparent, standardized analysis rather than through confidential settlement, and why the attribution methodology itself belongs among the Datacenter Measures a mature framework will need to specify.
3.4 Water Infrastructure Costs
The same two-part logic that separates commodity electricity from electricity infrastructure applies with equal force to water. A facility may pay its normal volumetric water tariff faithfully and still require new mains, expanded treatment capacity, additional pumping equipment, new reservoirs or storage, recycled-water facilities, drought contingencies, or alternative cooling infrastructure whose capital costs fall on the water utility and, through it, on the ratepaying community. The University of California, Riverside study of March 2026 quantified this previously uncounted category, estimating national datacenter-driven water-infrastructure requirements at 10 to 58 billion dollars depending on growth rates, and noting that in February 2026 alone three major technology companies secured multi-million-gallon-per-day supplies for projects in Virginia, Louisiana, and Indiana with associated water-infrastructure costs approaching one billion dollars [30]. The relevant Datacenter Measure is therefore the incremental water-system cost attributable to incremental AI load, measured at peak rather than average conditions, because cooling demand spikes on exactly the hot days when municipal systems are already stressed, and an annual average conceals the very hours in which the constraint binds.
3.5 Tax Incentives as Infrastructure Capital
Tax incentives constitute another accounting category that has traditionally been separated from infrastructure analysis and that Datacenter Measures deliberately reunites with it. A state may receive substantial capital investment while simultaneously granting property-tax abatements, sales-tax exemptions on equipment, or direct infrastructure support, and because these grants flow through tax expenditure rather than appropriation, they frequently escape the scrutiny applied to ordinary spending. This paper does not presume that such incentives are economically harmful; in some circumstances they may be decisive in attracting investment whose net fiscal contribution is strongly positive. What Datacenter Measures asks is simply that their value be measurable alongside the benefits used to justify them, within a single disclosure template that a legislator, a journalist, or a bond analyst can read in one sitting. It is notable that the political momentum now runs toward exactly this reunification: Governor Abbott announced that he will work with the Texas Legislature in its next session to eliminate financial incentives for datacenters entirely [5], and Pennsylvania’s executive order directs state agencies to condition tax-exemption eligibility on compliance with its GRID requirements [22].
| Public Ledger Line | Entries |
| Additions (benefits) | Corporate capital investment; expected tax revenue; construction employment; permanent employment; local procurement; utility investment supporting broader reliability |
| Subtractions (costs) | Property-tax abatements; sales-tax exemptions; infrastructure grants; identifiable public infrastructure commitments; incremental water and grid costs not borne by the developer |
| Result | Not a political score, but an economic disclosure enabling informed judgment |
3.6 Emergency Services and Municipal Capacity
Hyperscale campuses are industrial facilities, and their public-service footprint deserves the same systematic treatment as their electricity and water. That footprint can involve fire protection sized for facilities containing megawatt-hours of stored chemical energy in batteries and large volumes of diesel fuel; hazardous-materials response; policing; road access and pavement wear from years of heavy construction traffic; traffic management; emergency planning; water pressure adequate for fire suppression; communications; and disaster-response coordination. These costs may prove modest relative to a multibillion-dollar project, and in many communities the fiscal contribution of a datacenter will comfortably exceed them. But Datacenter Measures is built on completeness rather than on the assumption that any particular category will dominate, because the categories that dominate differ by place, and the entire purpose of a standardized statement is to let each community see which categories dominate in its own case.
3.7 Backup Generation and Local Air Impacts
Reliability requirements give rise to one of the least appreciated features of the modern AI campus: it contains not one energy system but two. The first is the grid-connected system through which the facility normally operates. The second is the emergency system it maintains in case the first fails, typically an extensive fleet of diesel or gas generators, together with batteries and fuel storage, capable of carrying hundreds of megawatts of load for extended periods. Both systems must be measured, because the second has its own land footprint, its own capital cost, its own fuel logistics, and, most importantly for host communities, its own air-permit profile during testing and emergency operation. Texas’s September directive explicitly reaches this second system, extending the permitting pause to behind-the-meter generation precisely so that developers cannot bypass the audit by building their own power [6]. The emerging AI factory is not merely a consumer of grid power; it is an energy complex containing transformers, batteries, generators, cooling systems, and sophisticated power-management equipment, and it should be disclosed as such.
3.8 Public Cost Versus Public Benefit
A rigorous Datacenter Measures framework must measure both sides of the ledger, and it is worth stating explicitly what this paper is not arguing. It is not arguing that datacenters are net fiscal burdens; in many jurisdictions they demonstrably are not, and the tax revenues, construction employment, permanent jobs, local procurement, utility investment, infrastructure improvements, and broader economic activity they generate can be substantial. Nor is it arguing that public costs are invariably hidden; Ohio’s tariff proceeding, discussed in Section 4, shows a regulatory process capable of surfacing and allocating them openly. What the paper argues is that the tradeoffs must be visible before resources are committed, because a community that discovers the costs after the substation is built has no remaining leverage, and a developer whose genuine benefits are drowned in unmeasured suspicion has no way to demonstrate them. The purpose of measurement is not to predetermine the conclusion. It is to make the conclusion possible.
Section 4: The Datacenter Resource Statement — Creating Comparable Infrastructure Accounting
4.1 Why Comparability Matters
A fundamental and largely unremarked problem in today’s AI infrastructure market is that projects are described inconsistently, and the inconsistency is not innocent, because each descriptive convention flatters a different constituency. One developer announces investment dollars, which impress governors. Another announces megawatts, which impress utilities and rivals. Another announces acres, which impress land markets. Another announces jobs, which impress county commissions. Another announces GPUs, which impress capital markets. Without comparable metrics, state officials, utilities, investors, and communities can spend months debating what are, in effect, fundamentally different objects, and no party can be confident that the ten-billion-dollar announcement in one state describes more or less real activity than the two-gigawatt announcement in another. Financial markets solved the analogous problem a century ago through standardized financial statements audited to common rules; no investor today would accept a corporation that reported only whichever metric made it look best. Datacenter Measures proposes the equivalent instrument for AI infrastructure: a standardized resource statement, organized into eight sub-statements, filed and updated at defined milestones across a project’s life.
4.2 Statement One: Electricity
The electricity statement is the foundation of the entire disclosure because every other resource claim scales with load. It should distinguish requested load from contracted load, since the gap between the two is precisely the speculative margin Texas is now auditing; expected average load from peak load, since infrastructure is sized to the peak while economics are driven by the average; and it should disclose the ramp schedule, utilization assumptions, flexibility capability, and expected annual consumption. Flexibility deserves particular emphasis, because a facility able to curtail or shift load during system peaks imposes materially lower capacity costs than an identical facility that cannot, and research at MIT’s Center for Energy and Environmental Policy Research has begun quantifying both the cost savings and the emissions tradeoffs of flexible datacenter operation. A disclosure regime that rewards demonstrated flexibility would convert what is today a private operational detail into a publicly valuable planning parameter, and this alone would help distinguish genuine infrastructure demand from speculative queue positions.
4.3 Statement Two: Generation
The generation statement identifies what stands behind the meter and behind the market: existing grid supply relied upon; incremental generation induced; the technology, expected commercial-operation date, and firmness of that generation; contracted clean-energy resources; onsite generation; backup generation; and storage. Its purpose is to reveal the true relationship between compute growth and power-system growth for the specific project at hand, because that relationship varies enormously. A campus contracted against an existing nuclear plant’s uprated output has a fundamentally different generation footprint from one whose demand triggers construction of new combined-cycle gas capacity, even if their consumption is identical, and present disclosure practice makes the two indistinguishable to everyone except the counterparties.
4.4 Statement Three: Grid Infrastructure
The grid statement is where Datacenter Measures intersects most directly with utility finance, and it should identify new substations; transformer requirements, given that transformer lead times have become one of the binding constraints of the entire buildout; transmission additions; distribution upgrades; estimated upgrade cost; construction schedule; the cost-responsibility mechanism under which those costs will be recovered; and, critically, stranded-cost protection, meaning the contractual and tariff provisions that determine who pays if the load never materializes or departs early. Ohio’s experience demonstrates that this last item can be specified with precision: the tariff approved by the Public Utilities Commission of Ohio on July 9, 2025 requires new datacenter customers above 25 megawatts to pay for a minimum of eighty-five percent of their subscribed capacity regardless of actual consumption, over terms of up to twelve years including a four-year ramp, backed by collateral requirements and exit fees for cancelled projects [27]. Whatever one thinks of the particular percentages, the structure shows that stranded-asset risk can be measured, priced, and allocated in advance rather than litigated after the fact.
4.5 Statement Four: Water
The water statement should include annual withdrawal; annual consumption; peak-day requirements; source; percentage recycled; cooling technology; drought-year requirements; required infrastructure upgrades; and watershed stress. Each element earns its place. Withdrawal and consumption diverge sharply for evaporative systems. Peak-day figures matter because, as the Riverside research emphasizes, annual averages conceal the short periods when cooling systems demand the most water and when municipal systems are least able to supply it [30]. Source matters because groundwater, surface water, and recycled water carry different scarcity and replenishment profiles. And watershed stress matters because it is the conversion factor between physical gallons and economic significance, permitting meaningful comparison between projects in California, Arizona, Texas, Virginia, and water-abundant regions where the same nominal gallon has entirely different opportunity costs. The European Commission’s proposal to require disclosure of water efficiency in relation to local water stress adopts precisely this logic at continental scale [17, 18].
4.6 Statement Five: Land and Environment
The land statement should measure datacenter acreage; energy-infrastructure acreage, including generator yards and battery installations; transmission corridors; water infrastructure; the noise envelope, which has become a leading source of community complaint around large campuses; the environmental review pathway applicable to the project; and adjacent land-use effects. Its unifying principle is that a server campus’s true footprint extends beyond the property line whenever supporting infrastructure requires additional land, and a disclosure regime that stops at the fence invites developers to push their most contested infrastructure just beyond it.
4.7 Statement Six: Public Finance
The public-finance statement should disclose state incentives; local incentives; sales-tax exemptions; property-tax abatements; infrastructure grants; public financing; expected tax revenue; and incentive duration. As Section 3.5 argued, the purpose is comparability rather than predetermined judgment, but comparability here has a distinctive democratic function: tax expenditures are public money by another name, and a community asked to weigh a project deserves to see the full fiscal exchange on one page, in the same units, over the same time horizon, before the exchange becomes irreversible.
4.8 Statement Seven: Workforce
The workforce statement enforces the separation this paper has already insisted upon: temporary employment, meaning construction and installation jobs measured in worker-years across the build period, must always be reported separately from durable employment, meaning expected operating positions once the campus is complete. Conflating the two, as promotional materials habitually do, weakens infrastructure analysis at its root, because it obscures both the genuine short-run labor-market draw of construction, which competes with every other regional project for scarce skilled trades, and the genuine long-run modesty of operating headcounts, which communities must understand when they weigh permanent fiscal benefits against permanent resource commitments. Virginia’s framework and Pennsylvania’s GRID requirements both move in this direction by attaching explicit local-workforce and local-procurement expectations to approval [19, 22].
4.9 Statement Eight: Community Infrastructure
The final statement gathers roads, fire services, emergency-response capabilities, housing pressure, training programs, and other identifiable local requirements. Not every project will materially affect every measure, and many entries will legitimately read as negligible. That is precisely why standardized disclosure is useful: a negligible entry, affirmatively disclosed, is information; a missing category is merely an unanswered question, and communities have learned to treat unanswered questions as concealment whether or not concealment was intended.
4.10 Seven States, Seven Emphases: California, Texas, Virginia, Ohio, Indiana, Pennsylvania, and Arizona
The resource-statement architecture becomes most instructive when laid against the actual behavior of the leading datacenter states, because each state has, in effect, begun building two or three statements of the eight while leaving the others unwritten. The comparison that follows is emphatically not a ranking. Its purpose is to demonstrate that the same AI megawatt means something different depending on where it is built, and that the national picture is currently the union of partial ledgers.
| State | Primary Analytical Emphasis | Signature Instrument (2025–2026) |
| California | Disclosure, water, electricity rates, environmental review, and cost allocation | Seven-bill package signed September 21, 2026: AB 1577, AB 2383, AB 2469, AB 2619, SB 886, SB 887, SB 1168 [1] |
| Texas | Load credibility, grid impact, water, and project verification | ERCOT/TWDB audits and statewide permitting pause; TCEQ directive of September 21, 2026 [5, 6] |
| Virginia | Transparency and cumulative-impact management in the world’s largest datacenter cluster | Data Center Accountability Framework and Executive Order 22, September 18, 2026; first state to limit behind-the-meter gas generation and to enact a statewide datacenter energy-consumption tax [19, 20] |
| Ohio | Utility cost allocation and large-load contracting | PUCO-approved AEP Ohio Data Center Tariff: 85% minimum-demand obligation, collateral, and exit fees, July 9, 2025 [27] |
| Indiana | Economic-development incentives, industrial land, utility expansion, and new gas-fired generation | AEP-territory generation additions within the $78B capital plan serving contracted large loads [25] |
| Pennsylvania | Energy abundance combined with environmental safeguards, community consent, and disclosure | Executive Order 2026-05: removal from Fast Track permitting, NDA prohibition, binding GRID requirements, August 18, 2026 [22, 23] |
| Arizona | Semiconductor manufacturing, datacenter expansion, and growth economics in a water-constrained region | Water-stress-driven siting scrutiny amid Colorado River scarcity; the reference case for peak-water disclosure [30] |
The Virginia entry deserves elaboration, because Virginia is the jurisdiction in which the datacenter industry’s past and future meet most directly. Governor Spanberger’s framework, organized around five pillars of transparency, environmental protection, energy costs, clean energy, and workforce, ends the state’s previous accommodationist posture, prohibits the nondisclosure agreements that had shrouded siting negotiations, removes datacenters from fast-track permitting, and directs legislation for the 2027 General Assembly session [19, 20]. Her framing at the announcement captured the shift in a single sentence:
“this industry won’t have carte blanche to play by their own rules”
— Governor Abigail Spanberger of Virginia, announcing the Data Center Accountability Framework, September 18, 2026 [21]
Pennsylvania’s trajectory is equally revealing, because Governor Shapiro had stood beside Amazon barely a year earlier to announce a twenty-billion-dollar datacenter investment in the Commonwealth, and by August 2026 he was describing speculative developers in the language of predation and signing an order that conditions every state environmental permit on binding commitments and prior local approval [22, 24]. His explanation reached back to the state’s industrial memory:
“industry running roughshod over our communities to make a buck”
— Governor Josh Shapiro of Pennsylvania, signing Executive Order 2026-05, August 18, 2026 [24]
What these seven partial ledgers demonstrate, taken together, is that every element of the eight-statement architecture already exists somewhere in American practice. California has written the water and rate statements. Texas is writing the credibility screen. Ohio has written the stranded-cost statement. Virginia and Pennsylvania have written the community and transparency statements. Arizona’s hydrology is forcing the peak-water statement into being. The national task is not invention. It is assembly.
Section 5: From State Experiments to a 2027–2030 National Datacenter Measurement Architecture
5.1 Why State-by-State Accounting May Become Difficult
A hyperscaler operating nationally may soon confront dozens of different disclosure standards, and the compliance geometry is worth pausing over, because it explains why the industry itself may eventually become a constituency for standardization. One state measures water withdrawal while another measures consumption. One state emphasizes grid cost allocation through tariffs while another emphasizes it through legislation. One state audits load credibility while another accepts developer forecasts. One state regulates backup generators through air permits while another reaches them through interconnection policy. One state focuses on incentives while another prohibits nondisclosure agreements. Each requirement is individually rational, and this paper has defended most of them. But as they multiply, both governments and corporations may come to see that the deepest value of measurement, which is comparability, is destroyed when every jurisdiction measures in its own units, on its own schedule, under its own definitions, and that common definitions can be adopted without surrendering an inch of state control over permitting and utility regulation.
5.2 The Case for Common Definitions
A national architecture would not require a single federal permitting regime, and this paper does not advocate one. It could begin far more modestly, with standard terminology of the kind that accounting standards boards, not legislatures, typically produce. What constitutes an “AI datacenter” for reporting purposes, and at what megawatt threshold do obligations attach, given that Ohio drew its line at 25 megawatts, Pennsylvania at 25 megawatts of peak demand, and the House bill at 100 megawatts [22, 27, 13]? What constitutes “requested load,” and when does requested load become “committed load”? How should average demand be reported relative to peak demand? How should water consumption be distinguished from water withdrawal? What qualifies as project-attributable grid infrastructure, and under what counterfactual methodology? How should tax incentives be valued and over what horizon? None of these questions requires federal preemption to answer. Once definitions are standardized, comparisons become possible without requiring identical state policies, exactly as standardized accounting permits comparison of corporations chartered under fifty different state laws.
5.3 A National Datacenter Resource Statement Across the Project Lifecycle
By 2027 to 2030, one plausible architecture would resemble a standardized infrastructure disclosure filed at major development milestones, so that the measurement follows the datacenter across its lifecycle rather than relying on a single forecast made years before operation and never reconciled against reality. The stages map naturally onto the maturity ladder of Section 2.5, and each filing would update the eight statements of Section 4 with the information genuinely knowable at that stage.
| Lifecycle Stage | Disclosure Filed | Principal Content |
| Stage 1 — Proposed | Initial Resource Statement | Land position, targeted capacity, preliminary resource requirements |
| Stage 2 — Interconnection | Load Statement | Requested and contracted electricity capacity; flexibility commitments |
| Stage 3 — Permitted | Resource Obligations Statement | Water, environmental, land-use, and infrastructure obligations as permitted |
| Stage 4 — Financed | Capitalization Statement | Evidence of project capitalization and binding equipment and power commitments |
| Stage 5 — Construction | Progress Statement | Actual construction progress and revised resource requirements |
| Stage 6 — Energized | Capacity Statement | Available electrical capacity as connected |
| Stage 7 — Operating | Utilization Statement | Actual electricity, water, and infrastructure utilization, reconciled against forecasts |
5.4 Measuring Ghost Demand Without Freezing Real Demand
A standardized maturity framework could address the hardest operational problem now facing utilities and grid operators: separating real projects from duplicated or speculative interconnection requests without imposing the blunt instrument of a moratorium on everyone. Texas demonstrates the scale the problem can reach, with requested large-load capacity vastly exceeding anything the electrical system could realistically accommodate, and Texas’s response, a statewide permitting pause, also demonstrates the collateral cost of lacking a graduated screen, because a pause delays credible projects and speculative ones alike [5, 7]. The alternative is progressive demonstration: rather than treating every queue position identically, projects would advance through defined gates of site control, capital commitment, equipment commitment, power contract, construction, and energization, with planning confidence and processing priority rising at each gate. The further a project progresses, the greater the confidence grid planners could place in its forecast demand, and the less reason any state would have to freeze its entire pipeline in order to audit it. ERCOT’s post-2026 practice of including in its load forecast only large loads that have executed qualifying interconnection agreements is an early, partial version of exactly this discipline [8].
5.5 Measurement Could Become a Financing Tool
Datacenter Measures should not be understood solely as regulatory compliance, because verified information is itself an economic asset, and the parties who would consume these disclosures extend far beyond regulators. Better data could reduce uncertainty for utilities planning rate base; for bond investors pricing utility debt against load forecasts; for banks and infrastructure funds underwriting transmission and generation; for insurers pricing construction and operational risk; for power producers sizing plants against credible offtake; for equipment manufacturers scheduling transformer and turbine production against demand that will actually arrive; and for datacenter developers themselves, whose credible projects currently pay a congestion penalty imposed by their speculative competitors’ queue positions. A lender considering financing a transmission project would have better information about the loads supporting repayment. A utility ordering transformers could distinguish committed projects from conceptual ones. A power developer could better estimate generation requirements. Measurement could therefore lower some forms of infrastructure risk even while imposing additional reporting requirements, and the historical analogy is again to financial disclosure, which the securities industry resisted in 1933 and could not function without today.
5.6 Federalism Rather Than Uniformity
The paper deliberately avoids assuming that Washington must replace state authority, because nothing in the record of 2025 and 2026 suggests either that Congress wishes to nationalize datacenter permitting or that it would do it well. A more plausible 2027–2030 architecture layers responsibilities: federal definitions and reporting standards; state permitting; state utility regulation; regional grid planning through the RTOs and ISOs; local land-use decisions; and corporate reporting to capital markets. Such a system would preserve the different institutional competencies where they actually reside while making the underlying resource information comparable across all of them, and it has the additional virtue of matching how American infrastructure federalism already works in banking, securities, and environmental reporting, where federal definitions coexist with state enforcement and local application.
5.7 Congress Begins Entering the Datacenter Cost Debate
The U.S. House’s passage of the Ratepayer Protection Act on September 16, 2026 demonstrates that datacenter infrastructure costs are no longer solely a state regulatory issue, and the anatomy of the vote repays study. The bill, H.R. 9340, introduced by Representative Gabe Evans of Colorado and shepherded by Energy and Commerce Chairman Brett Guthrie, passed 417 to 3 under suspension of the rules, with every Republican present voting in favor and only three progressive Democrats opposed, on the ground that the bill did not go far enough [13, 14]. Mechanically, the legislation works through Section 111(d) of the Public Utility Regulatory Policies Act of 1978, requiring state utility commissions to consider, though not necessarily adopt, a federal standard under which large computational loads above 100 megawatts pay the full incremental costs of serving them; it also codifies portions of the White House’s Ratepayer Protection Pledge of March 4, 2026, under which Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI, together with more than three hundred utilities, cooperatives, and state governments, committed to financing their own generation and covering their own grid-upgrade costs [16]. Chairman Guthrie framed the legislative theory in terms that could serve as a summary of this entire paper’s argument:
“keep building the needed infrastructure with appropriate accountability measures”
— Representative Brett Guthrie (R-KY), Chairman, House Committee on Energy and Commerce, on House passage of H.R. 9340 [15]
The bill’s consider-only mechanism means that actual ratepayer protection still depends on fifty state commissions, and its Senate prospects remain uncertain in a compressed election-year calendar [16]. But the 417-to-3 margin is significant for Datacenter Measures beyond its legal effect, because it reveals that measurement, transparency, and cost attribution have become one of the very few AI-related subjects on which an otherwise polarized Congress can act nearly unanimously. Federal policy may therefore initially evolve not through comprehensive AI-infrastructure regulation but through narrower questions of who counts what, and who pays for what is counted.
5.8 From Environmental Reporting to Infrastructure Reporting
It is worth being precise about the kind of document this architecture would produce, because it resembles neither of its two obvious ancestors. Environmental disclosure traditionally measures emissions, pollution, or resource use against ecological baselines. Corporate financial disclosure measures assets, liabilities, and flows against accounting standards. The AI Infrastructure Resource Statement proposed here is a hybrid of the two: it asks how much electricity, how much generation, how much grid, how much water, how much land, how much public support, how much employment, and how much infrastructure risk, and it asks these questions in physical units and dollars simultaneously, at project level, across a lifecycle, for the benefit of an audience that includes regulators, investors, lenders, communities, and competitors at once. The European Union’s evolving datacenter rating scheme, which began with energy-performance reporting under the recast Energy Efficiency Directive and expanded in September 2026 toward water-stress context and waste-heat reuse, shows one continental version of this hybrid taking shape in real time [17, 18]. The American version, if it emerges, will more likely grow from the state instruments catalogued in Section 4, harmonized through the common definitions of Section 5.2.
5.9 2030: From Gigawatts Announced to Gigawatts Accounted For
The most important transition by the end of the decade may therefore be linguistic as well as economic. The first AI infrastructure boom has been dominated by announcements, each larger than the last: ten billion dollars, fifty billion dollars, one hundred megawatts, five hundred megawatts, one gigawatt, five gigawatts, and, in the aggregate capital plans of the four largest hyperscalers, sums that Goldman Sachs now models at 5.3 trillion dollars through 2030 [35]. The next phase may demand a second question of every such number: accounted for how? Behind which contracts, which generation, which transmission, which water, which land, which incentives, and which allocation of risk? That question, asked systematically and answered comparably, is the heart of Datacenter Measures, and the remainder of this paper distills what the events of 2025 and 2026 have already taught about how it will be answered.
Section 6: What Have We Learned? Seven Pillars
Pillar 1 — Every AI Megawatt Has a Resource Shadow
A megawatt entering a datacenter does not begin at the datacenter wall. Behind it lies generation, transmission, substations, transformers, cooling systems, water, land, equipment, workers, and capital, and the length of that shadow varies by geography in ways a single number cannot express. The first principle of Datacenter Measures is therefore that compute capacity should be understood together with the physical system required to sustain it, because the AI infrastructure debate becomes incomplete, and eventually misleading, when it counts servers but not the infrastructure enabling those servers. The AI Megawatt Resource Envelope of Section 1 is the practical instrument of this pillar, and the eight-part Resource Statement of Section 4 is its formalization.
Pillar 2 — Measurement Must Separate Demand From Intent
Not every announced project will become an operating datacenter, not every interconnection request represents committed demand, and not every reserved megawatt becomes an energized megawatt; the 474-gigawatt ERCOT queue standing against a system whose record peak is roughly one-fifth of that figure is the era’s defining illustration [7]. The second pillar therefore distinguishes the full progression from announced through requested, reserved, contracted, financed, construction, and energized to utilized megawatts, and insists that electrical systems be built around credible demand rather than headlines. This progression may become one of the most important planning tools of the AI infrastructure era, because it converts the central epistemic problem of the boom, namely that nobody knows which demand is real, into an administrable sequence of verifiable gates.
Pillar 3 — Different States Are Measuring Different Risks
California, Texas, Virginia, Ohio, Pennsylvania, Indiana, and Arizona need not adopt identical policies to participate in the same larger transition, and the evidence of 2026 is that they will not. California is emphasizing resource disclosure and cost allocation; Texas is emphasizing verification and system impact; Virginia is emphasizing transparency and cumulative-impact management in a mature market; Ohio is emphasizing tariff-based stranded-cost protection; Pennsylvania is emphasizing binding community consent atop energy abundance; and the water-constrained Southwest is forcing peak-hydrology disclosure into being. The third pillar is therefore that America is developing Datacenter Measures through federalism, with different jurisdictions measuring different components of the same infrastructure problem, and the appropriate national response is harmonization of definitions rather than homogenization of policy.
Pillar 4 — Disclosure Can Become Economic Infrastructure
Information itself has economic value, and the fourth pillar holds that the more capital-intensive the AI economy becomes, the more valuable verified infrastructure information becomes. Utilities require credible forecasts; power producers require credible offtakers; lenders require credible cash flows; communities require credible resource estimates; governments require credible cost-benefit information; and equipment manufacturers require credible demand schedules against which to expand transformer and turbine production. Better measurement therefore need not exist solely to constrain development. It can also help distinguish projects that are financeable, buildable, and supportable from projects that remain speculative, and in doing so it functions as economic infrastructure in its own right, exactly as audited financial statements function as the infrastructure of capital markets.
Pillar 5 — Water Is the Hidden Binding Constraint
The electricity dimension of the AI buildout arrived first in public consciousness, but the scholarship and the legislation of 2025 and 2026 converge on a fifth pillar that the industry internalized late: in a growing number of communities, water, not power, is the binding constraint, and it binds at the peak hour of the driest year rather than at the annual average. The Riverside estimates of 4.2 to 6.6 billion cubic meters of AI-attributable withdrawal by 2027 and of 10 to 58 billion dollars in required water infrastructure quantify the physical claim [30, 31]; California’s AB 2469 and AB 2619, the European Commission’s water-stress disclosure, and the Texas Water Development Board’s audit all institutionalize it [1, 17, 5]. Money can build treatment plants and pipes, but, as Professor Ren has observed, it cannot buy more snowpack, and a measurement regime that reports withdrawal, consumption, peak-day demand, and watershed stress separately is the minimum required for communities to know whether a given campus fits within their hydrology.
Pillar 6 — Affordability Politics Has Made Measurement Bipartisan
The sixth pillar is political, and it would have seemed implausible as recently as 2024: datacenter cost accountability has become one of the only AI questions on which American politics is nearly unanimous. A Democratic governor of California and a Republican governor of Texas acted within hours of each other on the same day. A Democratic governor of Virginia and a Democratic governor of Pennsylvania adopted frameworks that Republican legislators in Congress echoed in a bill passing 417 to 3, with Republicans in competitive districts among its most eager co-sponsors as electricity bills joined groceries and housing among the affordability issues of the 2026 midterms [14]. The lesson for Datacenter Measures is that measurement is the natural common ground of this politics: parties that disagree profoundly about how much AI infrastructure to build can nonetheless agree on knowing what it costs and who pays, and frameworks built on disclosure and attribution are therefore likelier to survive electoral change than frameworks built on caps or mandates.
Pillar 7 — The Next AI Infrastructure Competition May Be About Accountable Capacity
The first stage of the AI race emphasized who could acquire the most GPUs. The second emphasized who could secure datacenter campuses and gigawatts. The seventh pillar anticipates the third stage: the competitive metric may evolve from how many megawatts a region or a company can announce to how many megawatts it can reliably power, finance, cool, connect, staff, and sustain, with the surrounding costs visible and allocated. That distinction defines accountable capacity, and accountable capacity may become considerably more valuable than speculative capacity as the Five-Layer AI Economy scales toward 2030, because every institution that the buildout depends upon, from utility commissions to bond markets to county boards, is learning to discount the unaccounted megawatt. The regions that master measurement first may, paradoxically, build fastest, because their projects will clear institutions that speculative projects can no longer clear.
Conclusion: Why “Datacenter Measures” Fits the Next Phase of the Five-Layer AI Economy
September 21, 2026 may eventually be remembered as a revealing moment in America’s AI infrastructure transition, less because of what either state did alone than because of what their simultaneity disclosed. On the same day, California enacted seven laws covering multiple dimensions of datacenter development while Texas extended its intervention into new datacenter permitting during an ongoing audit of electricity demand, water consumption, tax incentives, and community impacts, and, an ocean away, the European Commission proposed a continental disclosure and rating scheme for the same facilities [1, 5, 17]. The three governments did not adopt identical policies, their economic and electricity systems are profoundly different, and their political leaderships agree on very little else. Yet all three actions pointed toward the same emerging reality: AI infrastructure has become too large to describe with a single number.
A megawatt is important, but it is not enough. Behind that megawatt may stand generation plants, substations, transmission corridors, transformers, cooling equipment, water systems, backup generators, roads, construction workers, utility capital, and tax arrangements. Behind a gigawatt stand one thousand of those megawatts. And behind the hundreds of gigawatts now appearing in United States development pipelines, alongside hyperscaler capital plans approaching 760 billion dollars in 2026 alone and trillions across the decade, stands a physical transformation whose scale increasingly resembles a national industrial buildout rather than a conventional expansion of the software economy [34, 35]. That is why Datacenter Measures is the appropriate title for this paper, and why its two meanings must be held together.
The first meaning is government measures: laws, regulations, audits, utility tariffs, disclosure requirements, permitting procedures, environmental reviews, cost-allocation mechanisms, and infrastructure agreements. These are the institutional tools that governments from Sacramento to Austin to Richmond to Harrisburg to Brussels began deploying in earnest across 2025 and 2026, as AI datacenters became consequential participants in electricity, water, and land systems. The second meaning is quantitative measures: megawatts and megawatt-hours, gallons withdrawn and gallons consumed, acres, transmission miles, substation capacity, generation capacity, infrastructure dollars, tax incentives, construction workers, permanent jobs, and emergency-service requirements. These are the numbers that describe what the AI infrastructure boom physically requires. The title captures the essential convergence explored throughout this paper: as governments adopt more measures governing datacenters, datacenters themselves are being subjected to more measures of their resource footprint, and each process accelerates the other.
None of this means that every measurement should automatically become a regulatory limit, and it bears repeating that measurement and judgment are different functions. Datacenter Measures is fundamentally about improving the first so that corporations, utilities, investors, communities, and governments can make better-informed exercises of the second. That distinction matters because artificial intelligence is transforming the scale at which errors propagate. The IEA expects datacenters to account for almost half of United States electricity-demand growth through 2030, while Lawrence Berkeley National Laboratory’s 2026 modeling places their potential share of national electricity consumption around 11.8 percent by the end of the decade, within a wide band of genuine uncertainty [9, 12]. At those scales, measurement becomes economically consequential in itself. A forecast error of ten megawatts is manageable; a forecast error of ten gigawatts can misdirect billions of dollars of transmission, generation, and equipment investment. A speculative datacenter announcement affects primarily its developer; a speculative demand pipeline large enough to distort regional grid planning affects utilities, power producers, regulators, manufacturers, lenders, and customers far beyond the proposed campus. This is why the AI infrastructure conversation is evolving from capacity acquisition toward capacity accounting.
The first generation of the Five-Layer AI Economy could be summarized as a race upward: energy to chips to datacenters to models to applications and agents. Datacenter Measures introduces the complementary observation that every layer also produces obligations downward. More applications require more inference; more inference requires more model infrastructure; more model infrastructure requires more accelerators; more accelerators require more datacenters; and more datacenters require more electricity, land, cooling, transmission, capital, and physical infrastructure. The Five-Layer AI Economy therefore does not terminate at the application layer. At sufficient scale it loops back into the physical economy, as applications create compute demand, compute creates datacenter demand, datacenters create electricity demand, electricity demand creates infrastructure demand, and infrastructure demand creates public and private capital obligations. Datacenter Measures provides a way to observe that loop, layer by layer, project by project, before its costs harden into commitments.
The central question for 2027 through 2030 consequently should not be whether America, or any individual state, should build more or fewer AI datacenters, because that question, posed in the abstract, admits only ideological answers. The analytically more useful question is whether policymakers, utilities, corporations, investors, and communities can see clearly what each incremental unit of AI infrastructure requires before resources are committed around it. If the answer increasingly becomes yes, the AI infrastructure industry may move from today’s fragmented disclosures toward standardized resource statements; from speculative interconnection queues toward verified demand; from headline megawatts toward contracted megawatts; and from isolated project announcements toward infrastructure accounting that follows a facility from proposal through operation. The megawatt will remain indispensable, but it may no longer stand alone. It will arrive accompanied by measurements of the generation behind it, the transmission delivering it, the water cooling it, the land supporting it, the capital financing it, the incentives attached to it, and the community infrastructure surrounding it.
That is the transition captured by the title. That is why this paper is called Datacenter Measures. And that is why measures, in both senses of the word, the actions governments take and the quantities an industrial system must count, may become one of the defining concepts of the next stage of the Five-Layer AI Economy.
Footnotes and Endnotes:
[1] Office of Governor Gavin Newsom, “Governor Newsom signs most comprehensive data center laws in the nation, providing communities more control on water, electricity, and land use,” September 21, 2026. https://www.gov.ca.gov/2026/09/21/governor-newsom-signs-most-comprehensive-data-center-laws-in-the-nation-providing-communities-more-control-on-water-electricity-and-land-use/
[2] CalMatters (Khari Johnson), “What to know about 7 new data center laws Gavin Newsom signed,” September 21, 2026. https://calmatters.org/economy/technology/2026/09/new-california-laws-data-centers/
[3] KQED News, “Newsom Signs New Restrictions on Data Center Development,” quoting Senator Steve Padilla (D-San Diego), September 21, 2026. https://www.kqed.org/news/12100746/newsom-signs-new-restrictions-on-data-center-development
[4] SFist, “Newsom Signs Seven New Laws Regulating Data Centers Amid Statewide Backlash,” citing the July 2026 Public Policy Institute of California poll, September 21, 2026. https://sfist.com/2026/09/21/newsom-signs-seven-new-laws-regulating-data-centers-amid-statewide-backlash/
[5] Office of the Texas Governor (Governor Greg Abbott), “Governor Abbott Directs TCEQ To Halt Data Center Permits,” September 21, 2026. https://gov.texas.gov/news/post/governor-abbott-directs-tceq-to-halt-data-center-permits
[6] The Texas Tribune, “Gov. Greg Abbott broadens moratorium on data center approvals to include environmental permits,” September 21, 2026. https://www.texastribune.org/2026/09/21/texas-data-center-moratorium-water-energy/
[7] Robert Walton, Utility Dive, “Facing an estimated 474 GW of interconnection requests, Texas hits pause on data centers,” August 5, 2026. https://www.utilitydive.com/news/texas-hits-pause-data-center-interconnections/827046/
[8] Robert Walton, Utility Dive, “ERCOT’s large load queue jumped almost 300% last year,” quoting Kristi Hobbs, ERCOT Vice President of System Planning and Weatherization, January 6, 2026. https://www.utilitydive.com/news/ercots-large-load-queue-jumped-almost-300-last-year-official/808820/
[9] Arman Shehabi et al., Lawrence Berkeley National Laboratory, “United States Data Center Energy Usage Report: 2025 Update,” LBNL-2001758, June 2026. https://www.rtoinsider.com/wp-content/uploads/2026/06/data-center-energy-usage-2025-update.pdf
[10] Christa Marshall, E&E News by POLITICO, “US data centers’ electricity use could double by 2030, DOE lab says,” June 22, 2026. https://www.eenews.net/articles/us-data-centers-electricity-use-could-double-by-2030-doe-lab-says/
[11] U.S. Department of Energy, “DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers,” December 2024. https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers
[12] International Energy Agency, “Energy and AI” special report, and remarks of Executive Director Dr. Fatih Birol, as reported by S&P Global Commodity Insights, April 10, 2025. https://www.spglobal.com/commodity-insights/en/news-research/latest-news/electric-power/041025-global-data-center-power-demand-to-double-by-2030-on-ai-surge-iea
[13] National Association of Counties (NACo), “U.S. House passes the Ratepayer Protection Act” (H.R. 9340), September 17, 2026. https://www.naco.org/news/us-house-passes-ratepayer-protection-act
[14] NBC News, “House passes bill to shield consumers from data center price hikes on energy,” September 16, 2026. https://www.nbcnews.com/politics/congress/house-passes-bill-shield-consumers-data-center-price-hikes-energy-rcna597683
[15] U.S. House Committee on Energy and Commerce, “Ratepayer Protection Act Passes House with Strong Bipartisan Support,” statement of Chairman Brett Guthrie (R-KY), September 16, 2026. https://energycommerce.house.gov/posts/ratepayer-protection-act-passes-house-with-strong-bipartisan-support
[16] TechTimes, “House Voted 417-3 to Make Data Centers Pay Electricity Costs; States Can Still Say No,” on the PURPA Section 111(d) mechanism and the March 4, 2026 White House Ratepayer Protection Pledge, September 17, 2026. https://www.techtimes.com/articles/327645/20260917/house-voted-417-3-make-data-centers-pay-electricity-costs-states-can-still-say-no.htm
[17] Kate Abnett, Reuters, “EU to require data centres to disclose energy and water efficiency,” September 21, 2026. https://kfgo.com/2026/09/21/eu-to-require-data-centres-to-disclose-energy-and-water-efficiency/
[18] ESG News, “EU Proposes Energy and Water Labels for Data Centres,” September 21, 2026. https://esgnews.com/eu-proposes-energy-and-water-labels-for-data-centres/
[19] Office of Governor Abigail Spanberger, “Virginia’s New ‘Data Center Accountability Framework’,” September 18, 2026. https://www.governor.virginia.gov/newsroom/news-releases/2026/september-releases/name-1123696-en.html
[20] Commonwealth of Virginia, Executive Order 22 (2026), “Establishing New Virginia Data Center Accountability and Artificial Intelligence Initiatives,” September 18, 2026. https://www.governor.virginia.gov/media/governorvirginiagov/governor-of-virginia/pdf/eo/eo-22-establishing-va-data-center-accountability-and-ai-initiatives.pdf
[21] WTVR CBS 6 Richmond, “Gov. Spanberger proposes new data center rules for Virginia,” quoting Governor Abigail Spanberger, September 18, 2026. https://www.wtvr.com/news/local-news/spanberger-data-centers-ai-sept-18-2026
[22] DLA Piper, “Pennsylvania Executive Order establishes state compliance framework for data center development,” analysis of Executive Order 2026-05 signed by Governor Josh Shapiro, August 27, 2026. https://www.dlapiper.com/en-au/insights/publications/2026/08/pennsylvania-executive-order-establishes-state-compliance-framework
[23] Katie Meyer, Spotlight PA, “What Shapiro’s new data center order does — and doesn’t — force developers to do,” August 2026. https://www.spotlightpa.org/news/2026/08/pennsylvania-data-center-executive-order-shapiro-explainer-environment/
[24] AI Weekly, “Shapiro Signs Order Curbing ‘Predatory’ AI Data Centers in PA,” quoting Governor Josh Shapiro, August 18, 2026. https://aiweekly.co/alerts/shapiro-signs-order-curbing-predatory-ai-data-centers-in-pa
[25] American Electric Power Company, Inc., Q1 2026 earnings release (Form 8-K, Exhibit 99.1), “AEP Reports First-Quarter 2026 Earnings, Reaffirms Guidance and Increases Five-Year Capital Plan” to $78 billion, with remarks of Chairman, President and CEO Bill Fehrman, May 5, 2026. https://www.sec.gov/Archives/edgar/data/0000004904/000000490426000031/a1q20268kpressreleaseex991.htm
[26] Data Center Dynamics, “AEP sees contracted capacity surge to 63GW, 90% tied to data centers,” July 27, 2026. https://www.datacenterdynamics.com/en/news/aep-sees/
[27] POWER Magazine, “Regulator Approves AEP Ohio’s Landmark Data Center Tariff,” on the Public Utilities Commission of Ohio order of July 9, 2025. https://www.powermag.com/regulator-approves-aep-ohios-landmark-data-center-tariff/
[28] Eliza Martin and Ari Peskoe, Harvard Law School Environmental and Energy Law Program, Electricity Law Initiative, “Extracting Profits from the Public: How Utility Ratepayers Are Paying for Big Tech’s Power,” March 2025. http://eelp.law.harvard.edu/wp-content/uploads/2025/03/Harvard-ELI-Extracting-Profits-from-the-Public.pdf
[29] Governing (Floodlight reporting), “Power for Data Centers Could Come at ‘Staggering’ Cost to Consumers,” quoting Ari Peskoe, Director of the Electricity Law Initiative, Harvard Law School, March 2025. https://www.governing.com/resilience/power-for-data-centers-could-come-at-staggering-cost-to-consumers
[30] UC Riverside News, “Data center water spikes could cost billions,” on research by Professor Shaolei Ren (UC Riverside) with Yuelin Han, Pengfei Li (RIT), and Adam Wierman (Caltech), “Small bottle, big pipe: Quantifying and addressing the impact of data centers on public water systems,” March 9, 2026. https://news.ucr.edu/articles/2026/03/09/data-center-water-spikes-could-cost-billions
[31] Pengfei Li, Jianyi Yang, Mohammad A. Islam, and Shaolei Ren, “Making AI Less ‘Thirsty’: Uncovering and Addressing the Secret Water Footprint of AI Models,” Communications of the ACM, Vol. 68, No. 7 (2025). https://dl.acm.org/doi/10.1145/3724499
[32] MIT News / MIT Energy Initiative, “The multifaceted challenge of powering AI,” quoting Dr. Deepjyoti Deka, Research Scientist, MIT Energy Initiative, January 21, 2025. https://news.mit.edu/2025/multifaceted-challenge-of-powering-ai-0121
[33] MIT News, “Explained: Generative AI’s environmental impact,” quoting Dr. Noman Bashir, Computing and Climate Impact Fellow, MIT Climate and Sustainability Consortium and CSAIL, January 17, 2025. https://news.mit.edu/2025/explained-generative-ai-environmental-impact-0117
[34] Statista, “Big Tech’s AI Spending to Reach $760 Billion in 2026,” on the Q2 2026 earnings reports of Microsoft, Alphabet, Amazon, and Meta, July 31, 2026. https://www.statista.com/chart/35046/capital-expenditure-of-meta-alphabet-amazon-and-microsoft/
[35] Brian Sozzi, Yahoo Finance, “Meta, Microsoft, Amazon, and Alphabet are about to spend a shocking amount of money to dominate the AI era,” reporting Goldman Sachs’ $5.3 trillion FY2025–FY2030 hyperscaler capex projection, June 3, 2026. https://finance.yahoo.com/sectors/technology/article/meta-microsoft-amazon-and-alphabet-are-about-to-spend-a-shocking-amount-of-money-to-dominate-the-ai-era-115359575.html
[36] Latitude Media, “ERCOT’s large load queue has nearly quadrupled in a single year,” on speculative or “phantom” load requests, and Texas legislation directing PUC transparency rules for duplicate interconnection requests, late 2025. https://www.latitudemedia.com/news/ercots-large-load-queue-has-nearly-quadrupled-in-a-single-year/
