Introduction: The Electric Grid Could Finally See the Datacenter, but It Still Could Not Control Its Demand
On October 12, 2026, the American electricity industry crossed a regulatory threshold whose significance will be debated for years, not because of what it permitted but because of what it refused to permit. On that Monday morning, PJM Interconnection, the regional grid operator serving roughly 67 million people across thirteen states and the District of Columbia, began operating under newly effective tariff provisions that created a formal Large Load Registry, a database of every new electricity customer whose peak demand at a single site reaches fifty megawatts or more, including the enormous datacenter campuses being developed to train and serve artificial intelligence models. The Federal Energy Regulatory Commission had accepted those registry provisions three days earlier, in an order issued on Friday, October 9, in Docket ER26-3515, and had set October 12 as their effective date.[1] Yet in the very same order the Commission rejected the far more consequential element of PJM’s filing: the Interim Resource Adequacy Service, a framework under which new large loads that did not bring their own new generating capacity would have been required to accept curtailment of their electricity supply in the hours before a grid emergency, and would have been compensated for that interruption at a federally established hourly credit rate.[2]
The distinction between what was accepted and what was rejected is the intellectual starting point of this paper. The federal government accepted that the grid needed to know far more about the industrial loads arriving on its wires, and it accepted that a formal, continuously updated registry of those loads was a just and reasonable instrument for acquiring that knowledge. What the federal government did not accept was that the act of registering a load, or the fact that a load had declined to secure its own capacity, could by itself give a regional transmission organization the authority to decide which retail customers would be interrupted first when supply ran short. The Commission’s language was unusually direct about where it believed that authority resided. In rejecting the curtailment framework, FERC wrote that “several provisions of the proposed tariff appear to place obligations directly on eligible new large loads, which are, for the most part, retail customers; that certain other provisions appear to dictate which retail customers taking retail service will be curtailed; and that the proposed compensation framework appears to establish a retail rate.”[2] In other words, the problem was not that PJM wanted information, nor even that PJM wanted flexibility from datacenters; the problem was that PJM had attempted to translate information directly into a priority of service without the intervening consent of the states that regulate retail electricity.
Imagine, then, the position of a grid planner confronting this new landscape on the morning the registry went live. A developer announces a multibillion-dollar artificial intelligence campus requiring several hundred megawatts of electrical service. The proposed facility appears in corporate investor presentations, in a governor’s economic-development press release, in a county planning board agenda, in the utility’s integrated resource plan, and quite possibly in two or three separate interconnection studies filed with different utilities in different states as the developer hedges its bets on which jurisdiction can deliver power first. Investors applaud the anticipated project. State officials count the construction jobs and the property-tax revenue. The utility’s engineers begin evaluating substations, transformers, transmission corridors, and generating capacity, and they begin contemplating financial commitments that may extend decades into the future and that will, in one form or another, be recovered from the entire body of ratepayers.
But the questions that actually determine whether any of this electricity will be consumed remain unanswered, and they remain unanswered because no institution has yet been built to answer them. Has the developer secured financing for every announced construction phase, or only for the first building? Has it signed an enforceable electricity service agreement with minimum billing obligations and termination fees, or merely submitted a non-binding request? Have the required transformers, which currently carry lead times measured in years rather than months, been ordered, manufactured, and scheduled for delivery? Will the models that will eventually run in this facility require the entire projected capacity on the day the campus is energized, or will actual consumption ramp gradually over several years as equipment is installed and workloads migrate? Is this project’s anticipated demand already counted in a neighboring utility’s forecast under a different name? And what happens to all of the infrastructure that was built in anticipation of this load if the developer postpones construction, switches locations, changes semiconductor platforms, or abandons the project altogether?
Then there is the hardest question of all, the one that the October 9 order placed squarely at the center of American energy governance: if electricity becomes insufficient during an emergency, who possesses the lawful authority to decide whether that datacenter must reduce its consumption, and on what evidentiary and procedural basis?
These are no longer hypothetical concerns confined to the engineering departments of utilities. They have become the substance of gubernatorial directives, legislative hearings, federal rulemakings, and capital-market valuations. PJM itself has projected that datacenter demand within its footprint could add roughly thirty gigawatts of peak load by 2030, the overwhelming majority of a thirty-two-gigawatt forecast increase, and its board has cited an expectation of approximately seventy gigawatts of new large-load demand by 2038, arriving on a system that has lost roughly fifteen gigawatts of generation to retirements since 2022.[3][4][5] Three consecutive capacity auctions have cleared at the FERC-approved price cap, and the most recent auction, for the 2028–2029 delivery year, procured 138,318 megawatts of capacity at $325 per megawatt-day while falling 6,831 megawatts short of the region’s reliability requirement.[6] PJM’s own simulation indicated that without the cap the auction would have cleared at $555 per megawatt-day.[7] As PJM President and Chief Executive Officer David Mills put it when the results were released:
“These auction results show that demand for electricity continues to grow faster than electricity supply.”
— David Mills, President and CEO, PJM Interconnection [8]
Yet PJM’s October regulatory experience demonstrated something that the capacity numbers alone could not: the ability to identify large electricity consumers and the ability to establish enforceable curtailment obligations are different legal, operational, and institutional questions, and the first does not automatically produce the second.
The urgency extends well beyond PJM. In Texas, Governor Greg Abbott on August 3, 2026 directed the Public Utility Commission of Texas and the Electric Reliability Council of Texas to conduct a comprehensive verification and audit of every datacenter project advancing through ERCOT’s interconnection process, and he instructed that the audit be completed before any such project moves forward.[9] The figure that justified this extraordinary intervention was itself extraordinary: ERCOT was considering more than 474 gigawatts of requests to connect to the Texas grid, a figure exceeding five times the state’s historical peak electricity demand, with approximately ninety percent of those requests associated with datacenters.[9] Those requests are not equivalent to verified future consumption; indeed, that very distinction is why their credibility required examination. Governor Abbott framed the directive in the language of consumer protection:
“Our top priority is to protect Texans’ safety and quality of life. Simply put, Texans must come first.”
— Governor Greg Abbott, State of Texas [9]
Texas also established an October 12, 2026 deadline for developers of datacenters of twenty-five megawatts or more pursuing interconnection to ERCOT to respond to a Request for Information covering public financial assistance and tax benefits, dependency on grid power and use of on-site generation, water sources and consumption, cooling technologies, noise and traffic impacts, and ownership and controlling interests, with ERCOT and the Public Utility Commission committing to publish an analytical report on or before December 10, 2026.[10] Consequently, on the same day that PJM began operating under its accepted large-load registry provisions, Texas reached an important deadline in its own effort to verify the substance behind datacenter development proposals. Two of the largest electricity markets in the United States, operating under entirely different legal frameworks, had converged on the same conclusion: the grid could not be governed on the basis of announcements.
Elsewhere, the pattern repeated with regional variations. Governor Josh Shapiro in Pennsylvania signed Executive Order 2026-05 on August 18, imposing the Governor’s Responsible Infrastructure Development requirements on every datacenter proposal above twenty-five megawatts, prohibiting nondisclosure agreements between state agencies and developers, and directing the Department of Environmental Protection to publish a public map of every proposed project that has engaged with the agency.[11][12] Governor Abigail Spanberger in Virginia unveiled a Data Center Accountability Framework and signed Executive Order 22 on September 18, organizing the Commonwealth’s response around five pillars of transparency, environmental protection, affordability, grid reliability, and workforce benefit, while previewing legislation for the 2027 General Assembly session.[13][14] Governor Gretchen Whitmer in Michigan unveiled an Affordable and Responsible Growth Action Plan in July that paired a call for the legislature to codify Michigan Public Service Commission ratepayer protections with a voluntary corporate pledge that six artificial intelligence and datacenter companies signed within forty-eight hours.[15][16]
These initiatives are not identical. Their legal authorities, implementation mechanisms, and enforceability differ substantially, and this paper will devote considerable attention to those differences. Nevertheless, they point toward a common recognition that has crystallized across the American political spectrum in 2026: the nation cannot govern an electricity-intensive artificial intelligence economy by treating investment announcements as substitutes for verified infrastructure demand.
The problem is especially significant within what I have elsewhere described as the Five-Layer AI Economy, the vertically interdependent structure in which energy, semiconductors, datacenters, frontier models, and autonomous applications each depend upon the layer beneath them. At Layer 1, energy providers must decide what generation and grid infrastructure to construct, and they must make those decisions years before the load arrives. At Layer 3, datacenter developers must translate expected digital workloads into physical electrical requirements, specifying transformer ratings and cooling loads for equipment that may not yet have been manufactured. At Layer 4, model developers must determine how training, inference, and increasingly agentic reasoning systems will affect the timing and intensity of those requirements, and their decisions can change quarterly. A single change at the model layer, whether an efficiency breakthrough of the kind that briefly unsettled markets after the DeepSeek announcement or a shift toward inference-heavy reasoning workloads that run continuously rather than in discrete training campaigns, can therefore alter assumptions that were locked into transmission plans and generation procurements years earlier at the energy layer. The grid is being asked to commit permanent physical capital on the basis of the most volatile layer of the technological stack.
This paper introduces Demand Cartography as a proposed analytical and governance framework for measuring, verifying, and understanding the geography, timing, credibility, ownership, flexibility, and financial consequences of large-scale artificial intelligence electricity consumption. Its central research question can be stated in a single sentence: could verified electricity-demand information become as essential to artificial intelligence infrastructure finance as audited financial statements are to corporate finance? The answer to that question may determine which regions can support the next generation of artificial intelligence, which developers receive scarce infrastructure capacity, which investments become economically viable, and whether America’s electricity expansion strengthens industrial competitiveness without transferring excessive risk to households and small businesses that never asked to underwrite it.
Why I Chose the Title “Demand Cartography”
I chose the title Demand Cartography because America’s artificial intelligence infrastructure expansion has created a fundamental problem that traditional electricity forecasting was never designed to solve: the difference between electricity that developers announce, electricity they request, electricity they contract for, electricity that utilities can physically deliver, and electricity that datacenters actually consume. These five quantities have always differed to some degree for every industrial customer, but for most of the twentieth century the differences were small enough, and the customers numerous enough, that statistical aggregation smoothed them away. A steel mill that overstated its load by twenty percent was absorbed into a regional forecast built from thousands of other customers. A single artificial intelligence campus that overstates its load by twenty percent, when that campus alone requests as much power as a mid-sized city, is not absorbed by anything; it becomes the forecast.
The October 12, 2026 implementation of PJM’s Large Load Registry reveals why this distinction has become urgent and why it has become legally fraught. Regulators can increasingly require better information about future electricity demand, and the registry accepted by FERC will collect a facility’s location, its ramp schedule, its capacity supply arrangements, and related operational details.[2] Yet those same regulators do not automatically possess the legal authority to determine which datacenters must reduce consumption during an emergency, because that authority, as the Commission made clear, rests substantially with the states that regulate retail service. Demand Cartography captures this emerging challenge of mapping industrial electricity demand before making consequential decisions about reliability, costs, investment, and access to power, while recognizing that the map and the authority to act on the map are governed by different institutions.
The term also fits the Five-Layer AI Economy framework because energy is the foundation upon which semiconductor fabrication, datacenter construction, frontier-model development, and autonomous applications ultimately depend. Without credible demand information, governments may authorize unnecessary infrastructure, utilities may misallocate investment, financial institutions may misjudge project viability, and households may face costs associated with projects that never become operational. The scale of the capital now in motion makes these errors consequential in a way they have never been before. The four largest hyperscale cloud providers guided to combined 2026 capital expenditures of roughly $730 billion following their second-quarter earnings reports in late July, nearly double the approximately $410 billion they spent in 2025, with Amazon raising its outlook to approximately $220 billion, Alphabet raising its range to $195–205 billion, Microsoft guiding calendar-year spending to roughly $175 billion, and Meta narrowing its range upward to $130–145 billion.[17][18] NVIDIA reported second-quarter fiscal 2027 revenue of $96.2 billion, of which $89.0 billion came from its datacenter segment, and guided the following quarter to $108 billion.[19] Every dollar of that spending ultimately resolves into a physical demand for electrons at a specific substation on a specific date, and the grid has no reliable instrument for learning which of those dates are real.
From Pennsylvania and Virginia to Michigan, Texas, and California, the next phase of artificial intelligence expansion will therefore require more than constructing power plants and server campuses. It will require an independently verifiable understanding of where electricity is needed, when it will be consumed, who will pay for it, and how competing economic and social priorities should be reconciled. Cartography, properly understood, is not merely the drawing of maps; it is the disciplined representation of reality through measurable information, consistent definitions, and verifiable observations. That discipline is what the American electricity system now lacks, and what this paper attempts to describe.
Section 1: From Electricity Demand Forecasting to an AI Infrastructure Census
The first task of any new governance framework is to explain why the existing one has failed, and in the case of electricity demand the explanation requires some patience, because the existing framework did not fail through negligence or incompetence. It failed because the phenomenon it was asked to measure changed its nature. For nearly a century, electricity forecasting was a statistical discipline built upon the behavior of populations, and populations behave predictably in aggregate even when individuals do not. The artificial intelligence datacenter is not a population; it is a single actor whose decisions can be as large as a population’s, and single actors do not obey the law of large numbers.
1.1 The Limits of Traditional Electricity Forecasting
Conventional electricity forecasting developed around a set of well-understood drivers: historical consumption patterns, weather normalization, regional economic activity, industrial production indices, demographic growth, appliance saturation, and the slow diffusion of new end uses such as air conditioning in the 1950s and 1960s. Utilities and regional operators built econometric models that related these drivers to load, calibrated the models against decades of hourly data, and produced forecasts whose errors, while never zero, were bounded and well characterized. MISO, for example, has long required its load-serving entities to supply hourly forecasts extending seven days ahead and maintains its own five-minute short-term load forecast for real-time dispatch, and it has observed that a single degree of temperature forecast error can translate into as much as two thousand megawatts of load forecast error across its footprint.[20] That is a large number, but it is a number with a known cause, a known distribution, and a known remedy in the form of operating reserves.
The arrival of hyperscale artificial intelligence campuses broke this statistical foundation in three distinct ways. First, the magnitude of individual requests became comparable to that of established industrial districts or entire metropolitan areas. A single campus requesting five hundred megawatts to a gigawatt of service is not an incremental perturbation to a regional forecast; it is a discrete step change whose timing depends on the private decisions of one developer. Second, the requests arrived with unprecedented velocity. ERCOT was tracking approximately fifteen gigawatts of large-load requests in December 2022; by August 2026 the figure exceeded 474 gigawatts across more than 1,800 requests, and ERCOT had told the Texas Senate that roughly ninety percent of that volume was datacenters.[9][21][22] No econometric model calibrated on the 2000s could have anticipated a thirtyfold increase in a single category of request over four years. Third, and most fundamentally, the requests were not observations of behavior but declarations of intent, and declarations of intent are governed by strategy rather than by statistics.
The North American Electric Reliability Corporation captured the aggregate consequence in its 2025 Long-Term Reliability Assessment, released in January 2026, which projected that summer peak demand across the bulk power system would grow by 224 gigawatts over the following ten years, an increase of more than sixty-nine percent over the 132 gigawatts projected only one year earlier, with winter peak growth projected at 245 gigawatts.[23] The assessment was blunt about the historical novelty of these figures:
“Electricity peak demand and energy growth forecasts over the 10-year assessment period continue to climb higher than at any point in the past two decades.”
— North American Electric Reliability Corporation, 2025 Long-Term Reliability Assessment [24]
NERC identified new datacenters as the principal source of load growth in many assessment areas, with thirteen of twenty-three areas facing elevated or high resource-adequacy risk over the next five years, and it projected that PJM’s summer peak would grow by fifty-six gigawatts to 210 gigawatts by 2035 while ERCOT faced forecasted large-load interconnections totaling forty-five gigawatts by 2030, of which twenty-three gigawatts were datacenters.[23] Yet NERC was equally candid that its assessment was built from a mid-2025 “snapshot in time” based on utility and market projections, which is to say that its headline numbers inherited whatever credibility problems afflicted the underlying requests.[24]
It is important to distinguish ordinary forecast uncertainty from the fundamentally different problem of evaluating proposed facilities whose construction schedules, customers, financing, and ultimate electricity requirements may remain unsettled. Ordinary forecast uncertainty concerns how much a known population will consume under unknown future conditions. The datacenter problem concerns whether a proposed consumer will exist at all, and if so, when, where, and at what scale. These are not the same analytical task, and conflating them has produced forecasts that are simultaneously too high in aggregate and too low in the specific locations where real projects are actually being built. The central argument of this section can therefore be stated simply: forecasting future consumption and verifying the existence of future demand are related but distinct analytical tasks, and the American grid has institutions for the first but not the second.
The academic debate over the magnitude of the problem illustrates how much depends on this distinction. Stanford economist Frank Wolak, the Holbrook Working Professor of Commodity Price Studies, has argued that artificial intelligence demand fits a long historical pattern of electricity booms and that the harder challenge is cleaning the grid rather than expanding it, observing that American electricity use grew roughly seven percent annually from 1900 to 1970.[25] He offered a characteristically skeptical assessment of the forecasts now circulating:
“Certainly my view is that [AI demand] is fundamentally not much different.”
— Frank Wolak, Holbrook Working Professor of Commodity Price Studies, Stanford University [25]
Wolak further cautioned that “everybody forecasts a lot of demand,” but that many projects never materialize.[25] Princeton’s Jesse Jenkins, who leads the ZERO Lab, offered a similar perspective when he told reporters that the situation was “not a crisis” and that even a fourfold acceleration in demand growth would leave the country “a long way from getting back to the pace of growth that we saw” between 1980 and 2005, when demand grew 2.3 percent per year.[26] Jonathan Koomey and colleagues at the Bipartisan Policy Center estimated that datacenters would account for at most twenty-five percent of electricity demand growth by 2030 nationally, with onshoring, vehicle electrification, and building electrification contributing more, while emphasizing that the trajectory is uncertain because of improvements in model efficiency, unpredictable demand for AI services, and manufacturing limits on chips and servers.[27] These scholars are not denying that datacenters matter; they are observing that national aggregates obscure regional concentration, and that the forecasting apparatus is being asked to adjudicate a credibility question it was never built to answer.
1.2 Why AI Datacenters Require a New Infrastructure Census
If the forecasting problem is at bottom a problem of verifying the existence of specific projects, then the instrument that addresses it must be a census rather than a model. A census does not estimate; it enumerates. It records each covered entity individually, by location and by responsible party, and it updates those records as circumstances change. The concept I propose is a geographically organized, regularly updated registry of major industrial electricity consumers, and the design principles for such a registry can be derived directly from the questions a grid planner needs answered.
An effective census would record, for each covered project, the physical location of the facility and its electrical point of interconnection; the legal entity responsible for the project and, where appropriately protected, the identity of the ultimate customer or tenant; the requested capacity in megawatts, distinguished by phase; the construction stage, defined by observable milestones such as site control, permitting, financing close, groundbreaking, and equipment delivery; the expected energization date and ramp schedule; the planned power sources, including grid supply, on-site generation, storage, and backup capacity; and the intended expansion schedule with the contractual conditions attached to each expansion option. The census would also record what the project is not, which is to say whether the same commercial opportunity has been submitted to another utility or another region under a different name.
PJM’s Large Load Registry, accepted by FERC and effective October 12, 2026, is a significant regional step in this direction, and it deserves to be understood precisely. The registry applies to loads of fifty megawatts or more of cumulative peak demand at a single site, and it collects location, service territory, ramp schedule, capacity supply arrangements, and telemetry specifications.[1][2][28] PJM proposed to update the registry periodically while requiring parties to promptly report changes, and the registry was explicitly designed as the data foundation for the broader large-load programs that PJM’s board endorsed in July 2026.[4] The PJM Board of Managers had rejected a member-endorsed proposal for a purely voluntary registry on the ground that it “would not provide sufficient assurance” of covering the region’s capacity shortfall, and it adopted instead a mandatory registry of new customers with at least fifty megawatts of peak demand.[3] Interestingly, the board was candid about the limits of its own jurisdiction even as it proposed the registry; as PJM’s leadership noted in the filing, “Because PJM does not have jurisdiction to allocate retail costs directly to individual data centers, state action will be essential.”[3]
What the PJM registry is not must be stated with equal clarity. It is not a comprehensive national artificial intelligence datacenter census. It covers one regional transmission organization, and it covers new large loads rather than the full existing population. Its definitions are PJM’s definitions, not necessarily those of ERCOT, MISO, or CAISO, each of which uses different thresholds and different vocabularies. It does not, under the order as accepted, carry any curtailment consequence, which means that the incentive for a developer to report accurately is the ordinary incentive of tariff compliance rather than any direct link between the quality of the information and the quality of the service received. And it is not yet integrated with the state-level disclosure regimes that Pennsylvania, Virginia, Texas, and Michigan have begun constructing in parallel. The registry is a foundation stone, and foundation stones are important, but no one should mistake one for a building.
1.3 The Difference Between Megawatts, Megawatt-Hours, and Deliverable Capacity
Any discussion of datacenter electricity demand that does not begin with units will end in confusion, and much of the public debate of 2025 and 2026 has in fact been conducted in a state of unit confusion that serves no one well. The distinction between power and energy, between megawatts and megawatt-hours, is elementary to engineers and yet routinely collapsed in press releases, legislative hearings, and even some regulatory filings. A megawatt is a rate, the instantaneous capacity to consume or deliver electricity. A megawatt-hour is a quantity, the electricity actually consumed over time. A facility rated at five hundred megawatts that operates at an average utilization of seventy percent consumes 350 megawatt-hours in each hour and roughly 3.07 million megawatt-hours in a year, and the difference between the nameplate rating and the consumed energy is precisely the space in which both forecasting error and strategic overstatement live.
A campus requesting five hundred megawatts of grid service does not necessarily consume five hundred megawatts continuously, and in practice it almost never does. Its actual consumption depends on operating conditions, facility utilization, cooling requirements that vary with ambient temperature, maintenance cycles, and workload scheduling. MIT’s Christopher Knittel and his colleagues at the Center for Energy and Environmental Policy Research observed in their work on flexible datacenters that facilities do not operate consistently at full capacity and typically maintain utilization rates around eighty percent.[29] More recent research on hybrid artificial intelligence datacenters that mix training and inference workloads on shared accelerators has found that workload composition shapes grid-facing behavior in ways that cannot be inferred from annual energy use alone, with aggregate demand smoothing while short-horizon ramping persists; the authors argue that such facilities should be understood “not simply as large electric loads but as dynamic systems whose internal workload interactions shape grid-facing demand.”[30]
Demand Cartography therefore requires a precise technical vocabulary with at least four distinct quantities, and each should be reported separately rather than collapsed into a single headline number. The first is the maximum grid-service requirement, the contractual or tariff capacity the utility must stand ready to deliver, which determines transformer and feeder sizing. The second is the installed information-technology capacity, the electrical load that the installed servers, accelerators, and networking equipment could draw at full utilization, together with the mechanical overhead implied by the facility’s power usage effectiveness. The third is the contracted electricity supply, the energy and capacity the customer has secured through tariff service, power purchase agreements, or co-located generation. The fourth is actual metered consumption, the interval data recorded at the point of delivery. These four numbers can differ from one another by factors of two or more during a facility’s ramp-up, and only the fourth is an observation rather than a claim.
Table 1. Four distinct quantities that are routinely conflated in public discussion of datacenter electricity demand
| Quantity | Unit | What it measures | Who determines it | Evidentiary status |
| Maximum grid-service requirement | MW | Capacity the utility must be prepared to deliver at the point of interconnection | Customer request, utility study, tariff | Claim, subject to engineering review |
| Installed IT capacity (plus PUE overhead) | MW | Load the installed equipment could draw at full utilization | Developer equipment schedule | Claim, verifiable by inspection after installation |
| Contracted supply | MW and MWh | Energy and capacity secured through tariff, PPA, or co-located generation | Executed contracts | Commitment, verifiable by document review |
| Metered consumption | MWh and peak MW | Electricity actually delivered and recorded by interval meter | Utility metering | Observation, independently verifiable |
Source: Author’s synthesis of the cited regulatory filings, corporate disclosures, and literature.
The purpose of insisting on this vocabulary is not pedantry. When a governor announces that a state has “attracted” ten gigawatts of datacenter investment, the figure almost always refers to the first quantity, the requested grid-service capacity, and almost never to the fourth. When a utility files for a rate increase to serve “load growth,” the figure may refer to the first or the third. When a grid operator reports a capacity shortfall, the figure reflects a forecast built from some blend of all four. A framework that prevents large nominal capacity figures from being mistaken for observed electricity use is the minimum precondition for rational infrastructure planning, and it is also, as Section 3 will argue, the foundation for any financial attestation of demand.
1.4 The Problem of Duplicate, Speculative, and Contingent Demand
The most corrosive feature of the current interconnection landscape is not that developers lie about their intentions, for the most part they do not, but that the system has given them every incentive to describe multiple possible futures simultaneously while committing to none. A developer evaluating sites in three states, facing uncertainty about which utility can deliver power first and at what cost, rationally submits interconnection requests in all three. A developer uncertain whether a campus will ultimately house five hundred megawatts or a gigawatt of equipment rationally requests the larger figure to reserve the option. A developer whose anchor tenant has not yet signed rationally proceeds with site control and preliminary studies while the tenant deliberates. Each of these behaviors is individually defensible, and collectively they have produced interconnection queues that no one, including the developers themselves, believes.
University of California, Berkeley energy economist Severin Borenstein described the dynamic at the Technology Policy Institute’s Aspen Forum in August 2026, observing that datacenter developers routinely file duplicate interconnection requests across multiple utility territories without intending to build all of them, angling to secure a position wherever power becomes available first, and that the result is queue data that utilities, grid operators, and regulators all know is inflated but that is nonetheless treated in public debate as a straightforward preview of what is coming.[31] Borenstein drew a historical parallel to the late 1990s, when internet-driven electricity demand projections badly overshot reality because computing became far more efficient than expected, with usage rising twentyfold while electricity consumption less than doubled.[31]
The quantitative estimates of how much of the queue is real vary, but none of them is reassuring to anyone who has treated the headline figures as a forecast. Wood Mackenzie estimated in August 2026 that American grid operators would realistically commit to roughly twenty-eight percent of the 1,066 gigawatts requested for datacenter projects, leaving an estimated 768 gigawatts of “phantom load” that will never materialize, and its analysts did not expect the visibility problem to resolve before 2028.[32] Oxford Economics, applying a three-layer screening methodology to a national connection queue, concluded that six of every seven megawatts of connection requests could be characterized as phantom demand, with the removal of requests unlikely to proceed explaining almost ninety percent of the difference between total requests and expected draw at maturity, and with connection requests typically including a buffer above expected peak needs even for projects that do proceed.[33] An Electric Power Research Institute survey of utilities found that of the ten respondents reporting aggregate datacenter requests equal to fifty percent or more of their present peak load, none expected an actual five-year share above thirty-five percent.[34] Exelon reportedly cut its own datacenter pipeline by nearly forty percent, to an eleven-gigawatt queue it considered credible.[32]
The same underlying commercial opportunity can therefore appear in several planning documents or proposed locations, and the task of Demand Cartography is to develop methods for identifying such duplication without unfairly penalizing legitimate developers who are doing nothing more than prudent site selection. Several instruments have emerged. Texas Senate Bill 6, enacted in June 2025, requires large-load customers above seventy-five megawatts to pay a flat screening-study fee of at least $100,000, to demonstrate site control, to disclose on-site backup generation capable of serving at least half of their demand, and to disclose whether they are pursuing substantially similar interconnection requests elsewhere in Texas that could result in a material change to or withdrawal of the request under study.[35][36] The Public Utility Commission of Texas’s subsequent rulemaking has proposed financial security of $50,000 per megawatt of requested peak demand and a non-refundable interconnection fee of the same magnitude on contracted demand.[35] Ohio regulators approved a settlement requiring large customers to commit to twelve-year contracts before a utility invests in new substations or transmission upgrades on their behalf, and Google was the first company to accept those terms, a decision Borenstein suggested might prove competitively advantageous for companies with balance sheets large enough to absorb long-term commitment risk.[31] FERC’s June 2026 show-cause orders to all six jurisdictional grid operators directed attention to whether tariffs should define large loads, establish readiness requirements, deter speculative requests, and require timely studies.[37]
The limitation shared by all of these instruments is jurisdictional. Texas can require disclosure of duplicate requests within Texas; it cannot require disclosure of a duplicate request filed in Virginia. Ohio can condition its utilities’ investment on long-term contracts; it cannot see whether the same tenant has signed a competing contract in Georgia. A developer filing in two regional markets faces no obligation anywhere to reveal the duplication, and the only institution with a vantage point across all regions, the federal government, has so far chosen to act through region-by-region show-cause proceedings rather than through a uniform rule.[37] The objective of Demand Cartography is not to discourage industrial investment, and it is certainly not to punish developers for the ordinary practice of evaluating alternatives; the objective is to prevent unverified proposals from distorting infrastructure planning, and that objective cannot be achieved by any single state or any single grid operator acting alone.
1.5 Demand Cartography Within the Five-Layer AI Economy
Electricity planning cannot be separated from the economic activities that create the demand for electricity, and in the artificial intelligence economy those activities are organized in a vertical stack whose layers interact in ways that conventional load forecasting does not capture. At Layer 1, the energy system must establish the availability and cost of electricity at specific locations and times, committing capital to generation, transmission, and distribution assets with lives measured in decades. At Layer 2, semiconductor designers and fabricators determine the computational throughput and the thermal design power of each generation of accelerator, and the shift from Hopper to Blackwell to Vera Rubin has roughly doubled rack-level power density in successive generations. At Layer 3, the datacenter translates equipment deployment into electrical demand, specifying the transformers, switchgear, and cooling plant required to serve a particular density of compute. At Layer 4, frontier-model training, inference, and reasoning systems shape utilization patterns, determining whether a facility runs flat out for months during a training campaign or follows diurnal patterns driven by user demand. At Layer 5, applications and autonomous agents generate the ultimate demand for computation that flows down the stack.
The crucial insight for Demand Cartography is that the relationship between nominal infrastructure capacity and actual electricity consumption is determined not at Layer 1, where the grid lives, but at Layers 2 through 5, where the grid has no visibility. Changes in model efficiency, semiconductor performance, application demand, and agentic workloads can alter that relationship substantially and rapidly. The International Energy Agency’s analysis has emphasized the breadth of the uncertainty: its Energy and AI report projected that global datacenter electricity demand would more than double to approximately 945 terawatt-hours by 2030, with demand from AI-optimized facilities more than quadrupling and the United States accounting for nearly half of global demand growth, yet its April 2026 Key Questions update noted that datacenter electricity demand surged seventeen percent in 2025 as technology companies spent more than $400 billion on infrastructure, with on-site and gas-fired generation, battery storage, and the intermittent profile of AI loads all complicating the supply picture.[38][39] IEA Executive Director Fatih Birol framed the strategic stakes:
“The IEA was early in recognising that there is no AI without energy and that countries that provide secure, affordable and rapid access to electricity will be one step ahead.”
— Fatih Birol, Executive Director, International Energy Agency [39]
The Lawrence Berkeley National Laboratory’s work illustrates how rapidly the bottom-up picture has moved. Its 2024 report to Congress estimated that American datacenters consumed 176 terawatt-hours in 2023, or 4.4 percent of national electricity, and projected a range of 325 to 580 terawatt-hours by 2028, or 6.7 to 12.0 percent.[40] Its 2025 update, published in June 2026, raised the reference-case estimate to 649 terawatt-hours in 2030, or 11.8 percent of national electricity, within a sensitivity range of 9.5 to 15.3 percent, with the estimates derived from a bottom-up model of planned equipment shipments, per-device energy use, and cooling-system performance.[41] EPRI’s Powering Intelligence 2026 scenarios projected that datacenters would consume nine to seventeen percent of American electricity by 2030, a range sixty percent higher than EPRI’s prior scenarios, reflecting the accelerated pace of development.[42] NVIDIA’s chief executive, Jensen Huang, described the layer-four dynamics driving these revisions when the company reported its results in August 2026:
“AI has reached its inflection point. It’s doing useful work. Its tokens are productive and profitable. Now, compute is revenue.”
— Jensen Huang, Founder and CEO, NVIDIA [19]
Huang added that “this time last year, one lab alone was driving the buildout; today, we have a golden age of new AI labs and startups, multiple frontier labs scaling in parallel, a thriving open-model ecosystem and physical AI coming online.”[19] Each of those parallel scaling efforts is a separate claim on electricity at Layer 1, and the grid cannot distinguish among them.
The section thesis follows directly. America’s artificial intelligence economy requires a transition from forecasting electricity demand as an aggregate economic outcome, the twentieth-century approach that treated load as a statistical consequence of population and GDP, toward verifying major industrial loads as geographically specific, financially accountable projects whose existence, timing, and scale are established by evidence rather than inferred from announcements. That transition is the subject of the remainder of this paper.
Section 2: How PJM, ERCOT, MISO, and CAISO Measure and Verify Large Industrial Loads
The United States does not operate a single electricity market, and any framework that pretends otherwise will fail at the first jurisdictional boundary it encounters. The country’s bulk power system is organized into regional transmission organizations and independent system operators whose market designs, governance structures, resource mixes, and relationships with state regulators differ profoundly. PJM runs a mandatory forward capacity market under full FERC jurisdiction. ERCOT operates an energy-only market that lies almost entirely outside FERC’s reach because its grid does not synchronously interconnect with the rest of the country. MISO spans fifteen states and a Canadian province, most of them vertically integrated and state-regulated, with a residual capacity auction that plays a far smaller role than PJM’s. CAISO operates within a state whose energy agencies, under a 2022 memorandum of understanding, coordinate resource planning, transmission planning, and procurement in service of statutory decarbonization goals. The large-load problem has arrived in all four regions simultaneously, and each has responded in a manner shaped by its institutional inheritance. This section examines those responses not as a catalogue but as a comparative study of how different information problems arise from different market designs.
2.1 PJM: A Regional Registry Without Automatic Curtailment Authority
The sequence of events that culminated in the October 9 order began long before August 2026. FERC’s December 18, 2025 order in the PJM co-location proceeding had already found PJM’s tariff unjust and unreasonable for lack of consistency in serving co-located load, for failing to allocate costs to the entities that cause them, and for maintaining behind-the-meter generation rules designed for rooftop solar rather than gigawatt-scale datacenters, and it directed PJM to propose new transmission services and a new netting threshold for behind-the-meter generation.[43] In April 2026, FERC unanimously denied a complaint by PJM’s Independent Market Monitor that had sought a declaration that PJM possessed authority to deny transmission service to large loads that could not be reliably served; the Commission found the complaint procedurally deficient under section 206 but took pains to state that “our rejection of a procedurally inadequate complaint is unrelated to the gravity of the issues the deficient complaint raises.”[44] The Market Monitor’s position, stated in its complaint, framed the stakes starkly:
“PJM can and must say ‘no’ to the undisciplined interconnection of data center and AI large loads to the PJM transmission system to protect grid reliability unless and until there exists sufficient resource adequacy.”
— Independent Market Monitor for PJM (Monitoring Analytics), Complaint in Docket EL26-30 [44]
Against this background, PJM’s board in late July 2026 endorsed a package that departed from the member-endorsed stakeholder proposal. The package defined a Large Load as fifty megawatts or more of peak demand at a single site; established a mandatory Large Load Registry; created an Interim Resource Adequacy Service under which New Large Loads, defined as those entering service or adding incremental demand after June 1, 2027, that did not bring their own new capacity would be subject to curtailment before PJM deployed pre-emergency load management; permitted such loads to avoid curtailment by designating new capacity under a Bring Your Own New Capacity mechanism; excluded IRAS loads from the capacity-market demand curve; and proposed to compensate curtailed loads at a federally approved hourly credit rate.[4][3] PJM filed the package on August 13, 2026 in Docket ER26-3515, requesting an October 12 effective date.[1]
FERC’s October 9 order accepted the fifty-megawatt definition and the registry, finding both just and reasonable, and rejected the IRAS, the designation requirement, the BYONC rules, and the demand-curve exclusion.[1][2][45] The Commission’s reasoning rested on three grounds that are worth separating because each has different implications for Demand Cartography. The first was jurisdictional: the IRAS provisions appeared to place obligations directly on retail customers, to dictate which retail customers would be curtailed, and to establish what amounted to a retail rate, all of which lie within state rather than federal authority under the Federal Power Act’s division of jurisdiction.[2] The second concerned undue discrimination: by curtailing new large loads that lacked new capacity before curtailing other customers, the framework appeared to single out one class of customer for inferior service on the basis of a characteristic, the absence of a bilateral capacity arrangement, that other customers were not required to possess.[45] The third concerned market design: Commissioner David Rosner warned in a concurring statement that excluding IRAS loads from the demand curve could produce a capacity market that signaled abundance while the region shed load involuntarily.
“The result could be a capacity market signaling that PJM has more than enough capacity while the region sheds thousands of megawatts of demand on an involuntary basis. That’s not only economically inefficient, but also likely to prevent supply and demand from coming back into balance.”
— David Rosner, Commissioner, Federal Energy Regulatory Commission [2]
The Commission did not reject the goal; it rejected the instrument and reassigned the authority. “States have a crucial role to play in this challenge,” the order stated, “by designating which retail customers must be curtailed when the system is stressed, and we strongly encourage states to continue their important work on this issue.”[2] Rosner was more specific about the sequence he envisioned:
“We need state regulators to establish clear rules for which customers should be curtailed first when circumstances require it. And we need PJM to be ready to adjust their operations to this evolving paradigm. But such measures should only be taken when necessary to keep the system reliable, after voluntary demand response resources have been exhausted.”
— David Rosner, Commissioner, Federal Energy Regulatory Commission [2]
PJM’s response was measured. Its spokesman, Jeffrey Shields, said the organization was reviewing the order “with an eye toward how we can fulfill FERC’s requirements while recognizing the responsibilities FERC indicated belong to the states and honoring the concept that existing electricity customers should not be unfairly burdened by new, energy intensive customers in the form of higher costs or decreased reliability.”[2] Equity analysts at Jefferies observed that the ruling would potentially sustain higher capacity prices for longer, benefiting incumbent generators such as Constellation, NRG, Talen, and Vistra, and suggested that the curtailment concept might be “minimally revived in future iterations and likely led by efforts from states,” naming Ohio and Pennsylvania as the states to watch.[2]
For Demand Cartography the order’s lasting lesson is that four legal questions that PJM’s filing had bundled together must be kept analytically separate: the question of information gathering, which FERC resolved in PJM’s favor; the question of regional resource adequacy, which remains within PJM’s and FERC’s domain but which the capacity market has three times failed to resolve at the price cap; the question of electricity service obligations, which belongs to state-regulated retail service; and the question of what constitutes permissible differential treatment of customer classes, which implicates both federal and state non-discrimination principles. The research question this raises is precise: what additional procedural, evidentiary, and legal protections would be required before any grid operator could impose a special emergency curtailment framework on a category of industrial electricity consumers? The answer, this paper will argue in Section 4, almost certainly requires a state-level designation of curtailment priority, grounded in verified registry data, adopted through notice-and-comment proceedings, and applied through non-discriminatory criteria that attach to the characteristics of the load rather than to the identity of the customer.
2.2 ERCOT: Texas Confronts the Credibility of Its Interconnection Pipeline
Texas occupies a unique position in American electricity governance because the ERCOT grid, serving roughly ninety percent of the state’s load, is not subject to FERC’s plenary jurisdiction over interstate transmission, which means that the Legislature, the Public Utility Commission, and the Governor possess a concentration of authority over large-load policy that no other state enjoys. Texas used that authority twice in rapid succession. Senate Bill 6, signed in June 2025, created the statutory architecture: a seventy-five-megawatt large-load threshold, mandatory screening fees, site-control requirements, duplicate-request disclosure, backup-generation disclosure, a mandatory demand-management program, and explicit authority for ERCOT to direct large loads with qualifying backup generation to curtail or self-supply during energy emergencies after all other market services except frequency-responsive services have been exhausted.[35][36] The Texas Blockchain Council’s Lee Bratcher welcomed the statute’s disclosure provisions precisely because they addressed what the industry itself called phantom load, giving ERCOT “a more accurate picture of future load growth.”[36]
Governor Abbott’s August 3, 2026 directive then layered an executive audit on top of the statutory framework. The directive instructed the PUCT and ERCOT to verify every datacenter advancing through the interconnection process and stated that any project failing the verification and audit “must be denied” interconnection.[9][46] It also directed the agencies to obtain, from each project, information detailing the extent to which datacenters were paying their own way or depending on state financial assistance, the extent to which they were providing their own power or depending on the ERCOT grid, and the extent to which they were bringing and reusing their own water rather than consuming water needed by local communities.[9] The operational consequence was immediate. ERCOT’s August 3 market notice confirmed that it would not deliver its Batch Zero large-load classification notices by the August 7 deadline, it filed for good-cause exceptions on August 10, and the PUCT granted them on August 20 while approving conditional classification by August 31.[21] Under the ERCOT Planning Guide, each large load was to be sorted into one of three categories: load already sufficiently studied, load requiring additional study in Batch Zero, and load deferred to a future process; a load that submitted materially false information or failed to respond was to be classified as ineligible for Batch Zero.[46]
The magnitude of what was being audited bears repeating because it is the clearest illustration in the country of the gap between requested and credible demand. Approximately 205 gigawatts of large-load projects had been submitted through the Batch Zero process alone, against a historical ERCOT peak of roughly eighty-five gigawatts, and the full pipeline exceeded 474 gigawatts.[47][9] Bracewell’s Bryan Clark told the Houston Business Journal that ERCOT would request additional information from every proposed facility as part of the audit and that projects failing to respond could be removed from Batch Zero; his colleague Jared Berg captured the developer’s dilemma:
“It’s not a question of whether you get interconnected; it’s always a question of when. And to some extent, that timeline is going to dictate to what extent which of these projects are actually going to be viable in the first place.”
— Jared Berg, Partner, Bracewell LLP [47]
The October 12 Request for Information added a second layer of inquiry that is, in some respects, more revealing of where Texas policy is heading than the interconnection audit itself. The RFI, issued to all developers of datacenters of twenty-five megawatts or more pursuing interconnection, a threshold three times lower than SB 6’s, sought information on public financial assistance, grid dependency and on-site generation, water sources and consumption, cooling technologies, noise, light, and traffic, and ownership and controlling interests, with water questions developed jointly with the Texas Water Development Board pursuant to a September 14 gubernatorial letter.[10] ERCOT and the PUCT committed to publish an analytical report by December 10, 2026.[10] What Texas has constructed, in other words, is not merely a load-verification process but a state and community impact census, administered by the grid operator, covering dimensions of datacenter development that have nothing to do with electrons. This is Demand Cartography in a maximalist form, and it raises the question that Section 4 will examine: how can regulators separate legitimate high-growth investment from speculative demand without unnecessarily delaying the projects that can demonstrate financial and technical readiness?
The implications for Texas’s rapidly expanding artificial intelligence infrastructure are double-edged. The audit introduced months of uncertainty into a market that had been attracting investment precisely because of its reputation for speed, and developers who had already cleared every ERCOT gate except energization, reportedly seventeen large loads representing approximately 6.6 gigawatts ramping over five years, found themselves waiting alongside speculative requests.[21] At the same time, a credible audit that clears the queue of phantom load could, if executed well, accelerate the projects that remain by freeing study capacity, transmission planning capacity, and political goodwill that had been consumed by projects that were never going to be built.
2.3 MISO: From Annual Forecasting to Operational Demand Visibility
MISO’s approach to large loads has been shaped by two features of its institutional environment. The first is that most of its footprint consists of vertically integrated utilities whose resource adequacy is overseen by state commissions, which means MISO’s role is more operational and less market-centric than PJM’s. The second is that MISO’s engineers have been unusually forthright about the operational, as opposed to the planning, risks that computational loads pose. Through its Large Load Working Group, MISO has advanced a framework organized around interconnection reliability requirements, a Zero-Injection Generator Interconnection Agreement for co-located generation that would limit generator output so it does not exceed the associated load, reliable-operations requirements, and reserve-product enhancements.[20] The interconnection requirements focus on ramp limits, with an initial reference to a thirty-megawatt-per-minute net ramp-rate limit at the point of withdrawal; data and modeling requirements; telemetry and forecasting; and ride-through and stability assessments under which large loads would remain connected during voltage and frequency disturbances where technically feasible.[20]
The operational forecasting proposal that is most relevant to Demand Cartography is MISO’s 2026 stakeholder initiative on large-load operational forecast requirements, which contemplates that large-load operators would provide frequently updated operational forecasts, including hourly forecasts across a multiday horizon and five-minute forecasts for near-term operations, together with specifications on forecast granularity, submission timing, scheduling integration, and operating-horizon requirements.[48] This proposal responds directly to FERC’s June 18, 2026 show-cause order, which found that MISO’s tariff appeared unjust and unreasonable because it did not include provisions to mitigate operational risks associated with large loads, including provisions requiring transmission customers to provide MISO hourly forecasts, telemetry data, and other information sufficient to give MISO visibility into large-load operations, and because it did not require transmission owners to install, at the customer’s expense, equipment capable of monitoring fast-ramping loads and disconnecting them when necessary to maintain reliability.[49] MISO’s Missouri regulatory update in August 2026 reported that its large-load additions and forecasts continued to accelerate, with data-center load growing at a compound annual rate it estimated at 103 percent through 2030.[50]
It is essential to distinguish these proposed requirements from fully implemented regulatory obligations. As of this writing, the operational forecast requirements remain a stakeholder initiative with feedback solicited through the LLWG process, the reliability requirements were circulated in tariff-draft form with comments due in May 2026, and MISO’s own timeline contemplates 2027 as the expected implementation period for several operational requirements.[20] Stakeholders have raised unresolved questions about the technical basis for the ramp limit, applicability criteria, phasor-measurement-unit requirements, telemetry ownership, cybersecurity, data retention, and forecast-accuracy expectations.[20] The direction is nonetheless unmistakable, and it reveals a dimension of Demand Cartography that the planning-oriented registries of PJM and the audit-oriented approach of Texas do not address: whether large artificial intelligence workloads will require closer operational coordination between datacenter operators and electricity dispatch systems as inference and autonomous workloads become more dynamic. The July 22, 2026 event in which approximately 3.8 gigawatts of datacenter load tripped offline in Northern Virginia, the largest such event in PJM’s history, demonstrated that the operational question is not academic; a load that disconnects instantaneously in response to a voltage disturbance imposes a contingency on the grid comparable to the loss of several large generating units.[51] NERC has responded with a Large Loads Action Plan that contemplates registration criteria and reliability standards specifically for computational loads.[52]
2.4 CAISO: California’s Approach to Coordinating Large Loads and Long-Term Planning
California’s situation differs from the other three regions in a way that is easy to overlook: the absolute magnitude of its datacenter load growth is modest by comparison, but the institutional complexity of accommodating it is arguably the highest in the country. The California Energy Commission forecast as of January 2026 that datacenter load in the CAISO balancing area would increase by 1.8 gigawatts by 2030 and 4.9 gigawatts by 2040, figures that would barely register in ERCOT’s queue, yet CAISO reported that approximately 4.5 gigawatts of datacenter demand was being studied in its 2025–2026 transmission planning cycle and Pacific Gas and Electric has indicated requests for nearly two gigawatts of new datacenter demand in the San José area alone.[53][54][55] CAISO’s 2025–2026 Transmission Plan earmarked thirty-three projects to reinforce the state’s backbone network against a projected fifteen-gigawatt demand increase by 2035 driven by datacenters, industrial growth, and transport and building electrification.[56]
CAISO launched its Large Loads Initiative in January 2026 with an issue paper that candidly acknowledged the ISO plays “a relatively small role in the actual interconnection of individual loads, large or small,” because in California that function belongs primarily to the utilities under the jurisdiction of the California Public Utilities Commission.[56] A Large Loads Technical Requirements straw proposal, developed with the utilities interconnecting large datacenter loads, was posted on June 15, 2026, addressing ride-through capability, post-fault power recovery, ramp rates, monitoring and telemetry, modeling data, and commissioning.[53] Following FERC’s June 18 show-cause order, which raised concerns that CAISO lacked a clear and coordinated framework for evaluating large-load interconnection and transmission needs, lacked adequate safeguards against shifting transmission costs to existing customers, and offered limited transmission-service options for flexible or intermittent loads, CAISO designated the Large Loads Initiative as the primary forum for responding and posted a comprehensive straw proposal on August 12, 2026 with a stakeholder meeting on August 19 and comments due September 2.[57][53] The straw proposal addresses large-load definitions aligned with evolving NERC standards, data requirements and study procedures, consistency among participating transmission owners, aggregation of large-load additions in planning, planned network upgrades, mechanisms to charge large loads for ancillary services, and new transmission services for flexible large loads and for generators serving electrically proximate load.[57]
California’s approach must reconcile datacenter expansion with electrification of transportation and buildings, integration of renewable energy and battery storage, transmission construction constrained by permitting and cost, and reliability planning under Senate Bill 100’s clean-energy mandate, all within a governance structure in which the CEC forecasts demand, the CPUC orders procurement and regulates retail service, and CAISO plans transmission and operates the market under the 2022 memorandum of understanding.[56] Governor Gavin Newsom’s broader energy-policy environment, which has emphasized affordability alongside decarbonization in the face of rising retail rates, forms the political backdrop, but it is important to distinguish the roles: the Governor appoints commissioners and sets broad direction, the Legislature enacts mandates, the regulatory agencies implement them, and CAISO’s tariff is ultimately a FERC-jurisdictional document. One instructive data point from a May 2026 Stanford symposium on sustainable datacenters is that PG&E has estimated that every thousand megawatts of new datacenter demand could lower all customer bills by one to two percent as the larger customers assume a larger share of fixed grid costs, an outlook that runs counter to analyses finding that datacenters drive up bills in states like Ohio and Virginia.[55] That divergence is itself a Demand Cartography problem: whether a new load lowers or raises everyone else’s bills depends on whether the incremental infrastructure cost it causes is smaller or larger than the fixed costs it absorbs, and that comparison cannot be made without verified, project-specific demand data.
2.5 Four Grid Regions, Four Different Information Problems
Laying the four regions side by side reveals that they are not solving the same problem with different tools; they are solving different problems that happen to share a common cause. PJM’s problem is resource adequacy under a capacity-market design that has hit its price cap three times running, and its information need is a registry that can tell it which new loads have secured capacity and which have not, so that the shortfall can be attributed and, eventually, managed. ERCOT’s problem is queue credibility under an energy-only market with no capacity obligation, and its information need is verification that distinguishes real projects from options. MISO’s problem is operational stability in a footprint where fast-ramping computational loads interact with a rapidly changing resource mix, and its information need is near-real-time forecasting and telemetry. CAISO’s problem is planning coordination among multiple state agencies and utilities under a decarbonization mandate, and its information need is a consistent definition of large load and a consistent study procedure across its transmission owners.
Table 2. Comparative large-load information frameworks across four grid regions, as of October 2026
| Dimension | PJM | ERCOT | MISO | CAISO |
| Large-load threshold | 50 MW at a single site (FERC-accepted) | 75 MW under SB 6; 25 MW for impact RFI | Draft definition under LLWG; 50 MW in FERC show-cause | Definition under development, aligned to NERC |
| Primary instrument | Mandatory Large Load Registry (effective Oct. 12, 2026) | Batch Zero study process; gubernatorial audit; Oct. 12 RFI | Operational forecast, telemetry, ramp, ride-through requirements | Large Loads Initiative straw proposal (Aug. 2026) |
| Information collected | Location, ramp schedule, capacity supply, telemetry specs | Financing, site control, duplicate requests, backup generation, water, ownership, incentives | Hourly and five-minute forecasts, PMU and telemetry data, load models | Study data, aggregated additions, flexibility attributes |
| Curtailment authority | IRAS rejected by FERC; states to designate priority | Statutory: ERCOT may direct curtailment or backup deployment in emergencies | Under development; FERC seeks disconnect capability | Under development |
| Jurisdictional center | FERC (wholesale) with state retail authority | Texas Legislature, PUCT, Governor | FERC tariff with state resource adequacy | FERC tariff with CEC, CPUC, and MOU coordination |
| Status | Partially implemented | Statute in force; audit and RFI ongoing | Proposed; 2027 implementation expected | Proposed; stakeholder process ongoing |
Source: Author’s synthesis of the cited regulatory filings, corporate disclosures, and literature.
The comparison suggests where common data definitions would help and where regional differences are not merely tolerable but necessary. Common definitions would help in four areas: the threshold and method for defining a large load; the vocabulary for distinguishing requested, contracted, energized, and metered demand; the standard for disclosing duplicate requests across regions; and the format for reporting ramp schedules and phasing. Regional differences remain necessary in the design of curtailment authority, which depends on whether a state has retained retail regulation or adopted retail choice; in the treatment of capacity obligations, which exist in PJM but not in ERCOT; in the operational requirements, which depend on resource mix and weather exposure; and in the allocation of transmission costs, which depends on each region’s tariff and each state’s ratemaking. A framework that standardized the data while leaving the authority regional would respect both the physics, which is national, and the law, which is not.
2.6 What Governors Can Influence, and What They Cannot Directly Control
The year 2026 has been remarkable for the degree to which governors, rather than utility commissioners or federal regulators, have become the public face of large-load policy. This development is politically understandable, since electricity bills have become a salient issue in gubernatorial and legislative campaigns, but it carries an analytical hazard: the tendency to treat an executive directive, a legislative proposal, a utility tariff, and an independently enforceable federal reliability obligation as though they possess identical legal force. They do not, and Demand Cartography must be precise about which instruments produce which effects.
Governors possess substantial authority over permitting through the environmental and land-use agencies they direct, as Pennsylvania’s use of the Department of Environmental Protection illustrates; over economic-development incentives, which they can condition or withhold; over the appointment of utility commissioners, which shapes retail ratemaking over time; and over state agencies that gather information, as Texas’s use of the PUCT and ERCOT demonstrates. Governors do not possess authority over wholesale transmission tariffs, which are FERC-jurisdictional outside ERCOT; over regional capacity-market design; over NERC reliability standards; or, in most states, over the retail rates themselves, which are set by independent commissions through adjudicatory proceedings. A governor can direct an agency to create a map; a governor cannot, by executive order, change the order in which a grid operator sheds load, and FERC’s October 9 order makes clear that even the grid operator cannot do that without state regulatory action of a kind that typically requires either legislation or a commission rulemaking.
Table 3 arrays the instruments the four governors examined in this paper have actually deployed, together with their legal character. The point of the table is not to diminish any of these initiatives but to make visible the difference between a commitment that binds and a commitment that persuades.
Table 3. Gubernatorial instruments on datacenter electricity demand, 2026, by legal character
| State and Governor | Instrument | Date | Legal character | What it directly controls | What it cannot directly control |
| Pennsylvania (Shapiro) | Executive Order 2026-05; GRID requirements; DEP consent orders; public project map; NDA ban | Aug. 18, 2026 | Binding on executive agencies; conditions on state permits above 25 MW | DEP permitting, state incentives, agency disclosure, public map | PJM tariff, PUC rate design, curtailment priority |
| Virginia (Spanberger) | Executive Order 22; Data Center Accountability Framework; AI Task Force; NDA ban; 2027 legislative package | Sept. 18, 2026 | Binding on executive agencies; legislative proposals pending | Agency directives, toolkit, methane recommendations, cost-allocation advocacy | By-right local approvals (requires legislation), SCC rate orders, PJM rules |
| Michigan (Whitmer) | Affordable and Responsible Growth Action Plan; voluntary pledge; call to codify MPSC safeguards | July 15, 2026 | Voluntary pledge; legislative request; existing MPSC orders are binding | Public commitments of six signatories, political agenda | Enforceability of pledge, MISO rules, legislative outcome |
| Texas (Abbott) | June 10 and Aug. 3 directives; PUCT/ERCOT audit; Oct. 12 RFI; SB 6 statutory framework | Aug. 3, 2026 | Executive directive to state agencies; SB 6 is binding statute | Interconnection audit, information collection, denial of non-compliant projects | Private financing decisions, out-of-state duplicate requests |
Source: Author’s synthesis of the cited regulatory filings, corporate disclosures, and literature.
The section thesis can now be stated. Demand Cartography should produce information that is comparable across regional markets without pretending that the United States operates under a single electricity-market design or a single center of regulatory authority. Comparability is a property of data; uniformity is a property of law. The first is achievable and desirable; the second is neither, and attempts to impose it, as PJM’s IRAS experience suggests, are likely to be rejected by the institutions whose authority they would displace.
Section 3: The Four Stages of AI Electricity Demand: Announced, Contracted, Energized, and Consumed
The organizing idea of this section is that electricity demand from a datacenter is not a single quantity that exists or does not exist, but a process that unfolds through identifiable stages, each of which produces a different kind of evidence and each of which should be assigned a different degree of planning confidence. A framework that collapses these stages into a single number, whether that number is the announced capacity in a press release or the requested capacity in an interconnection application, will systematically misjudge both the magnitude and the timing of real demand. The four stages proposed here, announced, contracted, energized, and consumed, are not arbitrary; they correspond to the points at which a project crosses from one evidentiary regime into another, and they map onto the distinct institutions that observe each crossing.
3.1 Announced Demand: When Corporate Ambition Becomes a Megawatt Forecast
Announced demand is the stage at which a project exists primarily as a statement. The statement may take many forms: a corporate press release describing a multibillion-dollar campus, a land acquisition recorded in county deed books, an incentive application filed with a state economic-development agency, a preliminary development application before a planning commission, or an interconnection request filed with a utility. What unites these forms is that none of them obligates the announcing party to consume a single kilowatt-hour. The announcement is a declaration of ambition, and ambition, however sincere, is a poor foundation for a transformer order.
The magnitude of announced demand in 2026 is without precedent, and it is useful to understand its sources. The hyperscalers’ second-quarter earnings reports in late July provided the most authoritative view of the capital behind the announcements. Alphabet reported second-quarter capital expenditures of $44.9 billion, roughly double the prior year, and raised its full-year 2026 guidance to $195–205 billion from $180–190 billion, with its chief financial officer noting that about sixty percent of technical infrastructure investment went to servers and forty percent to datacenters and networking.[18][17] Microsoft reported fiscal fourth-quarter capital expenditures of $41 billion including finance leases and guided calendar-2026 spending to approximately $175 billion, a figure that reflects a lease-accounting change rather than a reduction in building.[18] Meta’s quarterly capital expenditures reached $31.08 billion, up eighty-three percent, consuming roughly ninety-eight percent of operating cash flow, and the company raised the floor of its 2026 range for the second time.[18] Amazon reported its first $200 billion revenue quarter, with AWS growing thirty-seven percent, the fastest in eighteen quarters, and raised its 2026 capital expenditure outlook to approximately $220 billion.[18] The market’s reaction was instructive: shares of Amazon, Meta, and Microsoft fell after Alphabet’s raise, and analysts spoke of “growing AI fatigue” among investors who had begun to question whether the spending would convert to revenue.[17]
Table 4. Hyperscaler capital expenditure guidance following second-quarter 2026 earnings
| Company | Report date | Quarterly capex | 2026 guidance (prior → updated) | Notable disclosure |
| Alphabet | July 22, 2026 | $44.9B | $180–190B → $195–205B | Cloud backlog reported at $514B; ~60% of infrastructure spend on servers |
| Microsoft | July 29, 2026 | $41B (FY Q4, incl. finance leases) | ~$175B calendar 2026 | Lower headline reflects lease-accounting change, not reduced building |
| Meta | July 29, 2026 | $31.08B (incl. finance-lease principal) | $125–145B → $130–145B | Capex ≈ 98% of operating cash flow; EPS miss; shares fell ~9–10% |
| Amazon | July 30, 2026 | $53.1B cash capex | ~$200B → ~$220B | First $200B revenue quarter; AWS +37% YoY |
| Four-company total | — | — | ≈ $730B at midpoints | ≈ 78% above 2025’s ≈ $410B |
Source: Author’s synthesis of the cited regulatory filings, corporate disclosures, and literature.
Every dollar in that table is a claim on future electricity, and yet the conversion from capital expenditure to megawatts is neither linear nor immediate. Servers purchased in 2026 may be installed in facilities energized in 2027 or 2028; facilities announced in 2026 may be the second or third phase of campuses whose first phase is still under construction; and a substantial share of the spending, particularly Meta’s and Alphabet’s, is being directed toward on-site or co-located generation that will never appear in a utility’s load forecast at all. The Ratepayer Protection Pledge that seven technology companies signed at the White House on March 4, 2026, committing to “build, bring, or buy” all new energy needed for their datacenters and to pay the full cost of infrastructure upgrades, has by design pushed a portion of announced demand behind the meter.[58][59] Google’s Ruth Porat described the strategy explicitly: “We’re committed to continuing to invest to add capacity where we build, and to that end, we will very shortly close on an acquisition of Intersect Power, which is going to enable us to build data centers right next to those new additional power generation capabilities where possible.”[58]
An announced project’s ultimate capacity may therefore represent a multiyear aspiration rather than an immediate requirement, and the distinction between corporate capital-expenditure intentions and independently validated electrical demand is the first distinction that Demand Cartography must preserve. Announced demand should be recorded, because it is the earliest signal of where development pressure is forming, but it should be recorded in a category of its own, with its own confidence weighting, and it should never be allowed to migrate into a planning forecast without crossing into the next stage.
3.2 Contracted Demand: When Proposed Consumption Becomes a Commercial Obligation
Contracted demand is the stage at which a project acquires a counterparty who can enforce its promises. The instruments that mark this crossing are familiar to utility lawyers: an executed electric service agreement specifying contract demand and minimum billing obligations; a financial deposit or letter of credit securing the utility’s investment; a construction commitment with milestones; a completed system-impact or facilities study with an executed interconnection agreement; and, increasingly, a long-term contract of the kind that Ohio now requires before a utility will build substations or transmission upgrades for a large customer.[31] The common feature of these instruments is that they create a legal obligation whose breach has consequences, and consequences are what convert an aspiration into a commitment.
The question of how utilities and grid operators can distinguish binding financial commitments from preliminary reservations has been answered, in practice, by the structure of the contracts themselves. The Michigan Public Service Commission’s safeguards, which Governor Whitmer has asked the legislature to codify, consist of minimum billing demand, contract termination fees, credit and collateral requirements, minimum contract terms, and related measures intended to ensure that datacenters pay the full costs they place on the system.[15] Texas’s proposed rules contemplate financial security of $50,000 per megawatt and a five-year operational requirement for capital recovery.[35] These instruments serve two functions simultaneously: they protect existing ratepayers against stranded infrastructure costs if the project fails, and they produce a credible signal of the developer’s own confidence, because a developer who posts collateral against a gigawatt of demand it does not expect to use is a developer who is losing money.
Harvard’s Ari Peskoe and Eliza Martin, in their widely cited paper on how utility ratepayers are paying for Big Tech’s power, reviewed nearly fifty regulatory proceedings and recommended that new datacenters be required to take service under published tariffs rather than secret special contracts, that new tariff classes reflect the unique risks datacenters pose to other ratepayers, and that datacenters be permitted to take service only if they commit to flexible operations or locate within energy parks with co-located new generation.[60] Peskoe explained the structural problem in terms that apply directly to contracted demand:
“A lot of this is baked into the utility business model. None of the consequences are baked in, but some of the basic mechanisms and incentives are just inherent and not unique to data centers.”
— Ari Peskoe, Director, Electricity Law Initiative, Harvard Law School [61]
What is unique to datacenters, Peskoe added, “is just the scale of this growth, and therefore the potential scale of these cost shifts.”[61] The paper’s central observation was that “utilities’ narrow focus on expanding to serve a handful of big tech companies … breaks the mold of traditional utility rates that are premised on spreading the costs of beneficial system expansion to all ratepayers.”[61] The implication for Demand Cartography is that contract terms are not merely financial protections; they are evidence. The existence, duration, and collateralization of a service agreement tells the planner something true about the probability that the load will arrive.
This raises the section’s research question: should stronger financial commitments increase the confidence assigned to projected electricity demand without automatically granting preferential rights to scarce grid capacity? The answer this paper proposes is a qualified yes on the first clause and a firm yes on the second. Stronger commitments should raise the planning weight assigned to a project, because they are costly signals that speculative projects will not send. But they should not by themselves confer a right to be served ahead of other customers, because the principle of non-discriminatory service, which FERC’s October 9 order reaffirmed, does not permit a utility to sell priority of service to the highest bidder, and because a regime in which only the largest balance sheets can secure grid access would entrench incumbents in a market whose dynamism depends on new entrants. The relationship between commitment and priority is one of the hardest design problems in large-load policy, and it is addressed further in Section 5.
3.3 Energized Demand: When the Datacenter Becomes Physically Connected
Energization is the stage at which electricity can physically flow to the site, and it is a milestone distinct from contractual commitment because it depends on the completion of physical works whose schedules are governed by supply chains, construction labor, and commissioning procedures rather than by legal documents. A fully executed service agreement does not energize a substation; transformers do. And transformers, in 2026, remain among the most constrained components in the global electrical supply chain, with large power transformers carrying lead times that have stretched well beyond two years and with the Department of Energy reconsidering distribution-transformer efficiency rules partly on national-security grounds.[2] A completed campus may initially energize only a fraction of its proposed ultimate capacity, because the first substation bay is energized before the second, because the first building’s cooling plant is commissioned before the second building is framed, and because the equipment deployment that determines installed capacity proceeds rack by rack over months.
The distinction that matters at this stage is between electricity that can physically be supplied to a site, which is determined by the rating of the energized interconnection facilities, and the electricity the site’s installed equipment is capable of consuming, which is determined by how much of the information-technology load has actually been deployed. A five-hundred-megawatt interconnection serving a facility with fifty megawatts of installed servers is energized at five hundred and consuming at fifty, and a planner who records the former as demand will overstate reality tenfold. Pacific Gas and Electric’s Flex Connect program, which allows larger distribution-connected loads to energize in months rather than years by accepting curtailable service during the interim, illustrates one way utilities are decoupling energization from firm capacity, and CAISO’s straw proposal contemplates analogous transmission services for flexible large loads.[57] MISO’s “Firm Service Step Up” concept similarly contemplates earlier energization at reduced load levels.[20] Each of these mechanisms creates a category of load that is energized but not firm, and Demand Cartography must record that status explicitly rather than treating energization as a binary.
Energization is also the stage at which the operational requirements being developed by MISO, CAISO, and NERC attach. Ride-through obligations, ramp limits, telemetry, and controlled load pickup are conditions of energization rather than of contract, and the July 22 event in PJM demonstrated why: a facility that is energized and consuming 3.8 gigawatts across multiple sites and that drops that load in a fraction of a second in response to a transmission fault has become a reliability contingency regardless of what its contract says.[51] The registry record for an energized facility should therefore include not only its energized capacity but its compliance with the operational requirements that make that capacity safe to serve.
3.4 Actual Consumption: When Metered Evidence Replaces Projection
Metered consumption is the only stage at which the quantity of interest is observed rather than claimed, and its evidentiary superiority over every prior stage should be reflected in the weight that planners, regulators, and financiers assign to it. Interval metering records energy delivered in fifteen-minute or five-minute intervals; operating records establish availability and maintenance patterns; historical peak demand establishes the facility’s coincidence with system peak; and consumption variability establishes the facility’s ramping behavior. These are different measurements serving different regulatory purposes, and a Demand Cartography framework should resist the temptation to reduce them to a single figure. Monthly energy consumption informs revenue and fuel planning. Peak demand informs capacity obligations and transformer loading. Short-duration demand changes inform operating reserves and ride-through compliance. A facility can have modest energy consumption and severe ramping, or high energy consumption and benign ramping, and the grid cares about both.
The relationship between workload and consumption is the frontier of current research, and it is where the Five-Layer AI Economy becomes empirically visible at the meter. Model training produces sustained, high-utilization load that can persist for weeks and that exhibits synchronized power swings at checkpoint boundaries. Inference produces load that follows user demand, with diurnal and weekly patterns and with sharp ramps when a new model or feature is released. Workload migration between sites, which hyperscalers increasingly use for both cost and reliability reasons, produces step changes at one meter mirrored by opposite changes at another. Cooling load varies with ambient temperature and can account for a substantial fraction of facility consumption in hot climates. MIT’s research on flexible datacenters found that temporal flexibility, the ability to shift workload across time, could lower system costs substantially, though with the important caveat that in some regions, including the Mid-Atlantic, the shifted load would be served by higher-emitting generation.[29] Knittel framed the policy choice directly:
“The potential grid crisis posed by data centers could become an opportunity
— if only policymakers act swiftly to incentivize or require flexible data center operations, such as dynamic pricing, demand response programs, or performance-based incentives tied to load-shifting capabilities.” — Christopher Knittel, George P. Shultz Professor, MIT Sloan School of Management [29]
In a June 2026 follow-up, Knittel observed that incentives such as faster grid connections in exchange for flexible consumption could help accommodate rapidly expanding datacenters without significantly increasing peak demand, noting that “one big concern about these data centers now is how long it takes for them to connect to the grid.”[62] Duke’s Tyler Norris and colleagues quantified the headroom that flexibility could unlock, finding that seventy-six gigawatts of new load, equivalent to ten percent of national aggregate peak demand, could be integrated at an average annual curtailment rate of 0.25 percent, ninety-eight gigawatts at 0.5 percent, and 126 gigawatts at one percent.[63] Norris summarized the finding:
“Our study demonstrates that the existing US power system—designed to manage extreme peak demand fluctuations—could accommodate significant load additions with modest flexibility measures.”
— Tyler H. Norris, Nicholas School of the Environment, Duke University [64]
NERC’s own reliability staff have endorsed the logic. John Moura, NERC’s director of reliability assessment, said during the LTRA briefing that datacenters capable of offloading demand to backup sites or shifting load among facilities, “if that can happen and they are committed to not being there during peak conditions,” could be interconnected more quickly.[24] The word “committed” is doing considerable work in that sentence, and it is precisely the work that metered data can verify and announced data cannot.
The challenge is that independently verified operating data of the kind these analyses require is largely unavailable to regulators, because hyperscalers treat facility-level consumption as commercially sensitive and because the workload composition that explains consumption is treated as a trade secret. Demand Cartography must therefore ask how verified operating data could improve electricity forecasts without revealing sensitive model architectures or customer workloads. The answer, developed in Section 4, lies in the distinction between the meter and the workload: the grid needs to know how much a facility consumed, when, and how fast it changed, and it does not need to know which model was training or whose queries were being served. Metering data can be verified by the utility and attested to the regulator without disclosing the computational activity behind it, in the same way that a bank’s audited financial statements disclose revenue without disclosing the identity of each customer.
3.5 An Auditable Four-Status Demand Ledger
The four stages described above can be formalized as a ledger in which each qualifying project carries four separately reported quantities, each with its own date, unit, evidentiary requirement, and responsible party. The ledger would be maintained by the regional registry, populated by developers and utilities, verified by the procedures described in Section 5, and reconciled at regular intervals against observed outcomes. Table 5 sets out the proposed structure.
Table 5. Proposed four-status demand ledger for qualifying large loads
| Status | Quantity recorded | Unit | Evidence required | Responsible party | Planning weight (illustrative) | Reconciliation trigger |
| Announced | Ultimate requested capacity, by phase; target dates | MW | Public announcement, interconnection request, incentive application, site control | Developer | Low (tracked, not forecast) | Any change in request, site, or sponsor |
| Contracted | Contract demand; minimum billing demand; collateral; term; termination provisions | MW; $; years | Executed service agreement; interconnection agreement; financial security; completed studies | Developer and utility | Medium to high, scaled by collateral and term | Contract amendment, default, or milestone miss |
| Energized | Energized interconnection capacity; firm vs. flexible service; installed IT capacity; operational compliance | MW | Commissioning certificates; utility energization record; equipment installation attestation; ride-through and telemetry compliance | Utility and developer | High for energized capacity; installed capacity verified separately | Each phase energization; compliance event |
| Consumed | Monthly energy; monthly and annual peak demand; coincident peak; ramp statistics; curtailment performance | MWh; MW; MW/min | Interval meter data; operating records; demand-response performance records | Utility (metering), verified by regulator | Highest (observation) | Monthly update; annual reconciliation |
Source: Author’s synthesis of the cited regulatory filings, corporate disclosures, and literature.
The framework would not treat these categories as interchangeable, and it would not automatically assume that every announced facility advances through them on schedule. Its most important feature is the regular reconciliation of earlier forecasts against observed outcomes: at each reconciliation, the announced and contracted quantities recorded in prior periods would be compared with the energized and consumed quantities recorded now, and the resulting error would be published, at least in aggregate, by region and by developer class. Over time this reconciliation would produce the one thing the current system most conspicuously lacks, which is an empirical conversion rate from announcement to consumption, disaggregated by project type, developer type, region, and contract structure.
The paper’s proposition is that utilities, regulators, and lenders could use such records to calculate forecast error, track project maturity, and identify potential stranded infrastructure investment before it is incurred. A utility whose contracted-to-energized conversion rate for a particular developer class has historically been forty percent would plan accordingly rather than building to one hundred. A regulator evaluating a rate case would be able to see whether the load growth the utility invoked had materialized in prior periods. A lender underwriting a transmission project would have a basis for discounting the demand on which the project’s revenue depends. None of this requires new physics; it requires only that the data already generated at each stage be recorded in a common structure and compared over time.
3.6 From Financial Statements to Electricity-Demand Attestations
The central analogy of this paper is with financial accounting, and the analogy deserves to be developed carefully because it is both illuminating and imperfect. Financial statements do not eliminate uncertainty about a company’s prospects; investors who rely on them still lose money, sometimes catastrophically. What financial statements accomplish is more modest and more durable: they impose standardized definitions on the quantities being reported, they require that those quantities be derived from verifiable records, they subject the derivation to independent assurance, and they distinguish, through the structure of the statements themselves, between what has been realized and what is merely expected. Revenue recognized is distinguished from backlog; assets owned are distinguished from options held; audited historical results are distinguished from management guidance. The result is not certainty but comparability and accountability, and comparability and accountability are what make capital markets function.
Electricity-demand reporting today has none of these properties. Definitions vary by region and by utility. Quantities are self-reported without assurance. And the structure of the reporting does not distinguish the realized from the expected; a load forecast that blends metered consumption with announced ambition is the equivalent of a financial statement that adds backlog to revenue. The proposition that follows is that similarly structured electricity-demand attestations could become a normal requirement for financing major artificial intelligence infrastructure projects, in much the way that audited financial statements became a normal requirement for public securities offerings after 1933. An attestation would state, for a given project or portfolio, the quantities in each of the four ledger categories as of a given date, the evidence supporting each, and the procedures by which that evidence was verified. It would distinguish, as financial statements do, between verified historical consumption, which can be attested with high confidence, and independently reviewed forecasts, which remain uncertain even when supported by credible assumptions and which should be presented with their assumptions disclosed.
The commercial logic for such attestations is already forming, even if the instrument has not been named. The Ohio twelve-year contract requirement, the Michigan collateral requirements, the Texas financial-security rules, and the Ratepayer Protection Pledge are all, in effect, demands that developers put verifiable commitments behind their announced demand. The IMF, in its working paper on how artificial intelligence will drive energy demand, warned that uncertainty about the net effect on electricity demand “may delay energy investments, causing higher prices,” and that if supply responses are sluggish, power prices could rise steeply, hurting consumers and possibly curbing growth of the AI industry itself.[65][66] An attestation regime is a direct response to that uncertainty: it does not resolve the uncertainty about how much electricity the AI economy will ultimately need, but it reduces the uncertainty about which specific projects are real, and it is the latter uncertainty that delays the investment decisions that matter.
The section thesis follows. The credibility of America’s artificial intelligence infrastructure expansion should be assessed through observable project milestones and verifiable demand evidence, not merely through the magnitude of announced megawatts. A gigawatt that has been contracted, energized, and metered is a fact about the grid. A gigawatt that has been announced is a fact about a press release. The two should never be added together.
Section 4: How Governors, Utilities, and FERC Can Use Load Disclosure Without Exposing Commercially Sensitive Information
The objection most frequently raised against any registry of industrial electricity demand is that it would compel the disclosure of commercially sensitive information, and the objection is not frivolous. The identity of a datacenter’s anchor tenant, the timing of its expansion phases, the capacity it has reserved against future growth, and the terms on which it has contracted for power are competitive information in an industry where site selection and speed to power are the decisive advantages. A disclosure regime that ignored these concerns would either be evaded or would drive investment to jurisdictions that demanded less. Yet the alternative, a regime in which the public bears the cost of infrastructure built to serve loads it is not permitted to know about, is politically unsustainable, as the 2026 wave of gubernatorial action demonstrates. This section examines how four states have attempted to balance disclosure against confidentiality, and it proposes a tiered structure that distinguishes what must be public from what must be restricted and from what must be operationally protected.
4.1 Pennsylvania: Governor Josh Shapiro and Public Infrastructure Accountability
Pennsylvania’s approach is distinguished by its use of environmental permitting as the point of leverage and by its creation of a genuinely public project map. Executive Order 2026-05, signed August 18, 2026 and titled “Protecting Pennsylvania Consumers from Data Center Impacts,” directs the Department of Environmental Protection to incorporate the Governor’s Responsible Infrastructure Development requirements, first announced in February and detailed in May, into its permit review for every datacenter proposal with peak demand above twenty-five megawatts, a threshold notably lower than PJM’s fifty.[12][67] The GRID requirements are organized around energy affordability, transparency and community engagement, workforce and economic development, and environmental protection, and the executive order converts them from policy aspirations into practical conditions of permitting through a template Consent Order and Agreement with DEP under which the developer commits to comply.[67] The order creates a two-track permitting framework, removes all datacenter proposals from the state’s Permit Fast Track Program, prohibits agencies under the Governor’s jurisdiction from entering nondisclosure agreements in connection with datacenter projects, directs DEP to consider new regulations on cumulative environmental impacts including backup-generator emissions, and requires operating datacenters to file annual energy and water consumption reports beginning July 1, 2027.[12][67][11] Governor Shapiro described the purpose in terms of accountability:
“The Executive Order sends a clear message: developers must demonstrate that a project can be developed responsibly and the community approves of the project in order to move forward.”
— Governor Josh Shapiro, Commonwealth of Pennsylvania [68]
He grounded the action in public concern: “Over the past year, I’ve listened to the people of Pennsylvania – and I’ve heard directly from many residents across our Commonwealth who are concerned about what data center development could mean for our communities, our environment, and our utility bills.”[11] The Governor’s office cited reports of more than one hundred datacenter facilities proposed in the Commonwealth, and the Governor’s Special Counsel for Energy Affordability was directed to engage with the Public Utility Commission to ensure that datacenter development does not shift energy costs onto households and businesses or compromise reliability.[12][11]
The feature of Pennsylvania’s approach most relevant to Demand Cartography is the publicly accessible map of proposed datacenter projects that have engaged with DEP, which displays pending permitting information and tracking status and is updated as DEP becomes aware of new proposals.[11] This map provides a particularly useful case study for distinguishing a public development map from a comprehensive operational electricity-demand registry. The DEP map answers the questions the public most wants answered: where are projects proposed, how large are they, and where do they stand in permitting. It does not, and should not, answer the questions the grid planner needs answered: what is the contract demand, what collateral has been posted, what is the ramp schedule, and what did the facility consume last month. The two instruments serve different audiences with different needs, and a Demand Cartography framework should treat them as complementary layers of a single structure rather than as competing designs. Public project visibility and confidential engineering information can coexist if the framework specifies, for each data element, which layer it belongs to and who may access it.
4.2 Virginia: Governor Abigail Spanberger and the Politics of Datacenter Transparency
Virginia is the most important datacenter jurisdiction in the world, with Loudoun County alone hosting a concentration of hyperscale capacity that has no peer, and its policy choices carry a weight that no other state’s do. Governor Spanberger’s Data Center Accountability Framework, unveiled on September 18, 2026 alongside Executive Order 22, was described by her office as “the most comprehensive and aggressive data center accountability effort in the country.”[69] The framework is organized around five pillars, requiring that developers and operators engage transparently with communities, protect the environment, improve affordability for Virginians, advance clean energy and grid reliability, and deliver a better deal for the workforce, and it is led by the Office of the Chief Energy Officer.[14][13] The Governor framed its purpose plainly:
“Virginia’s Data Center Accountability Framework is about delivering a better deal for Virginians and holding data center companies, developers, and operators accountable.”
— Governor Abigail Spanberger, Commonwealth of Virginia [69]
The executive order itself states that “future data center development must be balanced and sustainable to ensure these companies are strong innovation and economic partners of the Commonwealth.”[70]
Executive Order 22 implements several measures immediately through executive authority: it prohibits executive-branch agencies from entering or enforcing nondisclosure agreements that would prevent public disclosure of material information about proposed developments, subject to limited exceptions such as national security; it directs the Chief Energy Officer to publish a Data Center Planning and Community Engagement Toolkit for local governments; it directs recommendations within 180 days on datacenter investment in reducing methane emissions from the natural-gas infrastructure that growing demand will draw upon; it directs reforms to shield residential ratepayers from datacenter infrastructure costs, including the assignment of PJM Reliability Backstop Procurement costs to the datacenters that cause them; and it rescinds the prior administration’s 2023 and 2024 orders on artificial intelligence while retaining their published standards until replaced.[14][13] It also creates an Artificial Intelligence Task Force addressing workforce displacement, data privacy, and cybersecurity.[69]
It is essential to distinguish what the order accomplishes from what it merely proposes. The framework’s stated goal of eliminating by-right local approvals and requiring public hearing and approval processes for facilities above a size threshold cannot be achieved by executive order; it requires legislation, and the Governor has said she will pursue it in the 2027 General Assembly session, which begins in January.[13][14] The order’s cost-allocation directives operate through advocacy before the State Corporation Commission and through PJM processes rather than through direct command. The Governor declined to impose a moratorium, and some organizations criticized the framework for that omission.[13] Virginia’s experience nonetheless informs other states managing concentrated clusters of hyperscale datacenters in a specific way: it demonstrates that transparency measures, cost-allocation advocacy, and local-control reforms require three different legal vehicles, and that a governor can deliver the first immediately, influence the second over time, and only propose the third. The concurrent hearings on the proposed Dominion Energy–NextEra Energy merger, at which local officials focused on rates and datacenters, illustrate how these questions are now being litigated in every available forum.[71]
4.3 Michigan: Governor Gretchen Whitmer and the Financial Responsibility of AI Infrastructure
Michigan’s approach is instructive precisely because it relies so heavily on voluntary commitment, and because the Governor has been candid about the limits of that reliance. The Michigan Affordable and Responsible Growth Action Plan, unveiled July 15, 2026, rests on two pillars. The first is a call for the legislature to codify the safeguards the Michigan Public Service Commission has already put in place through its orders, including minimum billing demand, contract termination fees, credit and collateral requirements, minimum contract terms, and other measures to ensure that datacenters pay the full costs they impose on the system.[15] The second is the Michigan Affordability and Responsible Growth Pledge, described as the first state-level pledge in the nation, under which signatories commit to create good-paying jobs and to pay all of their own costs, including water and energy, so that residents see no additional costs.[15] Six artificial intelligence and datacenter companies signed within a day, with Google, which is developing a one-gigawatt datacenter in Van Buren Township, among the first.[15][16] Casey Katims of the U.S. Climate Alliance placed the initiative in national context:
“As Americans confront the rapid growth of AI data centers, Michigan is stepping up with fellow states to raise the bar for ratepayer protections and environmental safeguards.”
— Casey Katims, Executive Director, U.S. Climate Alliance [72]
The plan was announced amid criticism of the Governor’s support for datacenter development and her opposition to a statewide moratorium, with the Sierra Club and other environmental groups continuing to support moratorium bills until the legislature enacts stronger regulations, and with several municipalities, including Lansing and Warren, adopting or considering local moratoria.[16] The analytical value of Michigan’s experience lies in the distinction it forces between four instruments that are frequently conflated: voluntary corporate pledges, which bind only reputationally; enforceable contracts, which bind the parties to them; utility rate structures, which bind through commission orders and are enforceable against the utility; and binding state requirements, which bind through statute and are enforceable against everyone. The Governor’s plan uses the first to generate political momentum for the fourth, while relying on the third in the interim. That is a coherent strategy, but its coherence depends on the legislature acting, and the November 2026 elections will determine whether it does.
Michigan’s experience also raises the question whether financial responsibility for electricity infrastructure should be assessed against independently verified incremental demand rather than simply against announced campus capacity. A pledge to “pay all of their own costs” is only as meaningful as the method for determining what those costs are, and that method depends on knowing what the facility actually causes the utility to build and what it actually consumes. Large-load reporting of the kind proposed in Section 3 is therefore not an alternative to Michigan’s approach but its necessary complement: the pledge establishes the principle, the MPSC safeguards establish the mechanism, and the ledger establishes the measurement without which the mechanism cannot be applied. The role of such reporting in protecting residential and small-business customers from costs directly attributable to major industrial developments is not rhetorical; it is arithmetical.
4.4 Texas: Governor Greg Abbott and the Verification of Industrial Growth
Texas has been examined in Section 2 from the perspective of ERCOT’s interconnection process, and this subsection considers it from the perspective of executive strategy. Governor Abbott’s 2026 directives, the June 10 letter on large-load projects and the August 3 audit directive, together with the September 14 letter to ERCOT and the Texas Water Development Board on water questions, constitute the most aggressive use of gubernatorial authority over datacenter development in the country, and they are aggressive in a particular direction: toward verification rather than prohibition.[9][10] The directive does not cap datacenter development, impose a moratorium, or restrict siting; it requires that every project prove its compliance with existing requirements before advancing, and it threatens denial for those that cannot.[46]
The relationship among ERCOT, the Public Utility Commission of Texas, environmental and water authorities, and local governments is the institutional substance of the Texas approach. ERCOT administers the interconnection process and the RFI; the PUCT adopts the rules under SB 6 and grants the good-cause exceptions that adjusted the Batch Zero timeline; the Texas Water Development Board authorized the water questions; and local governments, whose tax abatements and community impacts are among the subjects of the RFI, are being asked, in effect, to disclose what they have given away.[10][21] The decision to route the entire inquiry through the grid operator, rather than through a new state agency or the Comptroller, reflects the practical reality that ERCOT is the only institution to which every datacenter developer in Texas must already report.
The benefits of requiring stronger evidence before expensive interconnection and infrastructure decisions proceed are evident in the Texas numbers: a 474-gigawatt queue cannot be planned for, and a transmission system built to serve even a quarter of it would impose costs on Texas ratepayers that no legislature would tolerate. The costs are equally evident: the audit paused approvals for months, including for projects that had already satisfied every requirement, and introduced into a market whose competitive advantage was speed a form of uncertainty that developers can price only by discounting Texas relative to other jurisdictions.[47] Texas therefore offers the paper’s sharpest research question: how can regulators separate legitimate high-growth investment from speculative demand without unnecessarily delaying projects that can demonstrate financial and technical readiness? The answer this paper proposes is the ledger itself. A project that can document contracted demand with collateral, energized or scheduled interconnection facilities, and either metered consumption at an existing site or verifiable equipment orders for a new one has, by definition, passed the audit; it should be able to present that documentation and proceed. A project that cannot has, by definition, not passed, and the burden of proving readiness should rest with it. An audit designed around the four-status ledger would be faster and fairer than one designed around a blanket pause, because it would allow the ready to self-identify.
4.5 The Confidentiality Problem: What Should Be Public, Restricted, or Operationally Protected?
The experiences of Pennsylvania, Virginia, Michigan, and Texas converge on a three-tier structure for disclosure, even though none of the four states has articulated it in those terms. The structure distinguishes information by the harm its disclosure could cause and by the audience that legitimately needs it, and it assigns each data element to one of three categories with corresponding access rules.
Table 6. Proposed three-tier disclosure structure for large-load registry data
| Tier | Audience | Information included | Access and protection mechanism | Rationale |
| Public | General public, local governments, press, researchers | Generalized project location (county or municipality); permitting stage; aggregated capacity requirement by phase; expected development timeline; public infrastructure commitments; public incentives received; water source category; signatory status under state or federal pledges | Published map or registry; routine updates; no NDA permitted for these elements | Democratic accountability; community engagement; prevention of secret deals |
| Restricted regulatory | State commission, grid operator, FERC, utility, independent verifier | Exact point of interconnection and connection specifications; contract demand, collateral, term, and termination provisions; customer or tenant identity where protected; detailed financing milestones; equipment schedules; confidential operating assumptions; duplicate-request disclosures across jurisdictions | Protective orders; confidentiality agreements with regulators; secure data rooms; aggregation thresholds for any public reporting; defined retention periods | Verification requires detail; competitive harm from disclosure is real; regulators already handle such data in rate cases |
| Protected operational | Grid operator control room, utility operations, reliability coordinator, designated security staff | Sensitive grid configurations; telemetry streams; protection settings; ride-through and disconnect parameters; cybersecurity-relevant data; information meeting Critical Energy Infrastructure Information criteria | CEII designation; NERC CIP controls; need-to-know access; no disclosure in regulatory filings except under seal | Infrastructure security; national security; operational integrity |
Source: Author’s synthesis of the cited regulatory filings, corporate disclosures, and literature.
Public information under this structure would include generalized project location, permitting stage, aggregated capacity requirements, expected development timelines, and relevant public infrastructure commitments, which is approximately what Pennsylvania’s DEP map and Texas’s forthcoming RFI report will disclose. Restricted regulatory information would include exact connection specifications, commercially sensitive contracts, customer identities where appropriately protected, detailed financing information, and confidential operating assumptions, which is approximately what PJM’s registry collects and what state commissions review in rate proceedings under protective order. Protected operational information would include sensitive grid configurations, cybersecurity-relevant data, and information whose exposure could compromise infrastructure security, which is already governed by FERC’s Critical Energy Infrastructure Information regulations and NERC’s Critical Infrastructure Protection standards.
Four mechanisms make the structure workable. Independent audits, conducted by qualified engineering or assurance firms under confidentiality obligations, allow restricted information to be verified without being published; the verifier attests that the contracted demand exists and is collateralized without disclosing the counterparty. Secure regulatory access, through data rooms and protective orders of the kind state commissions routinely use, allows regulators to see what they need to see. Aggregation thresholds, under which public reporting of restricted data is permitted only when at least a minimum number of projects are combined, prevent inference of individual facility data from aggregate statistics. And legal protections for critical energy infrastructure information, already in place, govern the third tier. The nondisclosure-agreement bans adopted by Pennsylvania and Virginia should be understood as applying to the first tier: they prohibit secrecy about the existence and general character of a project, not confidentiality about its engineering and commercial details, and the distinction is one that both executive orders preserve through their exceptions.
4.6 Information Authority Is Not Curtailment Authority
The analysis returns to the October 9–12 FERC decision because that decision, more than any other event of 2026, established the principle on which this section rests. The creation of a registry does not automatically establish a lawful right to interrupt service, to discriminate among customers, or to impose new financial obligations. FERC accepted PJM’s registry and rejected PJM’s curtailment service in the same order, and the juxtaposition was not accidental; it reflected the Commission’s judgment that the two requests rested on different legal foundations. The registry was an exercise of FERC’s authority over the terms of wholesale transmission service, and the Commission found it just and reasonable. The IRAS was, in the Commission’s view, an attempt to regulate retail customers and to establish a retail rate, and the Commission found that it exceeded PJM’s and FERC’s own authority while also appearing unduly discriminatory.[2][45] An observer at Energy Central summarized the Commission’s concern that the proposal “may worsen the problems it is attempting to solve.”[73]
The procedural path to lawful curtailment authority can now be described with some precision, even though no jurisdiction has yet walked it to the end. It requires, first, appropriate federal and state proceedings: a state commission rulemaking or a legislative enactment, in each state within the region, designating the categories of retail customers subject to pre-emergency curtailment and the order in which they will be curtailed, as FERC and Commissioner Rosner expressly invited.[2] It requires, second, tariff authority: once the states have designated priorities, the grid operator must file tariff provisions implementing them, and those provisions, because they govern wholesale operations, will return to FERC for approval. It requires, third, due process: the designation must be adopted through notice-and-comment procedures in which affected customers can be heard, and the application of the designation to any particular facility must be subject to review. It requires, fourth, reliability evidence: the curtailment must be shown to be necessary to maintain reliability and must be triggered only after voluntary demand response has been exhausted, as Rosner specified. And it requires, fifth, non-discriminatory treatment: the designation must attach to characteristics of the load, such as its size, its interruptibility, its backup capability, or its date of interconnection, rather than to the identity or industry of the customer, so that it does not single out datacenters as a class for inferior service relative to other large loads with the same characteristics.
Texas, uniquely, has already walked part of this path, because SB 6 is a legislative enactment that designates non-critical loads above seventy-five megawatts connected after December 31, 2025 as subject to curtailment during firm load shed and that authorizes ERCOT to direct backup deployment or curtailment in emergencies after other services are exhausted.[35][36] The Texas statute is the closest existing model for what FERC’s order asks the PJM states to produce, and it is instructive that it was enacted by a legislature rather than imposed by a grid operator.
The question of who deserves power must therefore be translated into transparent legal and economic standards rather than subjective political preference. A registry answers the question of who is there; it does not and cannot answer the question of who should be served first, because that question is distributional, and distributional questions in a democracy are answered by legislatures and commissions acting under law. The section thesis follows: effective electricity disclosure should make infrastructure governance more accountable while preserving lawful confidentiality, due process, national security, and the legitimate commercial interests of developers. Disclosure is the precondition for governance; it is not a substitute for it.
Section 5: A Proposed 2027–2030 National Framework for Verifying AI Electricity Demand
The preceding sections have diagnosed a problem and described how four grid regions and four states have begun to address it. This section proposes a framework, and it does so with a deliberate modesty of ambition. The framework does not create a new federal agency, does not preempt state authority over retail service, and does not require Congress to act, although congressional action could accelerate it. What it proposes is a common vocabulary, a set of verification procedures, a method for linking demand to infrastructure readiness, principles for cost allocation, a phased timeline, measurable outcomes, and an honest account of the political economy that will determine whether any of it happens. The federal government’s own posture makes this approach necessary: having directed FERC in October 2025 to issue a uniform large-load interconnection rule by April 30, 2026, the Department of Energy saw FERC instead choose region-by-region show-cause proceedings in June 2026, and five of the six grid operators subsequently sought delays rather than filing reforms on the August deadline.[74][37][32] A framework that depends on federal uniformity is a framework that will wait indefinitely.
5.1 Establishing a Common National Demand-Reporting Vocabulary
The first element is a consistent reporting standard distinguishing requested, contracted, energized, and metered electricity demand, as developed in Section 3, applied across every regional registry regardless of the threshold or tariff under which that registry operates. The standard would specify, for each of the four categories, the unit of measurement, the date convention, the evidentiary requirement, and the responsible reporting party, so that a megawatt recorded as contracted in PJM means the same thing as a megawatt recorded as contracted in MISO. It would also specify the treatment of the cases that most frequently produce double-counting and confusion. Phased projects would be recorded phase by phase, with each phase carrying its own four-status record and with the ultimate campus capacity reported only as the sum of phases that have reached at least the contracted stage. Colocation arrangements, in which a datacenter connects behind the meter of an existing generator, would be recorded with both the gross load and the net grid draw, following the definitions FERC established in its December 2025 PJM order.[43] Self-generation, energy storage, and backup generation would be recorded as attributes of the facility with their capacities, fuel types, and the operational conditions under which they can be dispatched, consistent with Texas’s SB 6 disclosure requirements and MISO’s zero-injection framework.[35][20] Facilities obtaining electricity through multiple sources would report each source separately, so that the grid draw, the behind-the-meter supply, and the contracted off-site supply could be distinguished.
The objective would be interoperability among regional registries rather than the creation of an entirely new federal permitting authority. The model is the financial-reporting system, in which the Financial Accounting Standards Board establishes definitions and the Securities and Exchange Commission requires their use, but in which the reporting itself is performed by companies and verified by private auditors rather than by a federal bureau. The analogous institutions in electricity already exist: NERC, through its Large Loads Action Plan, is developing registration criteria and reliability standards for computational loads and could house the definitional standard; FERC, through its show-cause proceedings, could require the use of that standard in each regional tariff; and state commissions could adopt it by reference in their own large-load rules.[52][37] The Texas RFI, the PJM registry, and the Pennsylvania DEP map are already collecting overlapping data in incompatible formats, and the cost of harmonizing them now, while each is new, is a fraction of the cost of harmonizing them after each has accumulated years of records.
5.2 Developing Independent Verification and Assurance Procedures
The second element is independent assurance of the demand claims recorded in each category, and the design question is what evidence should be required and who should examine it. For announced demand, the evidence is documentary and minimal: the announcement itself, the interconnection request, and the site-control instrument, verified by the utility as received. For contracted demand, the evidence is the executed service agreement, the collateral or letter of credit, the completed system-impact study and executed interconnection agreement, and the financing milestones the developer is willing to disclose under protective order, verified by an independent assurance provider who attests to their existence and material terms without publishing them. For energized demand, the evidence is the commissioning certificate for the interconnection facilities, the utility’s energization record, and an engineering attestation of installed equipment capacity, verified by the utility and by the independent provider on inspection. For consumed demand, the evidence is interval meter data and operating records, verified by the utility’s metering function and attested to the regulator.
Whether assurance should be conducted by qualified engineering firms, independent auditors, regulated utilities, or combinations of those institutions is a question on which reasonable designs differ, and this paper proposes a combination calibrated to the nature of each category. Engineering firms are suited to attestations of installed capacity and operational compliance, because those are engineering facts. Financial auditors are suited to attestations of contract terms, collateral, and financing milestones, because those are financial facts and the audit profession already possesses the confidentiality infrastructure and professional-liability framework that such attestations require. Regulated utilities are suited to attestations of energization and metering, because those are facts within their custody. In each case, the attesting party should be independent of the developer whose claim it attests, should be subject to professional standards and liability, and should be compensated by the developer under a structure that does not permit the developer to select among attesters on the basis of leniency, which is the same structure that governs financial auditing.
The paper insists on a distinction that financial auditing also maintains: factual verification is different from assumptions about future commercial success. An auditor attests that a contract exists and says what it says; the auditor does not attest that the counterparty will perform. A demand attestation would similarly confirm that contracted demand is contracted and that metered demand was metered; it would not confirm that the facility will be profitable, that the tenant will renew, or that the model the facility was built to train will succeed in the market. Those are commercial risks that lenders and investors must price themselves, and an attestation regime that pretended to eliminate them would be worse than none at all, because it would invite misplaced reliance of exactly the kind that financial statements, properly understood, do not invite.
5.3 Connecting Demand Forecasts to Physical Infrastructure Readiness
The third element addresses a gap that verification alone cannot close: even a perfectly verified demand schedule is useless to a planner who cannot tell whether the generation, transformers, substations, and transmission upgrades needed to serve it will exist on the dates the demand arrives. The relationship between credible future demand and infrastructure readiness is bidirectional, since developers choose sites partly on the basis of utility delivery dates and utilities commit to upgrades partly on the basis of developer schedules, and a framework that recorded only one side of the relationship would reproduce the current dysfunction in which each party waits for the other to commit.
The proposal is that utilities and grid operators maintain, alongside the demand ledger, a readiness record for each covered project that lists the infrastructure required to serve each phase, the status of each item, and the realistic delivery date, and that the two records be compared at each reconciliation. The comparison would produce, for each project, a readiness gap: the difference between the date the developer expects to need the capacity and the date the utility expects to be able to deliver it. Periodic readiness assessments aggregating these gaps by region would identify mismatches between datacenter construction plans and the infrastructure required to supply them before either side had sunk the capital that makes the mismatch costly. NVIDIA’s disclosure that it carries $105 billion in supply and capacity obligations alongside infrastructure guarantees for AI cloud power and land suggests the scale at which the semiconductor layer is now underwriting the energy layer’s readiness, and it suggests how much is at stake in getting the readiness record right.[19]
The readiness record would also expose the constraint that Borenstein identified as decisive regardless of how the queue numbers resolve: the equipment needed to build new capacity is not available at the scale required, which means that near-term demand cannot be accommodated even if every megawatt of it were real.[31] A readiness record that showed transformer delivery dates in 2029 for projects seeking energization in 2027 would communicate that constraint to developers, lenders, and policymakers in a form that no aggregate forecast can, and it would redirect speculative requests away from locations that cannot serve them toward locations that can.
5.4 Developing Fair and Transparent Electricity Cost Allocation
The fourth element is the use of verified incremental demand to allocate the costs of grid investment, and it is the element on which the political legitimacy of the entire framework depends. The principle that new large loads should pay the costs they cause is now nearly universal in American energy politics, endorsed by the White House’s Ratepayer Protection Pledge, by Governors Shapiro, Spanberger, Whitmer, and Abbott, by FERC’s December 2025 finding that PJM’s tariff failed to allocate costs to the entities that cause them, and by the Department of Energy’s October 2026 letter pressing PJM on ratepayer protections.[59][43][75] What is not universal is any method for determining what those costs are, and that determination cannot be made without verified demand.
Verified incremental demand helps regulators identify which grid investments are necessary because of particular industrial projects in a straightforward way: an upgrade that would not have been needed but for a specific contracted load is attributable to that load, and the attribution can be documented in the system-impact study that precedes the contract. Cost-allocation principles protecting existing ratepayers while respecting obligations to provide non-discriminatory service then follow. Directly assignable facilities, such as radial lines and dedicated substation bays, should be paid by the customer, as Texas’s proposed rules require through contributions in aid of construction.[35] Network upgrades triggered by a specific load but benefiting other customers should be allocated in proportion to the benefit, with the triggering customer bearing the share that would not have been incurred absent its load. Generation resources that serve multiple customers should be allocated through the capacity-market or integrated-resource-planning mechanism of the region, with the registry providing the demand data that determines each customer’s share. And infrastructure costs that cannot reasonably be assigned to a single developer, such as system-wide reliability upgrades, should be socialized under the traditional rule, but only after the directly attributable costs have been removed.
The quantitative stakes are large. Industrial Info Resources estimated that datacenters in PJM’s footprint would create approximately $6.3 billion in capacity costs for the 2028–2029 delivery year alone.[76] Borenstein estimated the wholesale price impact of concentrated datacenter demand in PJM at two to three cents per kilowatt-hour at retail, a ten to twenty percent increase across states, and observed that industrial customers connected at distribution voltage might bear more of the burden than residential customers, because datacenters connecting at transmission voltage do not pay into distribution costs.[31] The IMF’s working paper estimated that American electricity prices could rise 8.6 percent under its AI scenario absent policy responses, while noting that the effects would be “manageable but varying” depending on infrastructure constraints.[66] The jointly beneficial grid improvements that datacenters can fund, and that PG&E argues could lower bills for all customers by one to two percent per thousand megawatts, are real; so are the cost shifts that Peskoe and Martin documented.[55][60] Which effect dominates in a given territory is an empirical question, and the ledger is the instrument that answers it.
5.5 A Phased Implementation Roadmap: 2027–2030
The framework can be implemented in four phases, each building on the instruments that already exist. These milestones are proposed research recommendations, not existing statutory deadlines or adopted federal requirements, and they are offered as a sequence rather than a schedule.
Table 7. Proposed phased implementation roadmap for Demand Cartography, 2027–2030
| Phase | Year | Objective | Principal actions | Building on | Illustrative measure of completion |
| Standardization and regional pilots | 2027 | Common definitions and reporting templates | NERC or a joint FERC–NARUC working group publishes the four-status vocabulary; PJM registry, ERCOT Batch Zero and RFI, MISO LLWG, and CAISO initiative adopt it; secure information-sharing protocols established; cross-regional duplicate-request disclosure piloted among willing states | PJM registry (Oct. 2026); Texas SB 6 and RFI; FERC show-cause compliance filings; state executive orders | Four regions reporting in a common template |
| Independent verification and reconciliation | 2028 | Verified records and the first conversion-rate data | Assurance procedures adopted; first annual reconciliation of 2027 announced and contracted demand against 2028 energized and consumed demand; procedures for correcting inaccurate information; readiness records paired with demand records | Michigan MPSC safeguards; Ohio long-term contracts; Texas financial security; Pennsylvania annual consumption reports (from July 2027) | Published regional conversion rates by project class |
| Financial and regulatory integration | 2029 | Verified demand used in decisions | Demand attestations accepted in project finance, utility resource planning, rate proceedings, infrastructure guarantees, and public incentive decisions; state curtailment-priority designations adopted where chosen and reflected in regional tariffs | FERC’s October 2026 invitation to states; Virginia 2027 legislation; Texas curtailment statute | Attestations referenced in at least one rate case and one financing per region |
| Interoperable demand intelligence | 2030 | Comparable national picture | Regional assessments aggregated into a national view of credible industrial demand, growth rates, and unresolved electricity constraints; forecast-error benchmarks published; framework reviewed against outcomes | NERC LTRA; EIA; LBNL and EPRI bottom-up models | National assessment distinguishing the four statuses by region |
Source: Author’s synthesis of the cited regulatory filings, corporate disclosures, and literature.
The 2027 phase builds on functioning regional and state initiatives rather than assuming a national registry already exists, because none does. The 2028 phase encourages participating jurisdictions to reconcile submitted forecasts against construction milestones and early operational consumption, and to develop clear procedures for correcting inaccurate information, since a registry that cannot be corrected will accumulate errors faster than it accumulates value. The 2029 phase evaluates whether verified demand records can inform project finance, utility resource planning, infrastructure guarantees, rate proceedings, and public incentive decisions, which is the point at which the financial-statement analogy becomes operational. The 2030 phase works toward comparable regional assessments that enable policymakers and infrastructure providers to understand where credible industrial demand exists, how rapidly it is growing, and which electricity constraints remain unresolved.
5.6 Testing Demand Cartography Through Measurable Outcomes
A framework that cannot be evaluated cannot be improved, and the fifth element is a set of performance measures for determining whether better demand information actually improves infrastructure planning. The measures proposed here are deliberately empirical and deliberately comparative, so that regions adopting the framework can be compared with regions that have not.
The first measure is forecast accuracy, defined as the error between the contracted demand recorded in a given year and the consumed demand recorded two years later, by region and by project class; a declining error over time would indicate that the verification procedures are filtering speculative demand. The second is the proportion of projects reaching stated milestones, which is the conversion rate from announced to contracted, contracted to energized, and energized to consumed; a rising conversion rate would indicate that the costs of entering the registry are deterring projects that would not have proceeded. The third is the difference between planned and actual peak demand at energized facilities, which tests whether ramp schedules are being reported honestly. The fourth is infrastructure delivery delay, measured as the gap between the readiness date recorded at contract and the actual energization date; a narrowing gap would indicate that the readiness record is improving coordination. The fifth is the cost assigned to existing ratepayers from large-load-driven upgrades, measured through rate proceedings; a declining share would indicate that cost allocation is working.
These measures would allow regulators to assess whether improved verification reduces unnecessary construction, strengthens investor confidence, accelerates legitimate projects, and improves reliability outcomes. The hypothesis of this paper is that it would do all four, but the hypothesis is testable, and the framework should be designed so that it can be rejected if the evidence does not support it. A particular test worth specifying in advance concerns the Texas audit: if ERCOT’s December 10 report and the subsequent Batch Zero classifications produce a credible queue that is a fraction of the 474 gigawatts requested, and if the projects that clear the audit energize on schedule at a higher rate than the historical conversion rate, then the verification approach will have demonstrated its value in the largest natural experiment the country has yet run.
5.7 The Political Economy of Verified Electricity Demand
No framework for electricity governance can be evaluated without attention to who gains and who loses from it, and the political economy of verified demand is more complex than its technical logic. For governors, verification offers a way to demonstrate to voters that datacenter growth is being managed rather than merely welcomed, and the actions of Shapiro, Spanberger, Whitmer, and Abbott in 2026 suggest that governors of both parties have concluded that the political cost of appearing to subsidize speculative demand now exceeds the political benefit of announcing investment. For Congress, the White House’s March 2026 policy framework called on legislators to “ensure that residential ratepayers do not experience increased electricity costs as a result of new AI data center construction and operation” and to codify the Ratepayer Protection Pledge, and Senator Hawley among others endorsed codification, which suggests that verification and cost allocation may be among the few AI policy questions on which bipartisan legislation is possible.[59] For FERC, the October 9 order has defined the Commission’s role as the guarantor of information and the referee of non-discrimination while deferring curtailment priority to the states, a division of labor that the Commission appears to find comfortable. For state utility commissions, verification is both an opportunity and a burden: it gives them the data they need to allocate costs, and it asks them to make the distributional decisions about curtailment priority that FERC has declined to make.
For local governments, verification promises relief from the pattern in which communities learn the scale of a project only after the incentives have been granted, which is the pattern that both Pennsylvania’s and Virginia’s nondisclosure bans target. For financial institutions, verification offers a basis for underwriting that the current system does not provide, and the Jefferies observation that FERC’s order favors incumbent generators illustrates how closely capital markets are now reading regulatory signals about demand credibility.[2] For AI developers, verification is a mixed blessing: it imposes costs and disclosure obligations, but it also promises that credible projects will no longer wait behind speculative ones, and the largest developers, having signed the Ratepayer Protection Pledge and Michigan’s pledge and having accepted Ohio’s contract terms, appear to have concluded that a verified queue is in their interest.
The November 2026 elections and the 2027 legislative sessions that follow will accelerate these debates, because electricity affordability has become a campaign issue in Pennsylvania, Virginia, Michigan, Texas, and beyond, and because the governors who have acted by executive order will need legislatures to make their actions durable. Pennsylvania’s Republican gubernatorial candidate called for a pause on datacenter approvals; Michigan’s environmental coalition supports moratorium bills; New York enacted the first statewide temporary moratorium; and more than a dozen states have considered moratoria.[11][16][77] The tension between attracting AI investment and demonstrating to voters that households, small businesses, and established industries will not unfairly subsidize speculative electricity demand is the defining tension of state energy politics in 2026, and verification is the only approach that resolves it without choosing one side.
The section thesis follows. A national Demand Cartography framework should make electricity infrastructure commitments more evidence-based, financially accountable, and operationally reliable without creating unnecessary barriers to productive industrial growth. The barriers it creates are barriers to speculation, and speculation, in a system where the public underwrites the infrastructure, is a cost that someone else pays.
Section 6: What Have We Learned? Seven Pillars
The analysis developed across the preceding five sections can be distilled into seven lessons, each paired with a core principle. These are not restatements of the section theses but syntheses across them, and each draws on the events of 2026 as evidence rather than illustration.
Pillar 1 — America Must Map Electricity Demand Before It Can Govern Industrial Expansion
The first lesson is that infrastructure planning cannot depend exclusively on aggregate electricity forecasts or on corporate development announcements, because modern artificial intelligence datacenters have become large enough to influence regional electricity prices, generation decisions, transmission investment, and the economics of every other industry that shares the grid with them. A single campus of a gigawatt is no longer a customer; it is a planning scenario. When ERCOT’s queue reaches five times the state’s historical peak and PJM’s capacity auction clears at the cap for the third consecutive year with a 6.8-gigawatt shortfall, the aggregate forecast has ceased to be informative, and the only path back to informative planning runs through the enumeration of individual projects.[9][6]
A reliable electricity map must therefore describe not only where power is requested but also how credible the request is, when demand is expected to materialize, and whether sufficient infrastructure exists to deliver electricity safely. PJM’s registry, Pennsylvania’s DEP map, Texas’s RFI, MISO’s operational forecasting proposal, and CAISO’s straw proposal are each partial maps of this kind, and their proliferation in a single year is itself evidence that the need is recognized. The core principle is that better electricity governance begins with verified information, and that no amount of governance can compensate for information that is not verified.
Pillar 2 — Announced Megawatts Are Not the Same as Consumed Electricity
The second lesson is that electricity-demand reporting must distinguish commercial intentions from observable industrial activity, because announced projects, binding contracts, energized facilities, and metered consumption represent different forms of evidence with different probabilities of becoming real. Treating them as equivalent creates risks for utilities, which build to loads that do not arrive; for investors, who finance infrastructure whose revenue depends on those loads; for customers, who pay for stranded assets; and for governments, which grant incentives against capacity that is never consumed. Wood Mackenzie’s estimate that seventy-two percent of requested datacenter load is phantom, Oxford Economics’ estimate of six in seven, and EPRI’s finding that no surveyed utility expected more than thirty-five percent of its requests to materialize within five years are three independent measurements of the same gap, and the gap is the single most important fact about American electricity planning in 2026.[32][33][34]
Future electricity planning should preserve these distinctions and regularly compare prior projections with physical and commercial outcomes, so that the conversion rate from announcement to consumption becomes a measured quantity rather than a guess. The core principle is that infrastructure commitments should reflect the maturity and credibility of the underlying demand, and that a planning system which cannot distinguish a press release from a meter reading will be systematically wrong in the direction of overbuilding.
Pillar 3 — Transparency Does Not Automatically Create Regulatory Authority
The third lesson emerges directly from the October 2026 FERC decision. Grid operators may require more accurate information about major industrial customers without automatically gaining the authority to interrupt service under newly proposed arrangements. The registry was accepted; the curtailment service was rejected; and the Commission explained that the difference lay in the division of jurisdiction between federal regulation of wholesale transmission and state regulation of retail service.[2] The power to collect information, to impose costs, to determine electricity-service priorities, and to authorize emergency curtailment must each rest on applicable legal authority and clearly established procedures, and those authorities are held by different institutions.
This lesson is uncomfortable for those who hoped that a single regulatory action could resolve the large-load problem, but it is also clarifying, because it specifies what must happen next. The states within PJM must designate curtailment priorities through their own processes, as Texas has already done by statute; PJM must then implement those designations through tariff provisions that return to FERC for review; and the resulting framework must attach to the characteristics of loads rather than to the identity of customers. The core principle is that demand visibility must be accompanied by lawful, transparent, and non-discriminatory governance, and that the map, however accurate, does not confer the authority to act upon it.
Pillar 4 — Regional Diversity Requires Comparable Information, Not Identical Electricity Markets
The fourth lesson is that Pennsylvania, Virginia, Michigan, Texas, and California face different economic circumstances, energy resources, regulatory systems, and infrastructure requirements, and that PJM, ERCOT, MISO, and CAISO operate under different market designs and governance arrangements that no federal action is likely to homogenize. The Department of Energy’s attempt to secure a uniform federal large-load rule by April 2026 produced instead six tailored show-cause proceedings, and five of the six grid operators sought delay.[74][37][32] The lesson is not that uniformity is undesirable but that it is unavailable, and that a framework which waits for it will wait too long.
A credible national framework should therefore enable meaningful comparisons while respecting those differences. Comparability is a property of data definitions, and it is achievable through the common vocabulary proposed in Section 5; uniformity is a property of law, and it is neither achievable nor, given the real differences among regions in resource mix, weather exposure, and retail-market structure, desirable. The core principle is that standardized reporting can improve national coordination without eliminating legitimate regional regulatory autonomy, and that the right model is accounting standards rather than a single national utility.
Pillar 5 — Verified Electricity Demand Could Become an Essential Financial Asset
The fifth lesson is that independently verified electricity-demand information may become increasingly important to the financing of artificial intelligence infrastructure, because banks, institutional investors, utilities, insurers, and public authorities all depend on assumptions about when industrial electricity demand will emerge and how much infrastructure will be required, and because those assumptions are currently unsupported by any standardized evidence. The $730 billion of 2026 hyperscaler capital expenditure, the $89 billion of quarterly NVIDIA datacenter revenue, and the $105 billion of supply and capacity obligations NVIDIA has disclosed are all claims on electricity whose realization depends on infrastructure that someone must finance, and the financing is currently being done against demand that no one has attested.[17][19]
Credible demand attestations could reduce uncertainty and improve project evaluation, even though they would never eliminate market, construction, technology, or regulatory risk, in exactly the way that audited financial statements reduce uncertainty about a company’s position without eliminating uncertainty about its future. The Ohio twelve-year contract requirement, the Michigan collateral rules, and the Texas financial-security provisions are the first instruments of this kind, and the willingness of the largest developers to accept them suggests that the market is already pricing verified commitment above unverified ambition.[31][15][35] The core principle is that in the Five-Layer AI Economy, the quality of verified demand information may become nearly as important as the quantity of electricity available, because capital flows to what can be verified.
Pillar 6 — Flexibility Is a Property That Must Be Measured, Not Assumed
The sixth lesson, which extends beyond the original scope of this inquiry, is that the flexibility of datacenter load, on which so much of the optimistic analysis depends, is a property that must be measured at the meter and verified in operation rather than assumed from the nature of the workload. Duke’s finding that seventy-six to 126 gigawatts of new load could be integrated at curtailment rates between 0.25 and one percent, MIT’s finding that temporal flexibility could lower system costs, and NERC’s observation that committed flexibility could accelerate interconnection all depend on the word “committed,” and commitment is observable only in performance records.[63][29][24] The July 2026 event in which 3.8 gigawatts tripped offline in Northern Virginia demonstrated that datacenters are already flexible in the sense that matters least, which is that they can disappear from the grid instantaneously and involuntarily, while the flexibility that matters most, the ability to reduce consumption on request during the hours of greatest stress, has yet to be demonstrated at scale by most operators.[51]
The four-status ledger should therefore record not only how much a facility consumed but how it performed when asked to reduce, and the value of flexibility in interconnection and cost allocation should be tied to that performance record. Facilities that have demonstrated curtailment capability should receive the faster interconnection and the favorable cost treatment that the Duke, MIT, and NERC analyses justify; facilities that have merely asserted it should not. The core principle is that flexibility earns its benefits through verified performance, and that an unverified promise of flexibility is simply another form of announced demand.
Pillar 7 — The Energy Layer Cannot Be Planned Without Visibility Into the Layers Above It
The seventh lesson returns to the Five-Layer AI Economy. The electricity a datacenter consumes is determined by decisions made at the semiconductor layer, where each accelerator generation roughly doubles rack-level power; at the datacenter layer, where cooling and power architecture set the overhead; at the model layer, where training and inference workloads set the utilization pattern; and at the application layer, where user and agent demand sets the volume. The grid sees only the aggregate result, and it sees it only after the fact. The IEA’s observation that datacenter demand surged seventeen percent in 2025, the LBNL update raising its 2030 reference case to nearly twelve percent of national electricity, and Huang’s description of “multiple frontier labs scaling in parallel” are all reports from the upper layers that arrived at the energy layer as surprises.[39][41][19]
A Demand Cartography framework cannot give the grid visibility into model architectures or customer workloads, and it should not try, for the confidentiality reasons developed in Section 4. What it can do is require that the upper layers translate their plans into the vocabulary of the energy layer, phase by phase, with evidence, so that a change in model efficiency or chip density shows up in the ledger as a revised contracted or energized quantity rather than as an unexplained deviation at the meter. The core principle is that the energy layer can be planned only if the layers above it report their intentions in verifiable form, and that the Five-Layer AI Economy is, in the end, a single system whose foundation cannot be built without knowledge of what will stand on it.
Conclusion: Demand Cartography and the Next Geography of American Artificial Intelligence
The October 2026 regulatory decision involving PJM Interconnection offers an important glimpse into the future of America’s artificial intelligence economy. By accepting a large-load registry while rejecting PJM’s proposed curtailment service, the Federal Energy Regulatory Commission demonstrated that the ability to identify industrial electricity demand and the authority to control that demand are separate dimensions of energy governance, held by separate institutions and governed by separate bodies of law. That distinction will become increasingly consequential between 2027 and 2030, as the states within PJM decide whether to designate curtailment priorities, as Texas completes the largest audit of industrial electricity demand ever attempted, as MISO and CAISO convert their proposals into tariffs, and as the hyperscalers convert their capital-expenditure guidance into physical facilities whose consumption the grid will finally be able to measure.
America’s hyperscalers, semiconductor manufacturers, frontier-model developers, datacenter operators, utilities, and industrial investors are building an increasingly interconnected economic system. NVIDIA, AMD, Amazon, Microsoft, Google, Meta, OpenAI, Anthropic, Oracle, xAI, and many other enterprises ultimately depend on the availability of reliable electricity, and their investments and technological innovations will influence where electricity is consumed, how rapidly demand increases, and how urgently new infrastructure must be constructed. Seven of them signed a pledge at the White House to build, bring, or buy their own power. Six of them signed Michigan’s pledge within a day. One of them accepted Ohio’s twelve-year contract terms before any other. These are commitments, and they matter. Yet even the largest financial commitments cannot eliminate the need to distinguish proposed capacity from genuine physical demand.
A datacenter announcement is not an operating facility. A connection request is not a binding electricity commitment. A signed agreement does not guarantee completed infrastructure. An energized campus does not necessarily consume its maximum rated capacity. And a forecast of future consumption, however sophisticated and however well-funded, remains a forecast until actual operating evidence becomes available. These distinctions are fundamental to responsible industrial planning, and they are equally fundamental to public accountability, because every error in the direction of overbuilding is paid for by someone who was never asked.
Governors such as Josh Shapiro, Abigail Spanberger, Gretchen Whitmer, and Greg Abbott confront a difficult balance. They must support investment, technological innovation, economic development, and national competitiveness while responding to public concerns about electricity affordability, environmental impacts, infrastructure reliability, and the distribution of costs. Their instruments differ, from Pennsylvania’s permitting leverage and public map to Virginia’s five-pillar framework and legislative agenda, from Michigan’s voluntary pledge and call for codification to Texas’s statutory curtailment authority and executive audit, but their diagnosis is shared. The solution cannot simply be to approve every proposed project or to reject every new datacenter. Instead, governments need stronger evidence about what developers intend to build, what they are contractually obligated to deliver, what infrastructure can realistically support, and what electricity the resulting facilities actually consume.
Within the Five-Layer AI Economy, this requirement begins at the energy layer but influences the entire industrial system. The electricity needed to support datacenters is ultimately determined by the hardware they deploy, the models they operate, and the applications and autonomous agents that create demand for computation. Better demand verification at the energy layer can therefore improve investment decisions throughout the technological layers above it, because a semiconductor roadmap, a datacenter construction schedule, and a model-training plan that cannot be reconciled with deliverable electricity are plans that will not be executed on time, whatever their technical merit.
This is why the title Demand Cartography is appropriate. Cartography is not merely the act of drawing a map. It is the disciplined process of representing geographic reality through measurable information, consistent definitions, and verifiable observations. A reliable map distinguishes established roads from proposed routes, completed infrastructure from projected construction, and confirmed features from uncertain assumptions, and it does so through conventions that every reader understands and that every mapmaker is obliged to follow. The surveyor who records a proposed road as built is not merely mistaken; the surveyor has violated the discipline that gives maps their authority.
America’s electricity system now requires a comparable discipline. Its map must reveal where the next generation of AI infrastructure is likely to emerge, how much electricity is credibly required, when those requirements will materialize, which projects can be supported, and how the resulting financial and operational risks should be allocated. Such a map would not decide every electricity-policy dispute. It would not replace democratic accountability, state regulatory authority, federal oversight, engineering judgment, or market competition, and it would not tell a state commission whether to curtail a datacenter before a hospital, because that is a question of values rather than of measurement. But it could make all of those decisions substantially better informed, and in a system where the cost of uninformed decisions is measured in billions of dollars of stranded infrastructure and in percentage points of household electricity bills, better information is not a modest contribution.
For private corporations, Demand Cartography offers a framework for evaluating where infrastructure investment can succeed, and for distinguishing their own credible projects from the speculative requests of competitors that currently crowd the same queues. For financial institutions, it offers a basis for distinguishing substantiated industrial demand from speculative capacity, in the form of attestations whose structure they already understand. For state governments, it offers a foundation for attracting productive investment while protecting communities and ratepayers, and for converting executive orders into durable law. For federal policymakers, it offers a way to improve the quality and comparability of infrastructure information without assuming that every region requires identical regulation, which is the only approach the events of 2026 suggest is politically and legally available. And for the Five-Layer AI Economy, it offers a missing analytical connection between the physical geography of electricity and the emerging geography of machine intelligence.
The defining question of the next phase of American artificial intelligence expansion may no longer be simply how much electricity the nation can produce. It may be how accurately the nation can establish where electricity is actually needed, when demand will become real, who should bear the associated costs, and which legal institutions possess the authority to act on that information. That is the purpose of Demand Cartography. Before America can responsibly allocate the electricity that powers its artificial intelligence future, it must learn to map that future with evidence rather than expectation.
Footnotes and Endnotes:
[1] Federal Energy Regulatory Commission. Order Accepting in Part and Rejecting in Part Tariff Revisions, PJM Interconnection, L.L.C., Docket No. ER26-3515-000, October 9, 2026. https://elibrary.ferc.gov/eLibrary/filelist?accession_number=20261009-3055&optimized=false
[2] Ethan Howland, Utility Dive. “FERC rejects most of PJM’s large load interconnection plan,” October 12, 2026 (including statements of FERC, Commissioner David Rosner, PJM spokesman Jeffrey Shields, and Jefferies analysts). https://www.utilitydive.com/news/ferc-pjm-large-load-interconnection-iras-byonc/832716/
[3] Adam Aton, E&E News (Politico). “PJM proposes data center registry with teeth,” July 28, 2026. https://www.eenews.net/articles/pjm-proposes-data-center-registry-with-teeth/
[4] Data Center Knowledge. “PJM’s New Deal for Data Centers: Bring Power or Face Cuts,” August 2026. https://datacenterknowledge.com/energy-power-supply/pjm-s-new-deal-for-data-centers-bring-power-or-face-cuts
[5] Jersey Vindicator. “Power grid operator PJM launches ‘fast-track plan’ to confront surge in data center electricity demand” (PJM Board letter; 2025 long-term forecast), 2025. https://jerseyvindicator.org/?p=12708
[6] PJM Interconnection. “PJM Capacity Auction Procures 138,318 MW of Generation Resources as Work Continues To Address Supply Shortfall,” news release, July 14, 2026. https://www.pjm.com/-/media/DotCom/about-pjm/newsroom/2026-releases/20260714-pjm-capacity-auction-procures-138318-mw-of-generation-resources.pdf
[7] Modo Energy. “PJM’s 2028/29 capacity auction: prices remain at the cap as the crunch continues,” July 15, 2026. https://modoenergy.com/research/en/pjm-capacity-auction-2028-2029-reliability-price-cap-elcc
[8] Pennsylvania Capital-Star. “PJM Interconnection electricity price hits cap again in latest auction” (statement of David Mills, PJM), July 14–15, 2026. https://penncapital-star.com/?p=67929
[9] Office of the Texas Governor. “Governor Abbott Directs Comprehensive Data Center Audit,” press release, August 3, 2026. https://gov.texas.gov/news/post/governor-abbott-directs-comprehensive-data-center-audit
[10] Electric Reliability Council of Texas. “ERCOT Request for Information Regarding Data Center State and Community Impacts,” September–October 2026. https://www.ercot.com/about/legal/data-center-impact-rfi
[11] Office of Governor Josh Shapiro, Commonwealth of Pennsylvania. “Governor Shapiro Signs Executive Order Demanding Data Center Developers Comply with Strict Requirements and Blocking Speculative, Irresponsible Data Center Projects,” August 18, 2026. https://www.pa.gov/governor/newsroom/2026-press-releases/governor-shapiro-signs-executive-order-on-data-center-developmen
[12] Reed Smith LLP. “Pennsylvania governor signs executive order imposing new requirements on data center development,” Client Alert 2026-169, August 2026. https://www.reedsmith.com/articles/pennsylvania-governor-signs-executive-order-imposing-new-requirements-on-data-center-development/
[13] Charlie Paullin, Virginia Mercury. “Spanberger presents slate of data center regulation proposals, puts first directives in motion,” September 18, 2026. https://virginiamercury.com/2026/09/18/spanberger-presents-slate-of-data-center-regulation-proposals-puts-first-directives-in-motion/
[14] Greenberg Traurig LLP. “Virginia Executive Order 22 Sets New Requirements for Data Center Development and AI Governance,” September 2026. https://www.gtlaw.com/ar/insights/2026/9/virginia-executive-order-22-sets-new-requirements-for-data-center-development-and-ai-governance
[15] Office of Governor Gretchen Whitmer, State of Michigan. “Gov. Whitmer Secures Six Company Commitments to Michigan Affordability and Responsible Growth Pledge,” July 16, 2026. https://www.michigan.gov/whitmer/news/press-releases/2026/07/16/whitmer-secures-six-company-commitments-to-michigan-affordability-and-responsible-growth-pledge
[16] Michigan Advance. “Whitmer rolls out data center plan, calls on lawmakers to pass laws protecting energy customers,” July 15, 2026. https://michiganadvance.com/2026/07/15/whitmer-rolls-out-data-center-plan-calls-on-lawmakers-to-pass-laws-protecting-energy-customers/
[17] CNBC. “Amazon, Meta and Microsoft face skeptical investors this week after Google report sparked sell-off,” July 28, 2026. https://www.cnbc.com/2026/07/28/hyperscalers-face-higher-capex-scrutiny-after-alphabet-report-panned.html
[18] EffectStory. “Microsoft, Alphabet, Meta, Amazon: Q2 2026 AI Capex Actuals and Raised Full-Year Guidance,” compiled from company earnings releases, August 2026. https://effectstory.com/en/articles/hyperscaler-ai-capex-q2-2026-microsoft-alphabet-meta-amazon
[19] NVIDIA Corporation. “NVIDIA Announces Financial Results for Second Quarter Fiscal 2027,” August 26, 2026 (statement of Jensen Huang). https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Announces-Financial-Results-for-Second-Quarter-Fiscal-2027/default.aspx
[20] Zero Emission Grid. “MISO LLWG 05/14 Meeting Summary: Large Load Interconnection Reliability Requirements, ZGIA, Reliable Operations, and Reserve Products,” May 2026; and MISO, “Load Forecast Development and Use at MISO.” https://www.zeroemissiongrid.com/iso-rto/miso-llwg-05-14/
[21] Global Data Center Hub. “ERCOT data center interconnection pause: PUCT Docket 59220 and the Batch Zero timeline,” August 2026. https://www.globaldatacenterhub.com/p/ercot-data-center-interconnection-pause-docket-59220
[22] Dave Friedman. “Inside Texas’s AI Data Center Queue: How real is the 226 gigawatts of demand sitting on Texas’ interconnection queue?” January 20, 2026. https://davefriedman.substack.com/p/inside-texass-ai-data-center-queue
[23] POWER Magazine. “NERC Warns Long-Term Grid Reliability Risks Mounting from Surging Demand, Lagging Resources” (2025 Long-Term Reliability Assessment), January 2026. https://www.powermag.com/nerc-warns-long-term-grid-reliability-risks-mounting/
[24] Power Engineering. “NERC warns reliability risk is rising as load growth outpaces infrastructure” (2025 Long-Term Reliability Assessment; statements of John Moura and Mark Olson), January 2026. https://www.power-eng.com/business/policy-and-regulation/nerc-warns-reliability-risk-is-rising-as-load-growth-outpaces-infrastructure/
[25] Broadband Breakfast. “Experts Debate Whether AI Energy Demand Is a New Crisis or History Repeating” (Frank Wolak, Stanford University, at the Technology Policy Institute Aspen Forum), August 18, 2025. https://broadbandbreakfast.com/experts-debate-whether-ai-energy-demand-is-a-new-crisis-or-history-repeating
[26] Bloomberg, via Data Center Knowledge. “AI’s Power Needs Not as Bad as Feared, Princeton Professor Says” (Jesse Jenkins, Princeton University), April 2024. https://www.datacenterknowledge.com/energy/ai-s-power-needs-not-bad-feared-princeton-professor-says
[27] Jonathan Koomey et al., Bipartisan Policy Center. “Electricity Demand Growth and Data Centers: A Guide for the Perplexed,” 2025. https://bipartisanpolicy.org/report/electricity-demand-growth-and-data-centers/
[28] PJM Interconnection. “Large Load — Registry and Resource Adequacy Initiatives,” PJM Markets and Operations. https://www.pjm.com/markets-and-operations/large-load
[29] Christopher R. Knittel, Juan Ramon L. Senga, and Shen Wang, MIT Center for Energy and Environmental Policy Research. “Flexible Data Centers and the Grid: Lower Costs, Higher Emissions?” MIT CEEPR Working Paper 2025-14, July 2025; and Utility Dive coverage with Knittel’s statements. https://www.utilitydive.com/news/data-center-flexibility-can-save-consumers-money-but-may-come-with-higher-emissions-MIT/756266
[30] arXiv preprint 2604.10769. “Workload composition smooths aggregate power demand while sustaining short-horizon ramps in AI data centers,” April 2026. https://arxiv.org/pdf/2604.10769
[31] Jericho Casper, Broadband Breakfast. “Economist Warns Data Center Interconnection Requests May Be Inflated” (Severin Borenstein, University of California, Berkeley, at the Technology Policy Institute Aspen Forum), August 20, 2026. https://broadbandbreakfast.com/economist-warns-data-center-interconnection-requests-may-be-inflated/
[32] Capacity Media. “Wood Mackenzie: US grid operators set to serve 28% of data centre requests,” August 25, 2026. https://capacityglobal.com/news/wood-mackenzie-us-grid-operators-set-to-serve-28pc-of-data/
[33] Oxford Economics. “Estimating Data Centre ‘Phantom Demand,’” research blog, 2026. https://www.oxfordeconomics.com/?p=124495
[34] Brian Martucci, Utility Dive. “A fraction of proposed data centers will get built. Utilities just don’t know which ones” (EPRI utility survey; Clements and Freed), 2025. https://www.utilitydive.com/news/a-fraction-of-proposed-data-centers-will-get-built/748214/
[35] Texas Legislature, 89th Session. Senate Bill 6 (Enrolled), relating to the planning for, interconnection and operation of, and costs related to providing service for certain electrical loads, effective June 20, 2025; and Foley & Lardner LLP, “Public Utility Commission of Texas Issues Proposed Rules for Large Load Interconnections.” https://legiscan.com/TX/text/SB6/2025
[36] Brian Martucci, Utility Dive. “Texas law gives grid operator power to disconnect data centers during crises,” June 2025 (including statement of Lee Bratcher, Texas Blockchain Council). https://www.utilitydive.com/news/texas-law-gives-grid-operator-power-todisconnect-data-centers-during-crisi/751587/
[37] Foley & Lardner LLP. “FERC Issues Six Tailored ‘Show Cause’ Orders to Accelerate Large Load Interconnection,” June 30, 2026. https://www.foley.com/insights/publications/2026/06/ferc-issues-six-tailored-show-cause-orders-to-accelerate-large-load-interconnection/
[38] International Energy Agency. “Energy and AI,” World Energy Outlook Special Report, April 2025 (statement of Executive Director Fatih Birol). https://www.iea.org/reports/energy-and-ai
[39] W.Media. “Data center power demand will double by 2030: IEA” (IEA Key Questions on Energy and AI report; statement of Fatih Birol), April 20, 2026. https://w.media/data-center-power-demand-will-double-by-2030-iea/
[40] Arman Shehabi, Sarah J. Smith, Alex Hubbard, Alexander Newkirk, Nuoa Lei, Md AbuBakar Siddik, Billie Holecek, Jonathan G. Koomey, Eric R. Masanet, and Dale A. Sartor, Lawrence Berkeley National Laboratory. “2024 United States Data Center Energy Usage Report,” December 19, 2024. https://energyanalysis.lbl.gov/publications/2024-lbnl-data-center-energy-usage-report
[41] Sarah J. Smith, Alex Hubbard, Alexander Newkirk, Mohan Ganeshalingam, Billie Holecek, Dale A. Sartor, Michael Mills, and Arman Shehabi, Lawrence Berkeley National Laboratory. “United States Data Center Energy Usage Report: 2025 Update,” June 2026. https://escholarship.org/uc/item/33m6w3x0
[42] Thomas F. Wilson, Geoffrey J. Blanford, John Bistline, and Nils Johnson, Electric Power Research Institute. “Powering Intelligence 2026: Updated Scenarios of U.S. Data Center Electricity Use and Power Strategies,” 2026. https://esca.epri.com/products/3002034696/
[43] Mondaq (Day Pitney). “FERC Directs PJM To Facilitate Faster Integration Of Large Electric Loads, Including Data Centers” (193 FERC ¶ 61,217, December 18, 2025), January 13, 2026. https://www.mondaq.com/unitedstates/energy-law/1730084/ferc-directs-pjm-to-facilitate-faster-integration-of-large-electric-loads-including-data-centers
[44] RTO Insider. “FERC Denies Monitor Complaint to Establish PJM Authority to Reject Large Loads” (Docket EL26-30), April 5, 2026. https://www.rtoinsider.com/128871-ferc-denies-imm-complaint-establish-pjm-authority-reject-large-loads/
[45] Rao Konidena. “FERC’s rejection of PJM’s IRAS proposal requires a second glance,” Substack, October 12, 2026. https://raokonidena.substack.com/p/fercs-rejection-of-pjms-iras-proposal
[46] Tory Lauterbach and Adam Whitehouse, The Texas Lawbook. “What Gov. Abbott’s Data Center Audit Directive Means for ERCOT and the Batch Zero Study Process,” August 25, 2026. https://texaslawbook.net/what-gov-abbotts-data-center-audit-directive-means-for-ercot-and-the-batch-zero-study-process/
[47] Bracewell LLP (Bryan Clark and Jared Berg, quoted in Houston Business Journal). “ERCOT: Gov. Abbott’s Audit Pauses Texas Data Center Development for Months,” August 12, 2026. https://www.bracewell.com/news-events/ercot-gov-abbotts-audit-pauses-texas-data-center-development-for-months/
[48] Midcontinent Independent System Operator. “LLWG: Large Load Operational Forecast Requirements (LLWG-2026-4a),” stakeholder feedback request, 2026. https://www.misoenergy.org/engage/stakeholder-feedback/2026/llwg-large-load-operational-forecast-requirements-llwg-2026-4a/
[49] Akin Gump Strauss Hauer & Feld LLP. “FERC Directs Midcontinent Independent System Operator to Revise Tariff to Expedite Data Center Development,” Speaking Energy, June 2026. https://www.akingump.com/en/insights/blogs/speaking-energy/ferc-directs-midcontinent-independent-system-operator-to-revise-tariff-to-expedite-data-center-development
[50] Midcontinent Independent System Operator. “Missouri Public Service Commission Quarterly Update,” August 19, 2026. https://psc.mo.gov/CMSInternetData/Agenda%20Presentations/2026%20Presentations/8-19-26%20MISO%20Status%20Update.pdf
[51] Ethan Howland, Utility Dive. “PJM eyes data center, crypto reliability requirements after 3.8 GW of load trips offline,” August 12, 2026. https://www.utilitydive.com/news/pjm-nerc-data-center-crypto-reliability-standards/827653/
[52] North American Electric Reliability Corporation (Serna). “Large Loads Action Plan,” presentation, June 10, 2026. https://www.rer.org/wp-content/uploads/6_NERC_Large_Loads_Serna.pdf
[53] California Independent System Operator. “Large Loads Initiative: Straw proposal posted, meeting on 8/19/26,” market notice, August 12, 2026; and “Large Loads Initiative: Technical requirements straw proposal posted,” June 15, 2026. https://www.caiso.com/notices/large-loads-initiative-straw-proposal-posted-meeting-on-8-19-26
[54] Zero Emission Grid. “CAISO Large Load Interconnection Update: Data Center Growth, Planning Impacts & Technical Requirements” (CAISO Large Load Information Session, February 5, 2026). https://www.zeroemissiongrid.com/iso-rto/caiso-large-load/
[55] Redwood City Pulse. “‘We need more energy’: Stanford symposium explores data center growth” (Stanford Sustainable Data Centers Symposium; EPRI; PG&E), May 4, 2026. https://www.rwcpulse.com/stanford-university/2026/05/04/we-need-more-energy-stanford-symposium-explores-data-center-growth/
[56] California Independent System Operator. “The California ISO’s role in large loads,” Large Load webpage; and “Large Loads Interconnection and Energization: Issue Paper,” January 2026. https://www.caiso.com/generation-transmission/load/large-load
[57] California Independent System Operator. “Large Loads Considerations: Straw Proposal,” August 12, 2026. https://stakeholdercenter.caiso.com/InitiativeDocuments/Large-loads-Straw-Proposal-Aug-11-2026.pdf
[58] Washington Examiner. “Big Tech firms sign Trump pledge to have data centers pay for their own power use” (statement of Ruth Porat, Google), March 4, 2026. https://www.washingtonexaminer.com/policy/energy-and-environment/4480533/big-tech-firms-trump-pledge-data-centers-power-use/
[59] American Nuclear Society, Nuclear News. “The Ratepayer Protection Pledge: Who’s signed on?” July 29, 2026 (text of the White House pledge; signatory states and utilities). https://www.ans.org/news/2026-07-29/article-8255/the-ratepayer-protection-pledge-whos-signed-on/
[60] Eliza Martin and Ari Peskoe, Harvard Law School Electricity Law Initiative. “Extracting Profits from the Public: How Utility Ratepayers Are Paying for Big Tech’s Power,” March 2025. https://eelp.law.harvard.edu/wp-content/uploads/2025/03/Harvard-ELI-Extracting-Profits-from-the-Public.pdf
[61] Matthew Zeitlin, Heatmap News. “Who Pays for a Data Center?” (interview with Ari Peskoe), March 2025. https://heatmap.news/energy/data-centers-utility-infrastructure
[62] MIT Sloan School of Management. “How data centers can better manage energy use,” press release, June 26, 2026 (Christopher Knittel, Juan Ramon L. Senga, Shen Wang). https://mitsloan.mit.edu/press/how-data-centers-can-better-manage-energy-use
[63] Tyler H. Norris, Tim Profeta, Dalia Patiño-Echeverri, and Adam Cowie-Haskell, Nicholas Institute for Energy, Environment & Sustainability, Duke University. “Rethinking Load Growth: Assessing the Potential for Integration of Large Flexible Loads in US Power Systems,” February 2025. https://nicholasinstitute.duke.edu/publications/rethinking-load-growth
[64] Nicholas Institute for Energy, Environment & Sustainability, Duke University. News coverage of “Rethinking Load Growth,” including Tyler Norris’s statement to T&D World. https://nicholasinstitute.duke.edu/news/issue/9935
[65] International Monetary Fund (Ganchimeg Ganpurev and Andrea Pescatori). “AI Needs More Abundant Power Supplies to Keep Driving Economic Growth,” IMF Blog, May 13, 2025. https://www.imf.org/en/Blogs/Articles/2025/05/13/ai-needs-more-abundant-power-supplies-to-keep-driving-economic-growth
[66] International Monetary Fund. “Power Hungry: How AI Will Drive Energy Demand,” IMF Working Paper No. 2025/081, April 21, 2025. https://www.imf.org/en/Publications/WP/Issues/2025/04/21/Power-Hungry-How-AI-Will-Drive-Energy-Demand-566304
[67] K&L Gates LLP. “Pennsylvania Data Center Executive Order Adds New Development and Permitting Requirements,” August 19, 2026. https://www.klgates.com/thought-leadership/Pennsylvania-Data-Center-Executive-Order-Adds-New-Development-and-Permitting-Requirements-8-19-2026
[68] Pennsylvania State Association of Boroughs. “Governor Shapiro Signs Data Center Executive Order” (statement of Governor Josh Shapiro), August 18, 2026. https://boroughs.org/fullnotice.php?id=1035
[69] WDBJ7. “Governor Spanberger signs executive order on data center accountability, creates AI task force” (statements of Governor Abigail Spanberger), September 18, 2026. https://www.wdbj7.com/2026/09/18/governor-spanberger-holds-news-conference-data-centers-ai/
[70] WWBT 12 On Your Side. “Spanberger signs executive order taking on data centers, AI” (text of Executive Order 22), September 18, 2026. https://www.12onyourside.com/2026/09/18/watch-live-spanberger-address-data-centers-ai/
[71] Diana DiGangi, Utility Dive. “Dominion-NextEra merger hearing focused on rates, data centers,” October 12, 2026. https://www.utilitydive.com/news/local-hearing-dominion-nextera-merger-rates-data-centers/832739/
[72] WNEM. “Gov. Whitmer unveils Michigan Affordable and Responsible Growth Action Plan, urges data centers to sign pledge” (statement of Casey Katims, U.S. Climate Alliance), July 15, 2026. https://www.wnem.com/2026/07/15/gov-whitmer-unveils-michigan-affordable-responsible-growth-action-plan-urges-data-centers-sign-pledge/
[73] Energy Central. “NEWS: PJM’s plan to manage new large loads isn’t cutting it,” October 2026. https://www.energycentral.com/energy-management/post/news-pjm-s-plan-to-manage-new-large-loads-isn-t-cutting-it-GqQ6bEL9j8dlF6E
[74] Kirkland & Ellis LLP. “Potential Game-Changing Rulemaking for Data Centers and Other High-Demand Loads,” Kirkland Alert, November 2025 (DOE Section 403 directive and ANOPR). https://www.kirkland.com/publications/kirkland-alert/2025/11/potential-game-changing-rulemaking-for-data-centers
[75] Ethan Howland, Utility Dive. “DOE presses PJM on ratepayer protections from large load costs,” October 8, 2026. https://www.utilitydive.com/news/doe-pjm-ferc-ratepayer-protection-large-load-data-center/832479/
[76] Industrial Info Resources. “Report: Data Centers Create About US$6.3 Billion in Electric Costs for PJM Customers,” July 17, 2026. https://www.industrialinfo.com/iirenergy/industry-news/article/report-data-centers-create-about-us63-billion-in-electric-costs-for-pjm-customers–360334
[77] Mother Jones (from The Guardian). “Trump announces about 200 entities have signed nonbinding Ratepayer Protection Pledge,” 2026. https://www.motherjones.com/?p=1217343
