Introduction: The Week the Backlash Peaked and the Forecast Did Not Move
On October 5, 2026, Goldman Sachs published a research note that read, at first glance, like a contradiction sitting at the center of America’s artificial-intelligence infrastructure boom. Everything in the political environment appeared to be turning against the datacenter. Governors who had spent the preceding three years competing to attract hyperscale campuses were now imposing conditions, pausing permits and ending confidentiality arrangements. County boards that had once voted unanimously for rezonings were facing packed hearings. More than one hundred moratorium proposals were reportedly under consideration around the country, a figure that Amazon Web Services’ own chief executive would cite three days earlier in a lengthy public letter defending the industry. Public opinion had moved decisively: a June Reuters/Ipsos poll found that only about a third of Americans approved of the pace of datacenter construction, and only fourteen percent said they would support a facility being built in their own community to serve the AI projects of companies such as Meta, Alphabet, Amazon, Microsoft or xAI.[1] An NBC News poll cited in the Goldman note found that sixty-four percent of voters would be less likely to support a candidate who backed a local datacenter, against only eleven percent who would be more likely.[2]
And yet Goldman did not respond to this deteriorating environment by forecasting the collapse, or even the stall, of the American buildout. It did something more interesting. It raised its estimate of U.S. datacenter capacity at the end of 2026 by five gigawatts, to approximately 64 gigawatts. It reduced its year-end 2027 estimate by the same five gigawatts, but still expected capacity to reach roughly 90 gigawatts. It continued to forecast that U.S. datacenter power demand would grow 38 percent, or twelve gigawatts, in 2026, and another 38 percent, or seventeen gigawatts, in 2027, measured December against December.[1] The trajectory had bent slightly, shifting some capacity forward into the current year and pushing some out of the next. It had not broken.
“While US data center development faces increasing political and community opposition, project data and our updated capacity estimates suggest that the US data center growth outlook through 2027 remains largely unchanged.”
— Laura Cyr, Goldman Sachs[2]
That finding captures one of the most important, and least appreciated, structural features of the emerging Five-Layer AI Economy. Public opinion can move in weeks. Elections can change governments overnight. A governor can announce restrictions in an afternoon, a legislature can introduce a moratorium in a morning, a utility commission can reopen a tariff on a single agenda item, and a county board can suddenly find three hundred residents in a room demanding that another hyperscale campus be stopped. But a multibillion-dollar infrastructure system does not respond on the same timetable, because it was never built on that timetable in the first place. By the time the political argument reaches the ballot box, land may already have been assembled, substations planned, utility studies completed, transformers ordered, generation contracted, financing syndicated, construction labor mobilized, tax arrangements negotiated, and future GPUs reserved. Each of those steps was taken months or years earlier, in a different political climate, by actors who were responding to the incentives of that earlier moment rather than to the anger of the present one.
The gap between political time and infrastructure time is becoming especially consequential as the United States approaches the November 3, 2026 midterm elections, because the datacenter has migrated from an obscure question of industrial land use into the politics of household affordability. In the final days before the Senate left Washington to campaign, Republican leaders reserved one of the chamber’s last votes for legislation sponsored by Senator Jon Husted of Ohio that would have directed the Federal Energy Regulatory Commission to issue rules requiring large electricity customers to pay for the grid upgrades their demand necessitated. The bill had passed the House overwhelmingly. In the Senate it failed to reach the sixty-vote threshold, after Senate Democrats, who had offered competing and more aggressive proposals, dismissed it as inadequate.[3] Senate Democratic Leader Chuck Schumer did not mince words.
“The bill is a fraud, plain and simple.”
— Senator Chuck Schumer, Senate Democratic Leader[3]
Senate Majority Leader John Thune framed the vote as a test of whether Democrats would accept a partial victory on an issue voters cared about, and Republicans positioned the failed vote as evidence of Democratic obstruction.[3] But the underlying political dynamics were more uncomfortable for both parties than the floor rhetoric suggested. In August, the National Republican Senatorial Committee had circulated an internal memo warning that, more than any other factor in the Ohio race, datacenters were the anchor hanging around Husted’s neck, a reference to his support for datacenter tax incentives during his years as Ohio’s lieutenant governor.[4] Reporting in The Hill described a party increasingly worried that datacenters had become a pocketbook liability, caught between voter anger over power bills and a White House that had embraced the facilities as a matter of national strategy.[5] Lawmakers were expected to return to the question after the election, but not before.
The argument, in other words, is no longer simply whether the United States needs more artificial-intelligence infrastructure. The argument is increasingly about who pays for the electricity system required to support it, who decides where it goes, who is allowed to know what has been agreed, and what the communities that host it are owed in return.
The political backlash is real, and this paper does not minimize it. Virginia’s Governor Abigail Spanberger has unveiled a Data Center Accountability Framework and signed an executive order putting its core elements into immediate effect.[6] Pennsylvania’s Governor Josh Shapiro has removed every datacenter project from the state’s expedited permitting program, prohibited nondisclosure agreements, and conditioned state permits on prior local approval.[7] Texas Governor Greg Abbott has halted all state-issued permits sought by datacenter projects while ERCOT, the Public Utility Commission and the Texas Water Development Board audit their consequences for the grid, for water and for ratepayers, after the queue of datacenters and other large loads seeking connection to the Texas grid grew beyond 470 gigawatts, more than five times the state’s all-time peak demand.[8][9] California Governor Gavin Newsom has signed seven bills imposing additional requirements on electricity costs, water use, reporting and local oversight.[10] New York Governor Kathy Hochul has gone further still and signed the first statewide pause on new hyperscale datacenters in the nation’s history.[11]
And yet these measures coexist with extraordinary industrial commitments that were not reversed, slowed or even visibly dented by them. TSMC’s planned Arizona investment reached approximately $265 billion in July 2026 after the company added a further $100 billion for four more two-nanometer fabs and additional advanced packaging capacity.[12] Nvidia, OpenAI, SB Energy and SoftBank finalized in August 2026 the structure of an eight-gigawatt AI campus in Pike County, Ohio, with a twenty-year lease, a $105 billion residual-value guarantee from Nvidia, and at least ten gigawatts of new generation to be built alongside it.[13] Amazon activated its $11 billion Project Rainier campus in Indiana, committed a further $15 billion for northern Indiana, continued its development in Virginia, Louisiana, Mississippi and North Carolina, and on October 2, 2026 pledged more than $1 billion over five years to the communities hosting its datacenters.[14][15] Michigan has established affordability protections and secured pledges from six companies, including Anthropic, Google, Microsoft, OpenAI, Oracle and Verrus, that they will pay the full cost of their own development while the state simultaneously tries to ensure that large technology projects can still enter.[16] The four largest hyperscalers told investors during their second-quarter 2026 earnings calls in late July that they would spend, in combination, somewhere in the vicinity of $730 billion to $760 billion on capital expenditure this calendar year, up from roughly $413 billion in 2025, and three of the four raised their guidance in that very round of reporting.[17]
These developments suggest that America’s political response is shifting from the question “Should datacenters exist?” toward a much harder and more durable question: “Under what economic, environmental and infrastructure conditions will they be allowed to continue?”
This paper calls the phenomenon that produces this shift Buildout Inertia. The concept describes the tendency of large AI-infrastructure programs to continue advancing after sufficient physical, financial, contractual, regulatory and supply-chain commitments have accumulated, even when the political environment that originally welcomed them has deteriorated. The farther a project moves from announcement toward energization, the greater the economic cost of reversing it, and the more counterparties acquire a stake in its completion. Opposition can still delay construction, change cost allocation, impose environmental requirements, renegotiate community benefits, redirect generation sources or alter future phases. But cancellation becomes progressively more difficult as commitments multiply, because each commitment is simultaneously a promise made to someone else who has, in turn, made commitments of their own.
Buildout Inertia does not mean that every announced datacenter will be built. Quite the opposite. Power shortages, financing difficulties, interconnection delays, construction constraints and local opposition are already separating speculative projects from executable ones, and that separation is one of the central themes of this paper. The Federal Energy Regulatory Commission recently confronted a PJM capacity shortfall of 6,831 megawatts for the 2028/2029 delivery year, the first time in the history of the nation’s largest grid operator that the entire region fell short of its reliability requirement, and FERC’s September 29 order conditionally accepting PJM’s emergency procurement insisted that costs be allocated to the large new loads driving the shortfall rather than to existing customers.[18][19] Credit markets, which by early August had absorbed nearly $500 billion of AI-related debt issuance in 2026 alone, are similarly demanding stronger protections against permitting, electricity and construction delays.[20][21]
The more important point is that the projects which survive these filters become increasingly difficult to dislodge. A site with an option on land is relatively reversible. A site with approved zoning is less so. Add a signed utility service agreement, a twenty-year power contract, several billion dollars of debt, custom transformers with multiyear lead times, an anchor hyperscaler lease, reserved GPU deliveries and thousands of construction workers, and the project has entered a fundamentally different political category. What began as a proposal has become an industrial commitment, and industrial commitments have a constituency that proposals never do.
Seen through the Five-Layer AI Economy, this matters far beyond Layer 3, the datacenters themselves. Layer 1 energy investments, from gas turbines and nuclear restarts to substations and transmission, are increasingly planned around Layer 3 demand. Layer 2 chip orders, now running at a pace that produced $89 billion of Nvidia datacenter revenue in a single quarter, depend on the date electricity becomes available at a specific site.[22] Layer 4 frontier models require ever larger training and inference clusters whose construction schedules are set years in advance. Layer 5 applications and agentic systems ultimately depend upon the physical infrastructure beneath them, and are being designed today around capacity that will not be energized until 2028 or 2029. Once commitments begin propagating across all five layers, stopping one datacenter can mean unwinding contracts and investments extending from utilities and construction companies to Nvidia, AMD, TSMC, hyperscalers, banks, power generators, insurers and local governments.
The central argument of this paper is therefore that America’s datacenter politics are entering a second stage. The first stage, roughly 2023 through 2025, was dominated by competition to attract AI investment, and its characteristic instruments were sales-tax exemptions, fast-track permitting, confidential site-selection negotiations and ribbon cuttings. The second stage, visible throughout 2026, is dominated by the politics of managing infrastructure that has already acquired momentum, and its characteristic instruments are cost-allocation tariffs, disclosure mandates, local-approval requirements, water assessments, community-benefit agreements and audits of interconnection queues. Governors, utility commissioners, legislators and local governments are discovering that their greatest authority often exists before the commitment chain becomes deeply embedded. Afterward, they may have considerably more power to change the terms of development than to reverse its direction.
The remainder of the paper proceeds as follows. The section immediately following explains the choice of title. Section 1 develops the analytical framework: the Commitment Ladder, the distinction between political announcement time and infrastructure commitment time, the Reversibility Curve, and the mapping of inertia onto the five layers. Section 2 examines the quantitative foundations of the momentum, drawing on corporate earnings through the second quarter of 2026 and on the load-growth literature, while taking seriously the scholars who warn that much of the projected demand is phantom. Section 3 examines the state pushback through five cases: Virginia, Pennsylvania, Texas, California and New York. Section 4 examines projects that have already become difficult to reverse: Arizona, Ohio, Indiana and Michigan. Section 5 turns to the question of who pays for momentum, with PJM, the White House Ratepayer Protection Pledge, the emergence of the pay-your-own-way principle, community compensation and credit-market discipline. Section 6 addresses the November 2026 election and the politics of what can still be changed, and proposes the Infrastructure Intervention Window. Section 7 draws together eight pillars of lessons. The conclusion returns to the paradox with which the paper began.
Why I Chose the Title “Buildout Inertia”
I chose Buildout Inertia because the word buildout describes something larger than construction, and because the word inertia describes something more specific than momentum.
Buildout, in the vocabulary that has grown up around the AI economy since 2023, encompasses the entire chain through which artificial intelligence becomes physical. It begins with land acquisition and the quiet assembly of parcels, often through intermediaries and under nondisclosure agreements, in places where electricity, water, fiber and tax treatment align. It continues through electricity procurement, interconnection studies, transmission upgrades, substations, cooling systems and backup generation; through the allocation of semiconductors whose fabrication was itself scheduled years earlier; through project financing that increasingly runs through investment-grade bonds, private credit, asset-backed structures and vendor guarantees; through permitting at every level of government; through the mobilization of a construction labor force that in some projects numbers in the tens of thousands; and ultimately through the energization of racks that convert all of these prior decisions into tokens, inference and revenue. To speak of the buildout is to speak of this entire chain, and to recognize that a decision made at any one link constrains the choices available at every later link.
Inertia describes what happens after these commitments begin reinforcing one another. In physics, inertia is not motion; it is the resistance of a body to changes in its motion. A massive object already in motion can be slowed, redirected or made more expensive to move, but reversing it requires a force proportional to its mass and its velocity, and the political system rarely possesses force of that magnitude at the moment it most wishes to apply it. That is precisely the situation in which American governors, regulators and communities now find themselves. They are not confronting a proposal that can be declined. They are confronting a system in motion whose mass is measured in hundreds of billions of dollars of committed capital, gigawatts of contracted power, and twenty-year leases guaranteed by the most valuable company in the world.
The title also captures the paper’s political argument. Elections can change sentiment quickly, and the 2026 midterm cycle has demonstrated how quickly: a facility class that was an unalloyed economic-development trophy in 2024 had become, by the autumn of 2026, something that sixty-four percent of voters would hold against a candidate.[2] But power plants, transmission systems, semiconductor fabs and hyperscale campuses operate on multiyear timelines that are indifferent to the electoral calendar. Buildout Inertia is therefore the widening gap between the speed at which democratic politics can change its mind and the speed at which industrial infrastructure can change direction. That tension makes the term particularly appropriate for America’s AI economy heading into the November 2026 midterms and the much larger 2027–2030 infrastructure cycle, during which the capacity Goldman now forecasts for the end of 2027 will have to be designed, financed, permitted, powered and built under rules that are still being written.
Section 1: When Infrastructure Begins Moving Faster Than Politics
The first task of this paper is to explain why a forecast can remain intact while the politics surrounding it deteriorate, and to do so without resorting to the lazy explanation that industry simply overpowers democracy. The better explanation is temporal. Infrastructure and politics are governed by different clocks, and the capacity that will be energized in 2026 and 2027 was, for the most part, committed in 2024 and 2025. Political opposition therefore operates on a different inventory than the one the forecasts measure. It can and does affect what will be proposed, approved and financed in 2027, 2028 and beyond. It has very little purchase on a campus whose transformers were delivered last spring and whose GPUs are being racked this autumn. Understanding this distinction is the key to understanding everything else in the paper, and so this section develops it carefully, building a vocabulary that the later sections will use: the Commitment Ladder, the two clocks, the Reversibility Curve, and the mapping of inertia onto the five layers of the AI economy.
1.1 The 64-Gigawatt-to-90-Gigawatt Paradox
Begin with the Goldman Sachs forecast of October 2026, because it crystallizes the problem. U.S. datacenter capacity is expected to rise toward approximately 64 gigawatts by the end of 2026 and roughly 90 gigawatts by the end of 2027 even as political resistance intensifies, and even as the same note acknowledges that resistance explicitly, citing Governor Abbott’s Texas directive and the sharp deterioration in public sentiment.[2] Morgan Stanley, for its part, continues to estimate a U.S. datacenter power shortfall through 2028 on the order of a third of projected need, which is to say that the binding constraint on the buildout is still supply, not opposition.[1]
The apparent contradiction becomes the opening analytical problem of the paper: if opposition is rising so rapidly, why does capacity continue rising almost as rapidly? Three partial answers present themselves, and all three are correct.
The first answer is that political opposition primarily affects future optionality, while much near-term capacity originates from decisions already made months or years earlier. A hyperscale campus energizing in the fourth quarter of 2026 typically secured its land in 2023 or 2024, its interconnection agreement in 2024, its long-lead electrical equipment in late 2024 or early 2025, and its construction financing before the end of 2025. None of those decisions can be unmade by a moratorium adopted in September 2026, and most of them cannot be unmade by anyone at any price short of litigation.
The second answer is that the 2026 wave of restrictions is, for the most part, prospective rather than retrospective. Governor Abbott’s Texas directive halts new permits and pauses the “Batch Zero” interconnection process pending an audit; it does not disconnect campuses already drawing power.[23] Governor Shapiro’s Pennsylvania order removes projects from the Fast Track program and conditions future state permits on local approval; it does not revoke the five permits already issued, and Shapiro himself noted that no AI datacenter was yet operating in the Commonwealth.[24] Governor Spanberger’s Virginia framework proposes a 25-megawatt local-approval threshold for future projects; it does not shut down Loudoun County.[25] Governor Hochul’s New York pause applies to new facilities of fifty megawatts or more; it does not touch the operating fleet.[26] The restrictions are real, but they bite on the pipeline, not on the installed base, and Goldman’s near-term forecast is a forecast of the installed base.
The third answer is more subtle, and it is the one that gives this paper its structure. The restrictions themselves are increasingly designed not to stop projects but to change their terms. The common thread running through Virginia, Pennsylvania, Texas, California, Michigan, Indiana, the White House proclamation of March 2026 and the FERC order of September 2026 is the demand that large loads pay their own way: for generation, for transmission, for substations, for water infrastructure, and for the community impacts they produce. A project that can satisfy that demand is not stopped by it. A project that cannot was probably never going to be built in any case. The political system, in other words, is converging on a filter rather than a wall, and a filter lets the strong projects through.
1.2 The Commitment Ladder
To make these dynamics concrete, it helps to think of a datacenter project as climbing a ladder, where each rung converts a measure of optionality into a measure of obligation. The Buildout Commitment Ladder set out below is a stylized sequence; real projects skip rungs, climb several at once, or stall between them. But the ordering captures the essential logic, which is that the number of counterparties, the quantity of capital at risk, and the cost of walking away all rise monotonically as the project ascends.
| Rung | Commitment | What is now at stake | Typical reversibility |
| 1 | Land control (option or purchase) | Option premium; broker fees | High: option can lapse |
| 2 | Zoning and entitlement | Legal fees; community goodwill | High to moderate |
| 3 | Environmental and water approvals | Studies; consultants; time | Moderate |
| 4 | Interconnection study and queue position | Deposits; screening fees; years of waiting | Moderate: position has scarcity value |
| 5 | Utility service agreement / large-load tariff | Minimum demand charges; collateral; termination fees | Low: contract is binding |
| 6 | Generation procurement (PPA, co-location, on-site) | Multiyear or multidecade take-or-pay | Low |
| 7 | Tax and incentive agreement | Clawbacks; public commitments | Low to moderate |
| 8 | Project financing / bond issuance | Debt covenants; lender rights | Low |
| 9 | Long-lead equipment orders (transformers, switchgear, turbines) | Nonrefundable deposits; 2–4 year lead times | Very low |
| 10 | Anchor tenant lease | 15–20 year revenue obligation; guarantees | Very low |
| 11 | Construction mobilization | Labor; contractor claims; sunk civil works | Very low |
| 12 | Chip allocation and delivery | GPU reservations in a supply-constrained market | Very low |
| 13 | Energization | Operating revenue; depreciation; customer dependence | Effectively irreversible |
Early-stage projects remain easy to abandon, and developers abandon them constantly; the industry’s own practice of approaching several utilities and several jurisdictions with the same prospective load, which Jonathan Koomey and others have documented and which this paper discusses in Section 2, is precisely a strategy of holding many low rungs simultaneously while committing to few. Late-stage projects contain multiple counterparties with money, equipment, labor and political credibility already invested, and those counterparties are not passive. A utility that has ordered a substation, a lender that has syndicated a loan, a turbine manufacturer that has reserved a production slot, a union local that has dispatched two thousand electricians, and a county that has budgeted the incremental property tax are all, in effect, allies of completion, whether or not they were allies of the original proposal.
The policy implication, which Section 6 develops, is that the appropriate regulatory instrument depends on the rung. A moratorium is a reasonable instrument for projects on rungs one through four and an extraordinarily expensive one for projects on rungs nine through twelve. A cost-allocation tariff, by contrast, can be applied at almost any rung, which is why it has become the instrument of choice across the political spectrum in 2026.
1.3 Political Announcement Time Versus Infrastructure Commitment Time
The paper’s principal temporal distinction can now be stated precisely.
Political Announcement Time measures the moment at which elected officials publicly support, oppose or regulate a project: the press conference, the executive order, the bill signing, the campaign advertisement. It is the time that newspapers record and that voters remember. It is also, almost by construction, the time at which the political system is paying the most attention to the project.
Infrastructure Commitment Time measures the moment at which financial and physical obligations become sufficiently embedded that reversal becomes costly: the signing of the utility agreement, the issuance of the notice to proceed, the wire transfer for the transformers, the closing of the financing. It is recorded in contracts, filings and purchase orders rather than in headlines, and it is frequently concealed by the very nondisclosure agreements that governors are now racing to prohibit.
These clocks rarely coincide, and the lag between them runs in a consistent direction. Commitment almost always precedes announcement, because developers, utilities and financiers have every incentive to lock in land, power and equipment before public attention, and the competition it attracts, arrives. A governor announcing new restrictions in September 2026 is therefore, in practice, governing projects whose foundational decisions were made in 2024 or 2025, and is doing so with far less leverage than the public announcement implies. Governor Shapiro’s own trajectory illustrates the point: he consulted Amazon and the industry on his GRID principles early in 2026, celebrated the company’s planned $20 billion Pennsylvania investment, and then in August removed Amazon’s projects, along with everyone else’s, from the Fast Track program.[27] The projects did not disappear. They moved to a slower track, under new conditions, with their underlying commitments intact.
The distinction also explains why the 2026 restrictions are so heavily weighted toward transparency. Executive Order 22 in Virginia, Executive Order 2026-05 in Pennsylvania, the public permitting map Pennsylvania has now published, the disclosure requirements in California’s AB 1577 and SB 1168, and Amazon’s own October 2 announcement that it will stop requiring nondisclosure agreements with government agencies all address the same underlying problem: the political system cannot intervene in Infrastructure Commitment Time if it does not know that commitments are being made.[6][7][28][29] Transparency is the precondition for closing the gap between the two clocks, which is why it has become the least controversial and most rapidly adopted element of the new regulatory consensus.
1.4 The Reversibility Curve
The Commitment Ladder describes discrete steps. The Buildout Reversibility Curve describes the continuous relationship between a project’s progress and the cost of stopping it, and it is the conceptual device that links the paper’s descriptive sections to its policy conclusions.
At the earliest stage, political intervention can eliminate a project cheaply, in the sense that the developer loses an option premium and some professional fees, the community loses a hypothetical, and no third party is injured. The curve is flat and low. As commitments accumulate, the cancellation cost rises, and it rises faster than linearly, because each new commitment is made in reliance on the earlier ones and because the number of injured parties multiplies. By the time a project reaches construction, the cost of cancellation includes not merely the developer’s sunk investment but contractor claims, lender remedies, equipment that cannot be resold at par, a utility’s stranded substation whose cost will be sought from remaining ratepayers, and an anchor tenant’s lost capacity in a market where capacity is scarce. The curve is now steep and high.
Eventually the policy question changes in kind rather than degree. It stops being “Do we permit this project?” and becomes “Who must pay for the infrastructure already required by this project?” That transition is crucial to understanding contemporary utility regulation, because the overwhelming majority of the 2026 fights, from PJM’s backstop procurement to Ohio’s AEP tariff to California’s AB 2383 to Michigan’s codified collateral requirements, are fights at the top of the curve rather than the bottom. They are not fights about whether the load will arrive. They are fights about the allocation of costs the load has already imposed or will inevitably impose.
The Reversibility Curve also helps explain a pattern in the data on blocked projects. Data Center Watch, the research project that tracks local opposition, counted at least 75 projects worth approximately $130 billion blocked or delayed in the first quarter of 2026, the largest single-quarter total on record and roughly equal to all of 2025 in three months, followed by 45 projects worth about $68 billion in the second quarter.[30][31] Those are enormous numbers. But they are, almost by definition, numbers from the bottom of the curve. A project that can be blocked by a town board is a project that has not yet climbed past the rungs at which a town board has power, and the developers who lose those fights, as the industry’s own strategists have acknowledged, typically redirect the same capital to the next jurisdiction rather than abandon the deployment.[32] Opposition is therefore highly effective at the bottom of the curve and highly ineffective at the top, and the aggregate forecasts reflect the top.
1.5 Buildout Inertia Across the Five-Layer AI Economy
The argument so far has treated the datacenter as the unit of analysis. The Five-Layer AI Economy framework requires widening the lens, because the datacenter is only the middle layer of a vertically interdependent system, and inertia in one layer propagates to the others.
| Layer | Domain | Representative actors | How Layer 3 inertia propagates here |
| Layer 1 | Energy | Utilities, independent power producers, gas-turbine makers, nuclear operators, transmission owners, grid operators | Generation and transmission are planned, financed and contracted against datacenter load; stranded-cost risk if load disappears |
| Layer 2 | Chips | Nvidia, AMD, Broadcom, TSMC, Micron, SK hynix, equipment and packaging suppliers | Fab capacity, CoWoS packaging and GPU production are scheduled against energization dates; $89B of Nvidia datacenter revenue in a single quarter depends on sites coming online |
| Layer 3 | Datacenters | Hyperscalers, neoclouds, developers, REITs, construction and cooling firms | The locus of physical commitment: land, power, shell, racks |
| Layer 4 | Models | OpenAI, Anthropic, Google DeepMind, Meta, xAI and other frontier labs | Training runs and inference fleets are planned around capacity contracted years ahead; a 20-year lease at PORTS-Pike is a Layer 4 commitment |
| Layer 5 | Applications and Agents | Enterprise software, consumer services, agentic systems, robotics | Products are designed and sold against compute that may not energize until 2028–2029 |
Buildout Inertia becomes stronger when contractual commitments extend through several layers simultaneously, and the 2026 deal structures were explicitly designed to produce exactly that effect. At PORTS-Pike, SB Energy builds and owns the shell (Layer 3) and at least ten gigawatts of generation (Layer 1); Nvidia supplies the compute exclusively and guarantees the lease (Layer 2); OpenAI occupies the capacity for twenty years (Layer 4); and the applications and agents that OpenAI sells (Layer 5) are the revenue against which the entire structure is underwritten.[13][33] At Project Rainier in Indiana, Amazon builds the campus (Layer 3), designs its own Trainium chips (Layer 2), contracts with Indiana Michigan Power and NIPSCO for power (Layer 1), and dedicates the whole to Anthropic’s models (Layer 4).[34][35] In Arizona, TSMC’s $265 billion of fabs and packaging facilities (Layer 2) exist to serve exactly the demand that Layers 3 and 4 are generating, and the company’s chief financial officer has described the acceleration of the Arizona buildout as a response to a multiyear demand megatrend from its U.S. customers.[12]
The consequence is that a political decision aimed at a single datacenter is never, in practice, a decision about a single datacenter. It is a decision about a chain of obligations that runs upward into models and applications and downward into turbines and wafers, and the parties at each link have both the resources and the motive to defend the chain. That is the structural reason that the backlash of 2026, however genuine, has so far bent the trajectory rather than broken it, and it is the reason that the remaining sections of this paper focus less on whether the buildout will continue than on the terms under which it will.
1.6 What Inertia Does Not Mean
It is important to be precise about the limits of the concept before applying it. Buildout Inertia is not a claim that resistance is futile, that regulation is pointless, or that communities should accept whatever is proposed to them. The empirical record of 2026 refutes every one of those readings. Resistance has blocked or delayed well over a hundred projects in six months.[32] Regulation has fundamentally altered the economics of large loads in PJM, Texas, Ohio, Michigan and California. Communities have extracted community-benefit funds, water commitments, local-hiring pledges and transparency that did not exist two years ago. What inertia does mean is that the character of political power over the buildout changes as the buildout advances: from the power to say no, which is greatest at the bottom of the Reversibility Curve, to the power to set conditions, which persists all the way to the top. The rest of this paper is an extended study of that second kind of power, exercised in 2026 by actors who have discovered, often to their frustration, that the first kind arrived too late.
Section 2: The Numbers Behind the Momentum, and the Scholars Who Doubt Them
Before examining how states and regulators have responded to the buildout, it is necessary to establish what, exactly, they are responding to, because the magnitude of the commitments made in 2025 and 2026 is the raw material of inertia. A buildout financed at $100 billion a year could be slowed by a bad election cycle. A buildout financed at $700 billion a year, with a further $500 billion of debt issued in eight months and a chip supplier recording triple-digit revenue growth, has acquired a mass that no single election cycle is likely to arrest. At the same time, intellectual honesty requires acknowledging that some of the most credible energy researchers in the country believe the demand forecasts driving this spending are inflated, and that a meaningful share of the capacity in interconnection queues will never be built. Both things can be true, and the tension between them is exactly what the Reversibility Curve predicts: committed capital is inert, while speculative capital is not. This section lays out both sides.
2.1 Corporate Commitments Through the Second Quarter of 2026
The second-quarter 2026 earnings season, which ran from Alphabet’s report on July 22 through Amazon’s on July 30 and Nvidia’s on August 26, produced the clearest picture yet of how much capital has been committed to the physical AI economy, and of how little the political backlash had altered the trajectory of that capital.
| Company | 2025 capital expenditure (approx.) | 2026 guidance entering the year | 2026 guidance after Q2-2026 earnings | Q2-2026 quarterly capex |
| Amazon | ~$131B | ~$200B (Feb. 2026) | ~$220B (raised July 30) | ~$44B+ |
| Alphabet | $91.4B | $175–185B | $195–205B (raised July 22) | $44.9B |
| Microsoft | ~$118B (CY) | ~$120B then ~$190B (Apr.) | ~$175–190B (basis change; held in July) | ~$41B |
| Meta | $72.2B | $115–135B | $130–145B (floor raised July 29) | $31.1B |
| Four-company total | ~$413B | ~$600–650B | ~$730–760B | ~$160B |
Sources for the table are the companies’ second-quarter disclosures as compiled by Statista and independent capex trackers; figures are approximate and reflect differences in fiscal years and lease accounting.[17][36][37]
Several features of this table deserve emphasis. First, the direction of revision was uniformly upward or flat; no major hyperscaler cut its guidance in response to the political environment, and Amazon and Alphabet raised theirs for the second time in the year.[17] Second, the ratio of capital expenditure to operating cash flow has reached levels without precedent for companies of this size; Meta’s second-quarter capital spending of $31.1 billion was equivalent to roughly 98 percent of its operating cash flow in the quarter.[37] Third, a portion of the increase reflects price rather than volume: Microsoft attributed roughly $25 billion of its 2026 capital spending to higher component prices, a reminder that the Layer 2 supply chain is itself capacity-constrained.[17] Fourth, and most important for the argument of this paper, these are not annual decisions. A $220 billion capital program is a portfolio of multiyear site commitments, equipment orders and power contracts, the great majority of which were initiated before 2026 and will not conclude until after 2027.
On the supply side of Layer 2, Nvidia’s results for the quarter ended July 26, 2026 provide the clearest single measure of how much compute is actually being deployed rather than merely announced. Revenue was $96.2 billion, up 106 percent from a year earlier; datacenter revenue was $89.0 billion, up 117 percent, driven by the ramp of Blackwell Ultra systems; and the company guided to $108 billion of revenue in the following quarter while assuming no datacenter compute sales to China.[22][38] In the preceding quarter, hyperscalers accounted for roughly half of datacenter revenue, with the remainder coming from AI clouds, enterprises, industrial customers and sovereign buyers, a diversification that itself increases inertia by multiplying the number of counterparties with deployed capital.[39] Jensen Huang’s framing of the quarter was characteristically expansive, but the underlying claim that compute had become directly monetizable is central to why the buildout has resisted political pressure.
“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[22]
The point is not that Huang’s optimism is necessarily correct. The point is that a supplier recording $89 billion of quarterly datacenter revenue has shipped $89 billion of equipment to sites that have power, cooling and racks ready to receive it, and that equipment is now installed capacity rather than a proposal. The semiconductor industry’s own forward commitments reinforce the picture: TSMC raised its 2026 capital expenditure guidance to between $60 billion and $64 billion in July while committing an additional $100 billion to Arizona, and the company’s CFO explicitly tied both decisions to multiyear customer demand.[12]
2.2 The Debt Wave
Capital expenditure of this magnitude cannot be funded entirely from operating cash flow, and 2026 is the year in which the AI buildout became a credit story as much as an equity story. Goldman Sachs Research estimated that nearly $500 billion of AI-related debt had been issued in 2026 by early August, with hyperscalers themselves accounting for only about 40 percent of that total; direct hyperscaler issuance had risen from $108 billion in all of 2025 to $194 billion in the first seven months of 2026, and the firm expected a further $300 billion of project-finance and datacenter transactions in 2027 on top of direct corporate issuance.[20] Reuters Breakingviews put the AI share of higher-rated U.S. issuance at roughly one-fifth in 2026, against about one percent in 2024.[40] By the beginning of October, trade reporting indicated that lenders had begun to demand higher yields and explicit legal backstops against construction, grid and permitting delays, and that a single Texas pause had placed nearly fifty gigawatts of proposed projects at risk of delay.[21]
Debt matters for inertia in two opposite ways, and the paper returns to this duality in Section 5. Once issued, debt is among the most binding of all commitments, because bondholders and private lenders possess contractual remedies that voters and county boards do not. But before it is issued, debt is among the most discriminating of all filters, because lenders underwrite the counterparty, the power contract and the permit before they fund the shell. The result is an increasingly sharp divide between announced capacity, which is enormous and politically alarming, and financeable capacity, which is smaller, more concentrated, and far more durable.
2.3 The Load-Growth Literature, 2024–2026
The demand side of the equation has produced a substantial and fast-moving literature, and policymakers’ expectations have been shaped by a handful of institutional estimates.
The Lawrence Berkeley National Laboratory’s congressionally mandated 2024 Report on U.S. Data Center Energy Use, published in December 2024, found that datacenters consumed approximately 176 terawatt-hours, or 4.4 percent of U.S. electricity, in 2023, up from 58 terawatt-hours in 2014, and projected consumption of between 325 and 580 terawatt-hours, or 6.7 to 12 percent of the national total, by 2028, with compound annual growth of 13 to 27 percent over the forecast period depending on scenario.[41][42] The International Energy Agency’s first dedicated Energy and AI report, released in April 2025, projected that global datacenter electricity demand would more than double to around 945 terawatt-hours by 2030, that AI-optimized facilities would more than quadruple their consumption, and that in the United States datacenters would account for almost half of all electricity demand growth through 2030, by which point the U.S. economy would consume more electricity processing data than manufacturing aluminum, steel, cement and chemicals combined.[43]
“AI is one of the biggest stories in the energy world today – but until now, policy makers and markets lacked the tools to fully understand the wide-ranging impacts. Global electricity demand from data centres is set to more than double over the next five years, consuming as much electricity by 2030 as the whole of Japan does today.”
— Fatih Birol, Executive Director, International Energy Agency[43]
The International Monetary Fund, drawing on the commodity special feature of its April 2025 World Economic Outlook and a companion working paper, Power Hungry: How AI Will Drive Energy Demand, framed the same trend as a macroeconomic policy problem rather than merely an energy one. Its staff noted that global datacenter consumption of as much as 500 terawatt-hours in 2023 could triple to 1,500 terawatt-hours by 2030, comparable to the entire electricity use of India, that U.S. datacenter demand was likely to more than triple to over 600 terawatt-hours under a McKinsey medium scenario, and that under current policies the AI-driven increase in electricity demand could add 1.7 gigatons of greenhouse-gas emissions between 2025 and 2030.[44] The Fund’s conclusion was that the technology’s growth contribution depended on policies to expand supply, incentivize alternative sources and contain price surges, which is a reasonably precise description of the agenda that American governors adopted a year later.
Private forecasts are larger still. Structure Research, cited in the 2026 registration statement of an Asian datacenter operator, projects that the ten largest global hyperscalers will deploy $7.3 trillion between 2026 and 2030, more than five times the $1.4 trillion of the preceding five years, and that global installed capacity will grow from 90 gigawatts in 2025 to 225 gigawatts in 2030.[45]
2.4 Phantom Load: The Case for Skepticism
The scholars who doubt these forecasts are not fringe voices, and their critique is directly relevant to the distinction between speculative and committed capacity on which this paper rests.
Jonathan Koomey, the energy researcher best known for Koomey’s Law on the historical doubling of computing energy efficiency and a former consulting professor at Stanford, has spent the past two years arguing that the current forecasts repeat the errors of the dot-com era, when projections of internet-driven electricity demand proved wildly overstated. Addressing the Ohio Manufacturers’ Energy Conference on August 27, 2026, he argued that forecasts built on today’s technology and short-term trends systematically ignore efficiency gains and adoption uncertainty, that prospective datacenter load is routinely counted more than once as developers approach multiple utilities, and that interconnection queues may overstate the demand that ultimately materializes by three to five times.[46][47]
“But it turns out forecasts are not reality.”
— Jonathan Koomey, Koomey Analytics[46]
Koomey’s critique is reinforced by Tyler Norris and his colleagues at Duke University’s Nicholas Institute, whose February 2025 study, Rethinking Load Growth, introduced the concept of curtailment-enabled headroom and estimated that the existing U.S. power system could absorb roughly 76 gigawatts of new load, equivalent to ten percent of national peak demand, if that load accepted curtailment averaging only 0.25 percent of annual hours, and up to 98 gigawatts at 0.5 percent.[48][49] Norris’s point is not that demand is imaginary but that the grid’s capacity to serve it has been underestimated, which has the same policy implication: much of the new generation and transmission being planned, and billed to ratepayers, against datacenter forecasts may be unnecessary.
“…power system capacity—intentionally designed to handle extreme peak demand swings—could accommodate significant load additions with modest flexibility measures.”
— Tyler H. Norris, Duke University Nicholas School of the Environment[50]
The Texas data illustrate the phantom problem vividly. ERCOT was tracking approximately fifteen gigawatts of large-load requests in December 2022. By late 2025 it had received 225 new large-load interconnection requests in a single year, against 152 for the entire 2022–2024 period, and by the time Governor Abbott ordered his audit the queue had grown past 470 gigawatts, against an all-time system peak of roughly 86 gigawatts.[51][9] No serious analyst believes that 470 gigawatts of datacenters will be built in Texas by any date, and the state’s 2025 legislation requiring a $100,000 screening fee and disclosure of duplicate applications was itself an acknowledgment that the queue had become a reservation system rather than a construction forecast.[51]
From the perspective of this paper, the skeptics and the forecasters are describing different parts of the same distribution. The skeptics are right that the queue is largely phantom; the forecasters are right that the committed subset of the queue is real, financed and advancing. Goldman’s 64 and 90 gigawatt figures are not derived from interconnection queues but from project-level data on sites under construction and in late development, which is to say from the top of the Reversibility Curve rather than the bottom.[2] The policy task, which Governor Abbott’s audit is the most ambitious attempt to perform, is to tell the two apart before ratepayers are asked to pay for the difference.
2.5 The Economists’ Divide
Behind the energy forecasts lies a deeper disagreement among economists about whether the investment will pay off at all, and that disagreement bears directly on the durability of inertia, because capital that is not earning a return will eventually stop being committed regardless of how embedded it has become.
Daron Acemoglu of MIT, the 2024 Nobel laureate, has been the most prominent skeptic. His 2024 paper, The Simple Macroeconomics of AI, projected that the technology would raise total factor productivity by well under one percent and GDP by roughly 1.1 to 1.6 percent over a decade, figures far too small to justify the capital spending now under way, and he has repeated that assessment in interviews on the possibility of an AI investment bubble as recently as February 2026.[52] Erik Brynjolfsson of Stanford’s Digital Economy Lab has taken the opposite view, arguing that the economy is passing through the trough of a productivity J-curve in which heavy intangible and infrastructure investment temporarily depresses measured output before the gains appear.
“We’re creating the potential to have massive productivity gains and a lot more wealth.”
— Erik Brynjolfsson, Stanford Digital Economy Lab[53]
“We are now transitioning out of this investment phase into a harvest phase where those earlier efforts begin to manifest as measurable output.”
— Erik Brynjolfsson, Stanford Digital Economy Lab[54]
For present purposes the resolution of this debate matters less than its existence, because it means that the ultimate check on Buildout Inertia is more likely to be financial than political. If Acemoglu is right, the commitments of 2025 and 2026 will be honored, because contracts are contracts, but the commitments of 2028 and 2029 will not be made, and the curve of new capacity will flatten of its own accord. If Brynjolfsson is right, the revenue will arrive and the inertia will be self-reinforcing. In either case, the capacity already committed for 2026 and 2027 will be built, which is exactly what Goldman’s forecast says.
Section 3: The States Push Back: From Datacenter Recruitment to Datacenter Conditions
American federalism has placed the primary responsibility for governing the buildout on the states, because land use, retail electricity rates, water rights, environmental permitting and economic-development incentives are all overwhelmingly state and local prerogatives. The result in 2026 has been a remarkable natural experiment: five large states, governed by five governors with different political profiles and different exposures to the industry, have confronted the same structural problem within the same four months and have produced five visibly different responses. This section examines each in turn and then draws out what they share. The deep explanation that precedes the cases is this: in every one of them, the governor is acting late in Infrastructure Commitment Time and early in Political Announcement Time, and the shape of each response is determined by how far up the Commitment Ladder the state’s projects had already climbed when the political climate turned.
3.1 Virginia: Governing the World’s Most Concentrated Datacenter Market
Virginia represents the mature phase of the problem, and therefore the case in which inertia is strongest and the available instruments are most constrained. Northern Virginia’s Data Center Alley hosts the largest concentration of datacenters on the planet; the facilities drew more than a fifth of the state’s electricity even before the AI wave, and the industry has become woven into the commonwealth’s tax base, its transmission planning, its land market and its economic-development identity.[55] Virginia was also the state in which the politics turned first. Governor Spanberger won the 2025 gubernatorial election in significant part on a promise to lower utility bills that voters blamed on the datacenter boom, and a January 2026 survey found that nearly three-quarters of Virginia voters attributed rising electricity costs to the facilities.[56][55]
On September 18, 2026, Spanberger unveiled what her office described as the most comprehensive and aggressive datacenter accountability effort in the country, and signed Executive Order 22 to put its executive-branch elements into immediate effect.[6] The Data Center Accountability Framework operates in five areas: transparency, environmental protection, energy costs, clean energy and Virginia workers. Its most consequential provisions are a ban on nondisclosure agreements for commercial datacenter projects, which the executive order imposes immediately on every agency under the governor’s authority; the elimination of by-right approval and a requirement of local approval for any datacenter using more than 25 megawatts, regardless of underlying zoning; the end of state subsidies for datacenters in Virginia’s site-development programs and the removal of future large projects from the state’s fast-track permitting process; tools for localities to negotiate community-centered agreements; a commitment to ensure that costs arising from datacenter-driven demand are allocated to datacenters; and limits on on-site natural-gas generation paired with incentives for projects that bring clean energy faster than the Virginia Clean Economy Act requires.[25][57][58]
“Virginia families and small businesses should not foot the bill for the data center industry.”
— Governor Abigail Spanberger, Commonwealth of Virginia[25]
The 25-megawatt threshold is the detail that most clearly reveals the logic of the second stage. Twenty-five megawatts is small by hyperscale standards; a single Project Rainier building exceeds it many times over. By setting the local-approval trigger that low, the framework ensures that essentially every future AI datacenter in Virginia will pass through a discretionary local process, which is precisely the rung on the Commitment Ladder at which community power is greatest. It is a deliberate attempt to move the Infrastructure Intervention Window earlier.
The case study should ask: how does a state regulate an industry after the industry has already become deeply embedded in its tax base, transmission system, land market and economic-development strategy? Virginia’s answer is instructive. It does not attempt to reverse what has been built, and it could not do so without devastating consequences for Loudoun and Prince William Counties’ budgets and for Dominion’s rate base. It instead changes the form of its authority: from recruitment to conditions, from secrecy to disclosure, from state preemption to local consent, and from socialized to allocated cost. Virginia demonstrates that Buildout Inertia does not eliminate regulatory authority. It changes its form.
3.2 Pennsylvania: Josh Shapiro and the End of Unconditional Fast-Tracking
Pennsylvania offers a different model, because its projects were several rungs lower on the ladder when the climate turned, and its governor therefore retained more leverage and used it more aggressively. On August 18, 2026, Governor Shapiro signed Executive Order 2026-05, Protecting Pennsylvania Consumers from Data Center Impacts, in a Harrisburg ceremony at which he described his purpose as stopping predatory backers and preventing bad proposals from infecting the Commonwealth.[59] The order removed every datacenter project, including Amazon’s, from the PA Permit Fast Track Program and made future projects ineligible; prohibited agencies under the governor’s jurisdiction from entering nondisclosure agreements for datacenter projects; directed the Department of Environmental Protection to publish a public map of all proposed projects and their permit status, which the department has done; conditioned the state’s sales-and-use tax exemption for datacenter equipment on compliance with the administration’s GRID standards; directed DEP to collect annual energy and water data from existing facilities; and, most significantly, barred the state from issuing permits for projects over 25 megawatts until the developer has demonstrated compliance with local plans and secured all required municipal approvals.[7][60][61]
“Effective immediately, I am removing all data centers from the Fast Track permitting program – and this executive order makes any future data center project ineligible for Fast Track.”
— Governor Josh Shapiro, Commonwealth of Pennsylvania[24]
“Specifically, we’re going to ensure that any infrastructure costs caused by data centers are paid by the AI data centers – not Pennsylvania homeowners or business owners – even if a data center ultimately closes and can’t pay.”
— Governor Josh Shapiro, Commonwealth of Pennsylvania[24]
The political context was unmistakable. Polling cited by the Philadelphia Inquirer found that only 24 percent of Pennsylvania voters approved of Shapiro’s handling of datacenters, a dramatic gap below his overall favorability, and the governor’s own account of the order’s origins emphasized a visit to Archbald, a rural town in the northeast slated for extensive development.[27] Shapiro, widely regarded as a potential 2028 presidential candidate, had earlier consulted Amazon and the industry on his GRID principles and had celebrated the company’s planned $20 billion investment in the state; the August order marked his most dramatic shift.[27][32]
Yet Pennsylvania remains attractive, and the order was carefully written to preserve that attraction for compliant projects. The Commonwealth has abundant natural gas, a mature electricity system within PJM, proximity to the Northeast’s demand centers, and a construction workforce with deep experience in energy infrastructure. Shapiro’s remarks acknowledged that Fast Track had been an incredible asset for projects such as Eli Lilly’s Lehigh Valley investment, and the order’s structure, which expedites DEP review for developers who sign binding consent agreements built on the GRID standards, rewards exactly the kind of committed, transparent, cost-internalizing project that Buildout Inertia tends to produce.[24][60] Legal commentators have questioned whether a governor can unilaterally rewrite the eligibility criteria for a legislatively enacted tax exemption, and litigation remains possible.[61] Pennsylvania therefore illustrates the transition from investment attraction to negotiated industrial growth, with the state’s leverage derived precisely from the fact that, as the governor himself noted, no AI datacenter was yet operating within its borders.[24]
3.3 Texas: Greg Abbott and the Politics of an Oversubscribed Grid
Texas provides perhaps the strongest illustration of political reversal, and the most consequential, because Texas had been on pace to overtake Virginia as the largest datacenter market in the world by the end of the decade.[62] Governor Abbott’s reversal unfolded in three steps across the late summer of 2026. In early August he directed the Public Utility Commission and ERCOT to halt approvals for new datacenter interconnections and to audit every datacenter in ERCOT’s queue for its effects on reliability and residential rates, pausing the so-called Batch Zero large-load process. On September 14 he directed the Texas Water Development Board to compel datacenters to comply with water-use reporting requirements, impose consequences for past noncompliance, and join ERCOT’s audit. And on September 21 he directed the Texas Commission on Environmental Quality to halt all air and water permits sought by datacenter projects until the audit is complete, with the further instruction that no state agency should move forward with any regulatory approval related to datacenter development until the information is in hand.[8][23][9]
“Simply put, Texans must come first.”
— Governor Greg Abbott, State of Texas[8]
The governor’s statement set three conditions for any project to proceed: it must cover all electrical infrastructure costs, it must result in lower residential electricity bills, and it must complete the ERCOT audit. His office added that in the next legislative session he would work to eliminate any financial incentives for datacenters.[8] ERCOT and the Water Development Board are surveying more than four hundred facilities, and ERCOT expects to finish the electric audit in December, which means the freeze will span the election and most of the subsequent quarter.[9] Industry analysts estimated that the pause exposed nearly fifty gigawatts of proposed capacity and as much as $8 billion of revenue through the first quarter of 2027.[62]
But the Texas story should not be interpreted simply as the end of datacenter development in the state, and the three conditions reveal why. They are not prohibitions; they are a filter. A project that has secured its own generation, that is structured under ERCOT’s large-load rules to curtail during emergencies, that can document its water supply, and that is prepared to pay for its own interconnection will pass the audit, and the audit’s primary effect will be to clear the queue of the four hundred-odd gigawatts of reservations that were never going to be built. The more consequential question is which projects survive the audit and which have accumulated enough land, power, financing and corporate commitment to remain viable. Texas may therefore become a laboratory for distinguishing real demand from speculative queue demand, which is a service to the committed projects even as it is a blow to the speculative ones. The pause also carries a Layer 2 implication: TSMC has been evaluating a Texas expansion, and the state’s treatment of large loads will shape where the next generation of fabs, as well as datacenters, is sited.[63]
3.4 California: Gavin Newsom and Regulatory Layering
California’s response is distinctive because it came through the legislature rather than executive order, because it reversed the governor’s own position of a year earlier, and because it addresses a market that is, by most measures, already in relative decline. On September 21, 2026, Governor Newsom signed seven bills that his office described as the most comprehensive datacenter laws in the nation.[10] The package requires datacenters to disclose information about electricity use, water consumption, land use and workforce needs to local governments; establishes separate rate structures so that large datacenters pay their share of grid-upgrade costs and comply with state procurement requirements rather than shifting costs to other customers, with explicit protection for low-income ratepayers; requires proposed facilities to provide water suppliers with information on anticipated consumption, available supplies, efficiency and drought planning; bars cities and counties from approving new or expanded facilities without a water assessment and scarcity plan; and places the cost of any water-system upgrades on the project.[10][64] The measures include AB 1577 on reporting, AB 2383 on electricity, AB 2469 on water disclosure, AB 2619 on water resources and SB 1168 on transparency.[28]
“With these laws, we are ensuring that Californians remain in the driver’s seat — and that those profiting from data centers aren’t doing so at our expense.”
— Governor Gavin Newsom, State of California[65]
The reversal is notable. In 2025 Newsom had vetoed legislation requiring datacenter water-use reporting and had positioned himself as a supporter of the industry; by 2026, political pressure and rising consumer costs produced a comprehensive package.[64] Industry voices warned that the laws would further limit development in an already declining market and push jobs, clean-energy deployment and tax revenue to neighboring states.[65]
California demonstrates another feature of Buildout Inertia: regulation often accumulates around an expanding industry rather than eliminating the industry. The state can require greater disclosure, environmental compliance, ratepayer protections and local participation while the underlying demand for compute continues increasing, and the capacity that California declines to host will be built in Arizona, Nevada, Texas or Oregon rather than not built at all. For the Five-Layer framework, California matters less for its datacenters than for its Layer 4 and Layer 5 concentration: the state is home to most of the frontier labs and application companies whose demand the rest of the country is being asked to serve, and its regulatory choices therefore export physical load while retaining the economic returns, a pattern that is becoming its own source of interstate political friction.
3.5 New York: Kathy Hochul and the First Statewide Pause
New York merits inclusion because it took the step the other four did not. On July 14, 2026, Governor Hochul signed an executive order pausing state environmental permits for new hyperscale datacenters of fifty megawatts or more for up to one year, while the Department of Public Service completes a generic environmental impact statement on the industry’s effects on the grid, water, air quality and host communities.[11][26] It was the first statewide pause in the country, adopted after the legislature had passed its own, broader Responsible Data Center Development Act with a 20-megawatt threshold, new rate classes and public-hearing requirements; Hochul neither signed nor vetoed that bill but used executive action to impose a narrower version, in a maneuver attorneys compared to her handling of congestion pricing in 2024.[66]
“This pause will remain in place for up to one year, while New York establishes the strongest possible framework to protect our communities.”
— Governor Kathy Hochul, State of New York[11]
New York could afford a pause for the same reason Pennsylvania could afford to remove projects from Fast Track: its hyperscale pipeline was still near the bottom of the Commitment Ladder. The state had no operating AI campuses of the scale found in Virginia or Texas, and the political cost of freezing proposals was therefore low. The experiment is nonetheless significant, because it tests whether a one-year pause at the proposal stage can be converted into a durable regulatory framework before the pipeline resumes, which is exactly the use of the Infrastructure Intervention Window that Section 6 recommends. By August, more than one hundred cities, towns and counties nationwide had adopted some form of local pause.[67]
3.6 From “Yes or No” to “Yes, But Under These Conditions”
Virginia, Pennsylvania, Texas, California and New York should be compared not merely according to whether their governors are pro-datacenter or anti-datacenter. That binary has become inadequate, and the table below suggests why.
| State | Governor | Principal instrument (2026) | Threshold | Position on Commitment Ladder when acting | Core demand |
| Virginia | Spanberger (D) | Framework + EO 22; 2027 legislation | 25 MW | Very high: largest installed base in the world | Allocate costs; end secrecy; local consent |
| Pennsylvania | Shapiro (D) | EO 2026-05 | 25 MW | Low to moderate: five permits, none operating | Local approval first; no Fast Track; no NDAs |
| Texas | Abbott (R) | Executive directives; PUCT/ERCOT/TWDB/TCEQ freeze | All projects in queue | Mixed: large installed base, enormous phantom queue | Pay all costs; lower bills; pass audit |
| California | Newsom (D) | Seven statutes | Large facilities | Moderate: declining relative market | Disclosure; separate rates; water assessment |
| New York | Hochul (D) | Executive pause (up to one year) | 50 MW | Low: pipeline at proposal stage | Build framework before building facilities |
The emerging state model, across party lines and across very different market positions, is: yes to AI infrastructure, but not necessarily with unlimited secrecy, socialized grid costs, unrestricted water use, automatic tax incentives or unconditional permitting. Even Texas, whose freeze is the most sweeping, framed its conditions as a path to approval rather than a prohibition. This represents the political maturation of the Five-Layer AI Economy: the moment at which the host layer stops behaving as a supplicant to the layers above it and begins behaving as a regulator of them.
Section 4: The Projects Already Becoming Difficult to Reverse
If Section 3 described the political system discovering the limits of its power, this section describes the objects against which that power is being measured. The four cases that follow were chosen because each sits high on the Commitment Ladder, each binds multiple layers of the AI economy into a single contractual structure, and each demonstrates a different mechanism by which inertia is generated: ecosystem formation in Arizona, contractual interlock in Ohio, institutional integration in Indiana, and asset revaluation in Michigan. The deeper point that unites them is that none of these projects is any longer merely a datacenter or merely a fab. Each has become a node in a regional industrial system, and regional industrial systems are governed by a different political arithmetic than individual buildings. Before examining them it is worth stating the obvious: these projects were not immune from the backlash. Arizona’s water politics, Ohio’s Senate race, Indiana’s farmland fights and Michigan’s ratepayer anxieties all touched them directly. What distinguishes them is that the backlash changed their terms and not their direction.
4.1 Arizona: TSMC and the Infrastructure Surrounding the Chip Fab
TSMC’s north Phoenix campus is the clearest American illustration of how an infrastructure program accumulates mass over time. The project was announced in May 2020 as a single fab costing roughly $12 billion. It grew to approximately $40 billion for two fabs in December 2022, to $65 billion in April 2024, to $165 billion for three additional fabs, two advanced-packaging facilities and a research center in March 2025, and on July 16, 2026, to $265 billion, after chairman C.C. Wei announced on the company’s second-quarter earnings call a further $100 billion for four more fabs at two nanometers and below plus additional packaging capacity, supported by up to $6.6 billion in direct CHIPS Act funding.[63][68][69] The company describes the program as the largest single foreign direct investment in the history of the United States.
“This is to build several more semiconductor logic wafer fabs for 2-nanometer and below technologies, as well as advanced packaging fabs, to support the strong multi-year demand from our leading U.S. customers.”
— C.C. Wei, Chairman and CEO, TSMC[69]
Although fabs are not datacenters, Arizona belongs in this paper for two reasons. The first is that the Arizona fabs exist to serve Layer 3 demand: the wafers that will be processed in Phoenix are destined for the accelerators that will be installed in the campuses that Goldman is counting, and TSMC’s chief financial officer has been explicit that the Arizona acceleration is a response to a multiyear demand megatrend from U.S. customers and to U.S. government support.[12] The second is that Arizona demonstrates the larger principle of ecosystem formation. Semiconductor manufacturing attracts suppliers, skilled labor, housing, water infrastructure, transportation, utilities and political partnerships; the north Phoenix site now spans more than a thousand acres with some seventy cranes in operation, and the surrounding economy has reorganized around it.[69] Once such an ecosystem forms, the investment becomes much more than a collection of buildings, and the political constituency for its completion extends to every supplier, landlord and school district that has planned around it.
It is also important to be precise about how much of the $265 billion is yet inert. Satellite analysis in mid-2026 showed that only three of the ten announced fabs exhibited vertical construction, with one in production; seven existed on paper and in permits.[63] In the vocabulary of this paper, the first three fabs are near the top of the Reversibility Curve and the remaining seven are in the middle, which means that Arizona’s water politics, which have become a prominent midterm issue, retain genuine leverage over the shape of the later phases even as they have none over the first. The lesson is that even strategically important technology projects ultimately compete for physical resources, and that the point of political leverage is always the next phase rather than the current one.
4.2 Ohio: The PORTS-Pike Commitment Chain
Ohio provides what may become the defining case study of Buildout Inertia, because the PORTS-Pike Technology Campus was deliberately structured, in the summer of 2026, to interlock all five layers of the AI economy into a single set of mutually reinforcing obligations.
The project sits on private land and remediated Department of Energy property near the former Portsmouth Gaseous Diffusion Plant in Pike County, one of the poorest counties in the state. In August 2026, Nvidia, OpenAI and SB Energy, the SoftBank-backed infrastructure developer, disclosed its final structure. SB Energy will build, own and operate the datacenter and lease it to OpenAI for twenty years, delivering roughly eight gigawatts of IT capacity over time. Nvidia will be the exclusive compute supplier, deploying its full DSX platform of GPUs, CPUs and networking, and has invested $1.5 billion directly in SB Energy alongside SoftBank and OpenAI. SB Energy and SoftBank will build at least ten gigawatts of new generation, including approximately 9.2 gigawatts of natural gas, and invest at least $4.2 billion in regional grid infrastructure through a partnership with AEP Ohio that is structured so that the project pays the full cost of the upgrades and transmission it requires. The initial phase is 4.25 IT-gigawatts, with Nvidia holding an option over the remaining 3.75. SB Energy and OpenAI have each committed $40 million to a community benefits fund, and OpenAI’s total package for Pike County and Ohio exceeds $160 million including education credits.[13][70][71][72]
The most remarkable element is financial. To make the campus financeable, Nvidia has provided a residual-value guaranty of OpenAI’s lease obligations covering the initial 4.25 gigawatts at an aggregate guaranteed value of $105 billion, declining on a defined schedule over the twenty-year term, under which Nvidia may assume the tenancy, direct re-letting or, failing that, direct a sale of the premises if a trigger occurs. The guaranty terminates automatically if OpenAI achieves a designated credit rating or substitute credit support is provided. SB Energy’s own registration statement discloses that the eight gigawatts at PORTS-Pike constitute the entirety of its contracted-but-not-under-construction capacity and a substantial majority of its datacenter backlog.[33]
| Layer | Party | Commitment at PORTS-Pike | Duration / amount |
| 1 — Energy | SB Energy, SoftBank, AEP Ohio | ≥10 GW new generation; ≥$4.2B regional grid investment; project pays upgrades and transmission | Multi-decade |
| 2 — Chips | Nvidia | Exclusive compute supplier; $1.5B equity in SB Energy; residual-value guaranty of lease | Up to $105B cap; 20-year schedule |
| 3 — Datacenter | SB Energy | Build, own, operate ~8 IT-GW campus | 20-year lease |
| 4 — Models | OpenAI | Anchor tenant for ~8 IT-GW; $40M community fund; investor in SB Energy | 20 years |
| 5 — Applications | OpenAI’s customers | Revenue against which the lease and guaranty are underwritten | Ongoing |
| Host | Pike County, State of Ohio, U.S. DOE | Land, remediation, workforce, $80M+ community fund, ~35,000 construction jobs | Ongoing |
This is Buildout Inertia in concentrated form. Land, electricity, financing, GPUs, an infrastructure developer, a model company and a semiconductor company have become contractually intertwined, and the structure was designed so that no single party can exit without triggering obligations for the others. Canceling such a project would not mean canceling one building. It would mean unwinding an industrial network: ten gigawatts of generation already in development, a $105 billion guaranty on the balance sheet of the world’s most valuable company, the anchor asset of a company preparing an initial public offering, and the compute plan of the most prominent frontier lab.
And yet Ohio is also the state in which the datacenter has become the single most potent electoral issue in the country. The Senate race between Jon Husted and Sherrod Brown has turned on Husted’s record of supporting datacenter tax breaks, the National Republican Senatorial Committee has described datacenters as the anchor around his neck, and the failed Senate bill of September 30 was, in large part, Husted’s attempt to neutralize that liability.[4][3] The juxtaposition is the paper’s thesis in miniature: the politics of the datacenter in Ohio have turned sharply hostile, and the largest datacenter project in the world is proceeding in Ohio anyway, because it was structured to pay its own way and because the chain of commitments behind it had already closed before the campaign began. The project’s economics are not immune from the politics; SB Energy’s own prospectus notes that its OpenAI leases pass through operating costs, and any change in Ohio’s cost-allocation rules will flow directly into those leases.[33] But the direction is set.
4.3 Indiana: The Hyperscaler Becomes an Energy-Policy Participant
Indiana illustrates a third mechanism of inertia: institutional integration, in which the hyperscaler stops behaving like an ordinary commercial electricity customer and becomes a participant in the state’s energy planning.
Amazon’s $11 billion Project Rainier campus in New Carlisle, in St. Joseph County, went from farmland to operation in less than two years. By mid-2026 seven buildings were operating, filled with roughly half a million of Amazon’s own Trainium2 accelerators and dedicated to training and running Anthropic’s Claude models, with thirty buildings planned across 1,200 acres and an ultimate load of approximately 2.2 gigawatts.[34] It was the largest capital investment in Indiana’s history until Amazon exceeded it, in November 2025, with a further $15 billion commitment for new campuses in northern Indiana adding 2.4 gigawatts of capacity and 1,100 permanent jobs.[14][35]
The energy arrangements are what make Indiana analytically important. For the northern Indiana expansion, Amazon negotiated a framework with NIPSCO and its generation subsidiary under which the company pays fees to use existing lines and covers the cost of any new plants, lines or equipment required to serve its load; the agreement could add up to three gigawatts of new capacity against Amazon’s 2.4-gigawatt requirement, producing a surplus that NIPSCO estimated would save other customers roughly $1 billion over fifteen years.[35] For Project Rainier, Indiana Michigan Power has moved to acquire a natural-gas plant in Ohio and expects the campus alone to double its peak demand by 2030, while Amazon has expanded local wind and solar projects totaling 635 megawatts.[34] At gigawatt scale, in other words, the hyperscaler becomes a participant in generation planning, transmission planning, workforce development, tax policy, utility finance and regional industrial strategy. Indiana’s governor described the NIPSCO arrangement as a guarantee of surplus energy development that would deliver savings to ratepayers; the Harvard scholars discussed in Section 5 would describe it as a special contract whose terms deserve scrutiny.[35] Both descriptions can be accurate.
That transformation strengthens Buildout Inertia because the datacenter becomes integrated into institutions beyond the campus boundary. A utility whose integrated resource plan assumes the load, a generation subsidiary whose new plant is underwritten by the load, and a community college whose fiber-technician program was funded by the developer are all, in the vocabulary of Section 1, allies of completion. Indiana also demonstrates the limits of that integration: the New Carlisle town council president has spoken publicly about the difficulty of continuing to lose farmland, and the local politics of the next thirty buildings will be harder than the politics of the first seven.[34][55]
4.4 Michigan: Gretchen Whitmer, Affordability and Responsible Growth
Michigan offers a particularly valuable contrast because its governor has attempted to capture the benefits of inertia while preempting its costs, and because the state’s most symbolically important energy project illustrates how AI demand is revaluing existing assets.
On July 15, 2026, Governor Whitmer announced the Michigan Affordable and Responsible Growth Action Plan, built on two pillars. The first calls on the legislature to codify in statute the safeguards the Michigan Public Service Commission has already imposed on large loads: minimum billing demand, contract termination fees, credit and collateral requirements, minimum contract terms and related provisions designed to ensure that datacenters pay the full costs they place on the system. The second is the Michigan Affordability and Responsible Growth Pledge, a package of ten commitments under which signatory companies agree to pay the full cost of construction and operation, invest in the grid, use clean energy consistent with the state’s 100 percent clean-energy standard, hire Michigan workers, protect water resources and ensure that their infrastructure costs never raise household rates or taxes.[73][74] Within days, six companies had signed: Anthropic, Google, Microsoft, OpenAI, Oracle and Verrus.[16]
“It’s simple: any data center company that wants to invest in Michigan must ensure working families do not pay a single penny for data center development or operations.”
— Governor Gretchen Whitmer, State of Michigan[75]
Whitmer explicitly rejected a moratorium of the kind New York adopted, arguing that enforceable conditions would protect ratepayers better than a pause.[74] The approach is the purest expression of the second-stage model: it does not ask whether datacenters should come to Michigan, it specifies the price of admission, and it converts that price into contractual and, if the legislature acts, statutory obligations that will bind the projects at every rung of the ladder.
At the same time, Michigan is pursuing the restart of the Palisades nuclear plant, an 800-megawatt facility on Lake Michigan that was shut down for decommissioning in May 2022 and that Holtec International, backed by a Department of Energy loan guarantee and state support, has been working since 2023 to return to service. The Nuclear Regulatory Commission restored the plant’s operating license status in 2025, the first such transition in U.S. history; the Federal Energy Regulatory Commission cleared its grid reconnection; and the company loaded fuel in 2026.[76][77] In late September 2026, however, Holtec delayed the restart indefinitely after a partially used fuel assembly shifted during refueling, and the plant remains in testing and restoration.[78] The important connection is not that Palisades exists solely for AI; it does not, and its restart predates the AI load forecasts. It is that renewed electricity demand has changed the value assigned to existing generation assets, so that a plant judged uneconomic to operate in 2022 was judged worth more than a billion dollars to revive by 2024, and that NextEra’s Duane Arnold plant in Iowa and Constellation’s Crane facility in Pennsylvania are following the same path. The AI era is altering not only what America builds, but what infrastructure America decides is worth bringing back, and that revaluation is itself a form of inertia operating in Layer 1.
4.5 Industrial Ecosystems Create Their Own Momentum
These four cases demonstrate a broader phenomenon that the project-level framing of most political debate obscures. A hyperscale project rarely remains isolated. It attracts power-generation investment, transmission, fiber, construction suppliers, transformer and switchgear manufacturers, chip contracts, workforce programs, housing, tax commitments and secondary companies, and each of these in turn makes commitments premised on the project’s completion. Arizona has acquired a semiconductor supply chain; southern Ohio is acquiring ten gigawatts of generation; northern Indiana is acquiring a new gas plant and a reorganized utility; Michigan is acquiring a revived nuclear plant and a template for large-load contracts.
Buildout Inertia therefore operates at the ecosystem level, not merely the project level. That is why the cancellation cost at the top of the Reversibility Curve is so much higher than the developer’s own sunk investment would suggest, and why the political actors who have confronted these projects in 2026 have, almost without exception, chosen to negotiate their terms rather than to attempt their reversal.
Section 5: Who Pays for Momentum?
The preceding sections have argued that the buildout’s direction is set and its terms are open. This section examines the single most important of those terms: the allocation of cost. It is here that the abstract politics of artificial intelligence have become concrete, because electricity bills arrive every month, because they arrive in every household regardless of whether that household has ever used a chatbot, and because the mechanisms by which datacenter costs reach those bills are technical, opaque and, until very recently, almost entirely hidden from public view. The intellectual foundation for the 2026 cost-allocation debate was laid in a single Harvard Law School paper published in March 2025, and the regulatory architecture that debate has produced, from the White House to FERC to the state commissions, is converging on a principle that this paper calls pay-your-own-way. The section ends by examining the two forces, community compensation and credit-market discipline, that are turning that principle into a component of project economics rather than a political slogan.
5.1 The Political Center of Gravity Moves to the Utility Bill
The political debate is migrating from abstract enthusiasm about AI toward a concrete monthly question: will households subsidize the infrastructure needed by trillion-dollar technology companies? That issue has the potential to become more politically powerful than arguments about models, algorithms or even AI safety because electricity bills arrive every month, and because the evidence that they are rising is not in dispute even where the attribution is.
A Bloomberg News analysis of wholesale electricity prices found that costs had risen by as much as 267 percent over five years in locations with substantial datacenter activity, a figure that Senator Elizabeth Warren and others have used to describe household bills, though economists caution that wholesale prices affect only the supply component of a residential bill, which is typically thirty to fifty percent of the total.[79][55] A June 2025 analysis by Carnegie Mellon University and North Carolina State University’s Open Energy Outlook Initiative projected that datacenter and cryptocurrency growth through 2030 could raise average U.S. electricity generation costs by eight percent, and by more than 25 percent in central and northern Virginia, while keeping more than 25 gigawatts of aging coal plants in operation that would otherwise retire.[80] PJM’s independent market monitor estimates that datacenters have cost the region’s 67 million ratepayers roughly $29 billion over about two years, and one estimate puts the annual burden on a Maryland household at $168 to $216.[81] A Consumer Reports profile of a Manassas, Virginia homeowner whose January 2026 bill jumped from roughly $100 to $281 captures the lived experience behind these aggregates.[55]
“In general, these cost increases are spread to all ratepayers by the utility.”
— Kenneth Gillingham, Yale School of the Environment[79]
“…these facilities can use the same amount of electricity daily that a small or medium-sized city, around 80,000 houses, would use daily.”
— Hannah Wiseman, Penn State Dickinson Law[82]
It is important to record the counterevidence, because the industry relies on it and because it is not frivolous. A 2026 working paper by researchers at EPRI and Watershed, using causal methods, estimated that each ten percent increase in datacenter capacity was associated with a 0.4 percent decrease in average residential retail prices, consistent with economies of scale in fixed-cost recovery, and found that Virginia’s residential rate increases had tracked the national average.[83] Lucy Qiu of the University of Maryland’s School of Public Policy has emphasized that how much of any increase reaches households, and when, depends on utilities’ purchasing contracts, retail rate design and regulatory decisions about who pays, so that no single national number is possible.[81] The disagreement is less about whether datacenters add cost to the system, which they plainly do in the short run as new generation and transmission are built, than about whether well-designed tariffs can ensure that the datacenters themselves pay it. That is precisely the question the 2026 regulatory agenda is attempting to answer.
5.2 The Harvard Diagnosis: Extracting Profits from the Public
The intellectual turning point in the cost-allocation debate was the March 2025 publication by the Harvard Electricity Law Initiative of Extracting Profits from the Public: How Utility Ratepayers Are Paying for Big Tech’s Power, by Eliza Martin and Ari Peskoe. The authors reviewed nearly fifty regulatory proceedings concerning utility rates for datacenters and concluded that conventional ratemaking, which socializes the cost of system expansion across all customers on the premise that growth benefits everyone, was being used to transfer the energy costs of the world’s wealthiest companies to captive residential ratepayers, through special contracts, discounted rates, co-location arrangements and forecasts that justified investments whose costs would be borne by the public whether or not the datacenters materialized.[84][85]
“Without systematic changes to prevailing utility ratemaking practices, the public faces significant risks that utilities will take advantage of opportunities to profit from new data centers by making major investments and then shifting costs to their captive ratepayers.”
— Eliza Martin and Ari Peskoe, Harvard Electricity Law Initiative[85]
“We’re all paying for the energy costs of the world’s wealthiest corporations.”
— Ari Peskoe, Director, Harvard Electricity Law Initiative[86]
The paper’s influence is visible in almost every 2026 state action. Its recommendations, that new datacenters be required to take service under published tariffs rather than secret contracts, that regulators scrutinize special contracts under explicit guidelines, and that cost-causation principles be applied rigorously to large loads, are now the law or the executive policy of Virginia, Pennsylvania, Michigan, California and Texas. By October 2026 Peskoe was emphasizing that the single greatest obstacle to answering the public’s question about who pays remained secrecy.
“Everyone pays those costs. They’re spread across the region to every business and resident that has an electricity meter.”
— Ari Peskoe, Director, Harvard Electricity Law Initiative[81]
The Harvard analysis also explains why transparency and cost allocation have become inseparable in the 2026 agenda. A regulator cannot allocate costs it cannot see, and a public cannot judge an allocation it is not permitted to examine. The nondisclosure agreements that governors in Virginia, Pennsylvania and elsewhere have now prohibited, and that Amazon has volunteered to abandon, were the mechanism by which Infrastructure Commitment Time was hidden from Political Announcement Time.[6][7][29]
5.3 PJM and the 6,831-Megawatt Warning
The PJM Interconnection provides the crucial federal case, because it is where the arithmetic of inertia first produced an outright reliability shortfall and where the federal government was forced to decide who would pay to close it.
PJM, which operates the grid for 67 million people across thirteen states and the District of Columbia and is home to Virginia’s Data Center Alley, held its base capacity auction for the 2028/2029 delivery year in July 2026. For the first time in the organization’s history, the entire region fell short of its reliability requirement, by 6,831 megawatts, and only the price cap kept the clearing price at roughly $325 per megawatt-day. The shortfall was driven largely by datacenter load forecasts outrunning new supply; PJM estimates that datacenter and other large-load demand could grow by as much as 70 gigawatts by 2038.[87][88] In response, PJM’s board directed a one-time Reliability Backstop Procurement, filed with FERC on July 31, to secure new generation for terms of up to fifteen years at a weighted-average cap of $555 per megawatt-day, with costs allocated to the load zones and load-serving entities whose incremental large-load additions drove the need; estimates of the procurement’s cost ran to $20 billion.[18][87] PJM paired the procurement with a proposed Interim Resource Adequacy Service that would curtail large loads of fifty megawatts or more that did not bring their own capacity.
On September 29, 2026, FERC conditionally accepted the procurement, finding eleven of its fifteen key elements just and reasonable, but rejected PJM’s cost-allocation method as unsupported, offered an alternative tied to updated load forecasts, and refused to allow cooperatives and municipal utilities to opt out on the ground that doing so would discriminate against other load-serving entities with datacenters in their territories. FERC’s chair described the situation as a mess, and PJM suspended the procurement, which had been scheduled to open on September 30, the following day; it does not expect to resume before late February 2027.[19][88] FERC’s order emphasized that states have a central role in managing large-load growth, and PJM’s own statement committed it to allocating costs to the customers driving them.[19] Asked whether FERC’s directions did enough to keep datacenter costs off existing residential bills, PJM’s independent market monitor, Joseph Bowring, answered simply that they did not.[89]
This should become one of the paper’s central policy examples, because it demonstrates both halves of the argument at once. The shortfall is a product of inertia: the load was committed before the generation was, and the region now faces a 2028 reliability gap that no moratorium adopted in 2026 can close, because the facilities driving it are already under construction. But the response is a product of the second-stage politics: the entire federal and regional apparatus is now organized around the proposition that the large loads must pay. Buildout Inertia does not mean the public must accept every cost created by infrastructure momentum. It means regulators must decide where those costs should land, and in PJM they are deciding, slowly and contentiously, that they should land on the datacenters.
5.4 The White House Pledge and the Emergence of the “Pay Your Own Way” Principle
The most striking feature of the 2026 cost-allocation debate is its bipartisanship. On March 4, 2026, President Trump issued Proclamation 11014, the Ratepayer Protection Pledge, under which seven leading hyperscalers and AI companies accepted a set of commitments at a White House ceremony. The proclamation declared it the national policy of the United States that companies increasing electricity demand must pay the full cost of the energy and infrastructure needed to build and operate datacenters and must not pass that cost to the American people, and that signatories would build, bring or buy the new generation resources their demand required and pay for all new power-delivery upgrades to serve their facilities.[90] The pledge is non-binding, and critics have noted the absence of enforcement, but its language is nearly identical to the demands of Democratic governors in Virginia, Pennsylvania, California and Michigan, and it has been invoked by utilities designing the large-load tariffs that make the principle operational.[91]
“The hyperscalers and AI companies that increase electricity demand must pay for the full cost of the energy and infrastructure needed to build and operate data centers, and must not pass this cost on to the American people.”
— Proclamation 11014, The White House[90]
Across Texas, Michigan, Ohio, Pennsylvania, Virginia, California, the White House and the federal regulatory debate, a common principle has therefore emerged: large AI loads should increasingly finance the incremental infrastructure they require. The instruments through which that principle is being implemented are now reasonably standardized.
| Instrument | Function | Where adopted or proposed (2025–2026) |
| Dedicated or self-supplied generation | Load brings its own capacity | White House pledge; PORTS-Pike; PJM interim service; Texas conditions |
| Direct funding of substations and transmission | Removes network costs from rate base | Indiana–NIPSCO framework; SB Energy–AEP Ohio; California AB 2383; Spanberger framework |
| Minimum demand (take-or-pay) charges | Protects against under-utilization | Michigan PSC; Ohio AEP tariff; multiple state tariffs |
| Termination penalties and minimum contract terms | Protects against stranded assets | Michigan PSC; Pennsylvania GRID standards |
| Credit and collateral requirements | Protects against counterparty failure | Michigan PSC; Shapiro order (costs paid even if facility closes) |
| Separate rate classes | Isolates cost causation | California AB 2382/2383; New York legislation; Oregon, Utah |
| Screening fees and duplicate-application disclosure | Clears phantom load from queues | Texas SB 6; ERCOT Batch Zero audit |
| Curtailment obligations | Reduces capacity need | Texas large-load rules; PJM interim service; Duke “headroom” model |
| Community-benefit agreements | Compensates host communities | Pennsylvania; Virginia; PORTS-Pike; AWS Built Together |
| Disclosure and NDA prohibitions | Enables all of the above | Virginia EO 22; Pennsylvania EO 2026-05; California AB 1577/SB 1168; AWS policy |
This could become one of the defining policy developments of 2027–2030, because it reconciles the two facts that this paper has argued are both true: that the buildout will continue, and that the public is no longer willing to subsidize it. The Data Center Coalition, the industry’s principal trade association, has publicly stated that the industry is fully committed to paying the full cost of service for its electricity.[92] Whether that commitment is honored will be determined tariff by tariff and contract by contract, which is why the transparency provisions matter so much.
5.5 Community Compensation Becomes Part of Infrastructure Economics
Amazon Web Services’ announcement on October 2, 2026 that it will invest more than $1 billion over five years in the U.S. communities hosting its datacenters is particularly significant, less for the sum than for what it reveals about the pricing of social license. The Built Together program covers out-of-pocket community-college costs for an estimated 300,000 students in datacenter communities, expands free skilled-trades training through modular training centers, funds energy-efficiency upgrades targeting more than 30,000 homes and 300 public facilities with the aim of cutting their energy costs by twenty to forty percent, and advances water-replenishment projects toward a commitment to be water-positive by 2030. The company simultaneously announced that it would stop requiring nondisclosure agreements with government agencies and would disclose energy and water data earlier in the development process.[15][93][29] AWS chief executive Matt Garman, in an accompanying essay of more than three thousand words, acknowledged the apprehension in host communities, disputed the attribution of rising rates to datacenters, cited the more than one hundred moratoria under consideration, and warned that local opposition threatened the country’s AI lead.[15]
“We understand why there is apprehension in some communities where data centers are being built.”
— Matt Garman, CEO, Amazon Web Services[15]
The scale must be kept in proportion. The new commitment amounts to roughly $200 million a year against capital expenditure of approximately $220 billion in 2026, about one-tenth of one percent, and it follows more than $1 billion Amazon says it contributed over the preceding three years.[15] But the structural significance lies elsewhere. Together with the $80 million community fund at PORTS-Pike, the community-benefit agreements that Pennsylvania now effectively requires, the local-negotiation tools in Virginia’s framework and the industry alliance’s reported strategy of direct payments to households in target states, Amazon’s program suggests that the industry’s social license is acquiring an explicit economic price.[70][7][32] Community benefits are moving from public relations toward a component of project economics, underwritten in the same financial models as the transformers and the turbines.
5.6 Finance Can Create Inertia, and Impose Discipline
Artificial-intelligence infrastructure is now deeply intertwined with debt markets, and the relationship cuts in two directions that this paper has already introduced and can now state fully.
By early August 2026, AI-related debt issuance had reached nearly $500 billion for the year according to Goldman Sachs Research, with hyperscalers accounting for roughly 40 percent and the balance coming from datacenter developers, neoclouds, chip companies, utilities and project-finance vehicles; Goldman estimated that the three largest hyperscalers could add roughly $2 trillion of debt while remaining investment grade, but that the public bond market could comfortably absorb only about $510 billion of that before investors pushed back.[20][94] By September, lenders were demanding higher yields, signed leases, stronger counterparties and legal backstops against permitting, grid and construction delay; CyrusOne’s $10 billion August financing and Nvidia’s $105 billion PORTS-Pike guaranty are both products of that demand.[21][33] Even the European Central Bank has taken note: one of its governing-council members observed in early October that the volume of AI bond issuance was tightening financial conditions globally.[94]
Debt strengthens inertia once money has been committed, because bondholders and lenders possess remedies that no political actor can override, and because a project in default is more likely to be completed by a successor than abandoned. But credit markets also impose discipline before commitment, because no lender will fund a shell without a power contract, a permit and a creditworthy tenant. The result is a new and widening divide between announced capacity and financeable capacity. The latter is far more important, and it is the latter that Goldman’s forecasts measure. The projects that the 2026 backlash has blocked are, overwhelmingly, projects that had not yet reached financial close; the projects it has not blocked are, overwhelmingly, projects that had. In that sense the capital markets and the state regulators are performing the same function from opposite directions, and the capacity that passes both filters is capacity that has, almost by definition, agreed to pay its own way.
Section 6: November 2026 and the Politics of What Can Still Be Changed
The sections so far have been largely diagnostic. This section is prescriptive, and it begins from the observation that an election is approaching in which the datacenter will be on the ballot in all but name, and that the officials elected in November will inherit a pipeline whose composition is already largely fixed through 2027 and whose composition for 2028 through 2030 is being decided now. The question that should organize their thinking is not whether to support or oppose the buildout, a question that the Commitment Ladder has rendered partly moot, but how to allocate the finite political authority they possess across projects that differ enormously in how far they have climbed. The deep explanation that this section offers is that good datacenter policy is fundamentally a problem of timing: of recognizing that there exists, for every project, a window during which the project is concrete enough to evaluate but not yet so committed that meaningful changes become prohibitively expensive, and of designing institutions that act within that window rather than before or after it.
6.1 Datacenters Enter the Midterm Ballot
The November 3, 2026 election arrives as household affordability, energy prices and AI infrastructure increasingly intersect. The evidence that datacenters have become a national electoral issue is now overwhelming. About six in ten Americans support limiting the number of new datacenters, including majorities of both parties, and a majority describe themselves as extremely or very concerned about the facilities’ effects on electricity prices or water in the communities where they are built.[3] Sixty-four percent of voters say they would be less likely to support a candidate who backed a local datacenter.[2] The president has posted that he never wants Americans to pay higher electricity bills because of datacenters while simultaneously attacking opposition to them; Senators Hawley and Blumenthal have introduced the bipartisan GRID Act to prohibit datacenter-related rate increases and require new facilities to source power off-grid; Senator Heinrich’s GRID Savings Act would require loads above 150 megawatts to pay for their connections; Senator Sanders has called for a national moratorium; and Representative Landsman and others have introduced House measures.[95][82][92] The stalled Senate vote of September 30 demonstrated that the issue is now contested at the level of national party strategy, with each side attempting to position the other as the obstacle to relief.[3][5]
Governors, congressional candidates and state legislators can therefore no longer treat datacenters exclusively as ribbon-cutting opportunities. But the lesson of this paper is that they also cannot treat them as a single undifferentiated threat, because the appropriate response to a project depends almost entirely on where that project sits on the Commitment Ladder.
6.2 Politicians Must Distinguish Projects by Commitment Stage
A useful policy framework would divide projects into four categories corresponding to regions of the Reversibility Curve, and would assign to each a distinct regulatory posture.
| Stage | Description | Characteristic evidence | Appropriate instruments | Inappropriate instruments |
| Stage A — Speculative | Land discussions; preliminary proposals; unverified power requests; duplicate queue positions | Queue entry without deposit; no identified tenant; same load shopped to several utilities | Screening fees; duplicate-application disclosure; queue audits; moratoria where zoning is unprepared | Subsidies; fast-track treatment; utility capital spending in anticipation |
| Stage B — Development | Controlled land; permitting under way; interconnection study; preliminary financing | Zoning application; utility study agreement; community engagement | Local-approval requirements; water assessments; disclosure mandates; community-benefit negotiation; tariff assignment | Blanket bans that ignore project quality; secret incentive agreements |
| Stage C — Committed | Signed power arrangements; anchor tenant; financing closed; major equipment ordered | Executed service agreement; PPA or self-supply; bond issuance; transformer orders | Cost-allocation enforcement; collateral and termination terms; curtailment obligations; phased conditions on later phases | Cancellation; retroactive revocation of approvals |
| Stage D — Irreversible or near-irreversible | Active construction; energized infrastructure; deployed compute | Cranes; delivered GPUs; operating revenue | Rate-class design; ongoing reporting; environmental compliance; renegotiated community terms | Any instrument premised on stopping the project |
The regulatory instrument should differ by stage, because the costs and the counterparties differ by stage. A blanket policy treating all projects identically ignores the economics of infrastructure commitment, and it tends to produce the worst of both worlds: it imposes the political cost of appearing hostile to investment while delivering none of the benefit, because the Stage C and Stage D projects proceed regardless and the Stage A projects were never real. The Texas audit, whatever its other merits, is at bottom an attempt to sort the ERCOT queue into these four categories, and its most valuable output will be a defensible count of how much of the 470 gigawatts is Stage A.[23][51]
6.3 Stop, Slow, Redirect or Renegotiate
Political leaders really have four choices, and the framework above tells them when each is available.
They can stop early-stage projects, and in 2026 they are doing so at a record pace; the hundred and twenty projects worth roughly $200 billion that Data Center Watch recorded as blocked or delayed in the first half of the year are the evidence that this power is real where it is exercised early.[32][30]
They can slow projects while infrastructure consequences are studied, which is the New York and Texas approach; the risk is that a pause imposed without a deadline or a framework becomes a de facto prohibition for the Stage B projects it catches, while the Stage C projects it does not catch proceed to the top of the curve during the pause.
They can redirect projects toward locations with stronger electricity, water or transmission resources, which is the implicit effect of California’s statutes and the explicit goal of the clean-energy incentives in Virginia’s framework; the difficulty is that redirection within a federal system exports load to neighboring states without exporting the political cost, and PJM’s interstate cost-allocation fight is the predictable result.
Or they can renegotiate the economic terms of projects already too advanced to cancel efficiently, which is what Michigan’s pledge, Indiana’s NIPSCO framework, Ohio’s AEP tariff, Pennsylvania’s GRID consent agreements and FERC’s PJM order all represent. The fourth option may become increasingly common, because it is the only one available at Stages C and D, and because the industry, as its trade association’s statements and Amazon’s October announcement both suggest, has concluded that accepting renegotiated terms is cheaper than losing the social license altogether.[92][15]
6.4 Policymakers Need an Infrastructure Intervention Window
The most important policy lesson may be temporal. States need to identify the period during which a project is sufficiently concrete to evaluate but not yet so deeply committed that meaningful changes become prohibitively expensive. Call this the Infrastructure Intervention Window.
Before the window opens, there is typically insufficient information: the developer may not have identified a tenant, the utility may not have completed a study, and the project may be one of several the same developer is shopping. Intervening at this stage, except through generic rules such as screening fees and zoning standards, invites both error and litigation. After the window closes, Buildout Inertia has become too strong: the financing has closed, the equipment has shipped, the labor has mobilized, and the only questions that remain are questions of cost allocation and compliance. Good governance requires intervening between those points, which on the Commitment Ladder means roughly between the interconnection study and the closing of financing, or between Stage B and the beginning of Stage C.
The 2026 reforms can be read as attempts to pry that window open and to keep it open longer. Nondisclosure prohibitions open it by ensuring that the political system learns of the project while it is still at Stage B. Local-approval requirements with low megawatt thresholds, as in Virginia and Pennsylvania, keep it open by inserting a discretionary decision between the study and the financing. Water-assessment requirements, as in California, do the same for a second resource. Public project maps, as in Pennsylvania, allow communities to see the window before it closes. Queue audits, as in Texas, clear the Stage A projects out of the way so that the window can be focused on the projects that are real. Collateral, termination and minimum-demand requirements, as in Michigan, ensure that even when the window closes the public is not left holding the risk. None of these instruments stops the buildout. All of them shift the point at which the public exercises its power from the top of the Reversibility Curve, where it is weak, toward the middle, where it is strong.
6.5 The 2027–2030 Era: Conditional Expansion, Not Unlimited Expansion
The likely outcome is neither an unrestricted datacenter boom nor a nationwide shutdown. It is conditional expansion.
Projects with credible customers, secured power, transparent financing, strong grid arrangements and negotiated community benefits will continue, and the structures that Nvidia, OpenAI and SB Energy built in Ohio, that Amazon built in Indiana, and that six companies accepted in Michigan are templates for what credible will mean. Speculative projects that merely reserve land or occupy enormous interconnection queues will disappear, cleared out by screening fees, audits, lender diligence and the simple exhaustion of the communities they have been asking to host them. Goldman’s downward revision of its 2027 estimate by five gigawatts is the first measurable sign of that clearing, and the firm’s reasoning, that political and community resistance may gradually materialize over the medium and long term even as near-term execution remains intact, is exactly the pattern the Reversibility Curve predicts.[2]
That transition would actually strengthen the most credible parts of America’s AI infrastructure system by removing phantom demand while forcing genuine projects to internalize more of their costs. It would also, not incidentally, improve the quality of the load forecasts on which utilities, grid operators and regulators depend, which is the reform that Koomey, Norris and the Ohio manufacturers have been demanding and which no amount of political enthusiasm for AI could deliver on its own.[46][48]
Section 7: What Have We Learned? Eight Pillars
The preceding sections have moved from a paradox in a bank’s research note through the physics of commitment, the balance sheets of the hyperscalers, the executive orders of five governors, the contracts of four industrial ecosystems, the ratemaking of the largest grid in the country, and the calendar of a national election. It is now possible to state what the exercise has taught. The eight pillars below are offered in the spirit of the earlier papers in the Five-Layer AI Economy series: as lessons that are meant to travel beyond the specific facts of 2026 and to remain useful when the specific facts have changed.
Pillar 1 — Infrastructure Commitments Outlast Political Sentiment
The first and most fundamental lesson is that industrial infrastructure and electoral politics operate on different clocks. Political support can disappear between election cycles, and in 2026 it did: a facility class that governors competed for in 2024 had become a liability that sixty-four percent of voters would hold against a candidate by the autumn of 2026. A transformer order, a twenty-year lease, a nuclear restart agreement or a multibillion-dollar financing cannot disappear so easily, and the parties to those instruments possess rights that voters do not. That mismatch is the foundation of Buildout Inertia, and it explains why Goldman’s forecast could absorb Texas, Virginia, Pennsylvania, California and New York and move by only five gigawatts in each direction.
Pillar 2 — The Central Political Question Is Becoming Cost Allocation
America may increasingly agree that artificial-intelligence infrastructure is strategically important while disagreeing intensely over who should pay for it. The remarkable convergence of 2026, in which a Republican president’s proclamation, a Republican governor’s permit freeze, four Democratic governors’ executive orders and statutes, a bipartisan Senate bill and a federal regulatory order all demanded that large loads pay their own way, indicates that the cost question has displaced the existence question as the center of the debate. The most sustainable projects will therefore be those capable of demonstrating that their private economic gains do not require disproportionate public subsidy through household electricity bills, water systems or local infrastructure. The future political battle is less likely to be AI versus no AI than private benefit versus public cost.
Pillar 3 — States Can Change the Terms Even When They Cannot Easily Change the Direction
Governors Spanberger, Shapiro, Abbott, Newsom, Hochul and Whitmer illustrate six responses to the same structural problem. None of them has reversed a major committed project, and most have not tried. All of them have altered cost allocation, transparency, permitting, water requirements, generation obligations, community compensation, environmental standards or future project eligibility, and several have altered all of these at once. Buildout Inertia limits one form of political power, the power to say no, while increasing the importance of another: the power to rewrite the conditions of growth. The governors who have understood that distinction have been more effective than those who have not, and the industry’s acceptance of their conditions, from Michigan’s pledge to Amazon’s abandonment of nondisclosure agreements, is evidence that the second form of power is real.
Pillar 4 — AI Policy Is Becoming Energy, Utility and Industrial Policy
Artificial intelligence can no longer be governed solely by technology agencies or by debates over algorithms, model weights and safety evaluations. When a frontier model requires gigawatts of electricity, AI policy becomes energy policy, utility regulation, construction policy, industrial policy, water policy, land-use policy, capital-market policy and regional economic strategy, and the institutions that matter most are public utility commissions, grid operators, county planning boards and the Federal Energy Regulatory Commission rather than the agencies that most AI policy discussions have assumed. That is precisely what the Five-Layer AI Economy framework reveals: the intelligence at Layer 4 and Layer 5 ultimately depends upon physical commitments made across Layers 1 through 3, and the politics of those commitments are the politics of electricity and land.
Pillar 5 — The Earlier Government Acts, the More Options Government Retains
Waiting until a project has secured land, power, debt, customers, chips and construction contracts dramatically reduces the range of economically sensible policy choices, and this paper has argued that the reduction is not linear but steep. Government therefore needs rules before the next wave of commitments becomes embedded, which in practice means before the 2027 legislative sessions conclude, because the capacity that Goldman forecasts for the end of 2027 is being financed now and the capacity for 2028 through 2030 will be financed during those sessions. By 2027–2030, the most effective datacenter policy may not be the policy that blocks the most projects. It may be the policy that establishes clear cost, power, water, community and transparency requirements before Buildout Inertia takes hold.
Pillar 6 — Transparency Is the Precondition of Every Other Condition
A lesson that the 2026 record makes unusually clear is that nondisclosure agreements were not an incidental feature of the first-stage politics but their central mechanism. Secrecy was what allowed Infrastructure Commitment Time to run ahead of Political Announcement Time; it was what permitted utilities to build rate cases on forecasts the public could not examine; and it was what prevented regulators, in Peskoe’s phrase, from producing a single clean number for what datacenters were actually paying. The prohibition of such agreements by Virginia, Pennsylvania and, voluntarily, by Amazon, together with the public permitting map in Pennsylvania and the disclosure statutes in California, is therefore not merely one reform among many. It is the reform that makes cost allocation, local consent, water assessment and community negotiation possible, because each of those depends on the public knowing, in time, that a commitment is being made.
Pillar 7 — Phantom Demand and Financeable Demand Must Be Governed Differently
The scholars who doubt the demand forecasts and the forecasters who publish them are, this paper has argued, describing different parts of a single distribution, and the policy error of the first stage was to treat them as a single quantity. Utilities planned, and sought to recover from ratepayers, investments premised on queues that were three to five times larger than the load that would materialize; communities organized against proposals that were never going to be built; and credible projects were lumped together with speculative ones in both the enthusiasm and the backlash. The second-stage reforms that work best, Texas’s audit, screening fees and duplicate-application disclosure, Michigan’s collateral requirements, lenders’ insistence on signed leases, are those that separate the two. Governing phantom demand means clearing it from the queue before it imposes costs. Governing financeable demand means ensuring it pays. Conflating them produces both stranded assets and unnecessary opposition.
Pillar 8 — The Buildout Is Becoming an Arena of Federal–State and Interstate Tension
The final lesson is institutional. The PJM backstop dispute, in which FERC rejected a regional cost allocation and insisted that states play a central role; the Senate’s failure to pass legislation directing FERC to act; California’s export of physical load to neighboring states while retaining the economic returns; Texas’s unilateral freeze of a queue that affects the national supply of compute; and the White House’s attempt to resolve the cost question through a non-binding pledge all point to a system in which authority over the buildout is fragmented across levels of government that have not yet agreed on who decides what. The 2027–2030 cycle will test whether that fragmentation can be resolved through FERC rulemaking, through interstate coordination among governors, or through federal legislation that the 2026 Congress could not pass. Until it is, the buildout will continue to be governed by the actors who happen to control the particular rung of the ladder on which a given project sits, and Buildout Inertia will continue to favor the projects that have climbed highest.
Conclusion: The Politics of Infrastructure Already in Motion
America’s artificial-intelligence infrastructure boom is entering a more complicated phase, and the complication is not that the boom is ending. The first years were dominated by announcements: billions of dollars of investment, enormous GPU orders, hyperscale campuses, semiconductor fabs, new transmission, nuclear restarts and ambitious projections of future compute demand. Governors competed for projects. Communities were promised jobs. Technology companies competed to secure electricity, land and silicon as quickly as possible, frequently under agreements that kept the competition out of public view until it was concluded.
Now the consequences of those decisions are becoming visible, and they are visible in the most politically sensitive place imaginable, which is the monthly bill. Electricity demand is rising. Utility commissions are debating who pays for new infrastructure. The largest grid operator in the country has fallen short of its reliability requirement for the first time in its history and cannot agree with its federal regulator on how to allocate the cost of closing the gap. Governors are imposing safeguards. Communities are organizing in more than eight hundred groups across forty-nine states. Environmental groups are calling for limits. Datacenters have entered the national midterm debate, and a bill addressing them was the last substantive vote the Senate took before leaving to campaign. Texas has halted new approvals while it studies an enormous grid queue. California has imposed additional oversight. Virginia and Pennsylvania are rewriting development rules. New York has paused. Michigan is telling developers that they must protect existing ratepayers, and six of the largest companies in the industry have signed its pledge.
If these developments were examined individually, they might suggest that the American AI infrastructure boom is approaching a political wall.
Goldman’s forecast suggests something more complicated. Even amid this resistance, U.S. datacenter capacity is still expected to rise dramatically, from roughly 64 gigawatts at the end of 2026 toward approximately 90 gigawatts one year later, with power demand growing 38 percent in each year. The reason is not that politics has become irrelevant. The reason is that infrastructure has memory.
Every parcel assembled, interconnection study completed, turbine contracted, transformer ordered, financing agreement signed, GPU allocation reserved, tax incentive approved and construction crew deployed carries an earlier decision forward into the future. One commitment creates another. Layer 1 begins preparing electricity for Layer 3. Layer 2 reserves silicon for the datacenter. Layer 4 commits future models to the resulting compute, for twenty years at a time. Layer 5 begins designing applications and autonomous systems around infrastructure that may not become operational for another two or three years. That interconnected chain creates momentum, and the momentum is measured in the $730 billion to $760 billion that four companies will spend this year, in the $500 billion of debt that was issued in eight months, in the $89 billion of accelerators that one supplier shipped in a single quarter, in the $265 billion committed to a single campus in Arizona and the $105 billion guaranty behind a single campus in Ohio.
It is why Buildout Inertia is the right title for this paper. The word buildout captures the transformation of artificial intelligence from software into an enormous physical industrial system. The word inertia captures what happens once that system has accumulated enough mass, financial, contractual, electrical, political and physical, that changing its direction becomes much harder than changing the rhetoric surrounding it.
But inertia does not mean inevitability, and this paper has been at pains to show the difference. A moving object can still be redirected, and in 2026 it was redirected in ways that would have seemed implausible two years earlier. America’s governors, regulators, utilities, communities and federal policymakers still have substantial authority over how the AI infrastructure boom develops. They can require technology companies to fund the electricity infrastructure they need, and in Indiana, Ohio, Michigan and under the White House pledge the companies have agreed. They can protect residential ratepayers, and FERC has told PJM to do so. They can impose water standards, and California and Texas have. They can demand transparency, and Virginia, Pennsylvania and Amazon itself have ended the secrecy that made the first stage possible. They can direct investment toward regions with stronger grids. They can distinguish credible projects from speculative interconnection requests, and Texas is attempting the largest such exercise ever undertaken. They can negotiate community benefits, and the price of social license is now written into project budgets. And they can establish rules for projects that have not yet crossed the point at which cancellation becomes economically destructive, which is the work of the 2027 legislative sessions and the single most consequential thing the officials elected in November can do.
That distinction should become one of the central lessons for policymakers heading into November 2026 and beyond. Politics is most powerful before infrastructure becomes commitment. Once commitment becomes construction, and construction becomes an interconnected industrial ecosystem spanning energy, chips, datacenters, models and applications, government increasingly governs not whether the AI buildout will exist, but what obligations will accompany it.
That is the paradox at the center of America’s emerging AI geography. The backlash is getting stronger. The rules are getting stricter. The utility bills are becoming political. And yet cranes continue moving, transformers continue being ordered, substations continue being planned, semiconductor capacity continues expanding, financing continues being assembled, and tens of gigawatts of additional compute remain on the horizon for 2027 and beyond.
The infrastructure is already in motion. The work that remains is to decide, before the next wave of commitments closes, what it will owe to the people who live beside it.
That is Buildout Inertia.
Footnotes and Endnotes:
[1] Reuters, “Goldman sees US data center growth intact despite opposition,” October 5, 2026 (via SRN News). https://srnnews.com/goldman-sees-us-data-center-growth-intact-despite-opposition/
[2] Yahoo Finance, “AI data center growth won’t be slowed down by angry neighbors: Goldman Sachs” (quoting Goldman Sachs strategist Laura Cyr), October 5, 2026. https://finance.yahoo.com/technology/article/ai-data-center-growth-wont-be-slowed-down-by-angry-neighbors-goldman-sachs-161404366.html
[3] Associated Press / PBS NewsHour, “Data center bill falls short as Senate holds final votes before the midterms,” September 30–October 1, 2026. https://www.pbs.org/newshour/politics/watch-live-data-center-bill-falls-short-as-senate-holds-final-votes-before-the-midterms
[4] CNN, “Republicans’ last-ditch effort to lower costs ahead of midterms fails in key Senate vote,” September 30, 2026. https://www.cnn.com/2026/09/30/politics/data-centers-affordability-senate
[5] The Hill, “GOP wrestles with rising data center backlash weeks before midterms,” October 1, 2026. https://thehill.com/homenews/senate/6121776-gop-worried-data-center-midterms/
[6] Office of Governor Abigail Spanberger, “Governor Spanberger Unveils Nation’s Most Comprehensive & Aggressive Data Center Accountability Standards, Signs Executive Order to Immediately Put Into Action,” September 18, 2026. https://www.governor.virginia.gov/newsroom/news-releases/2026/september-releases/name-1123696-en.html
[7] Office of Governor Josh Shapiro, “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
[8] Office of the Texas Governor, “Governor Abbott Directs TCEQ To Halt Data Center Permits,” September 21, 2026. https://gov.texas.gov/news/post/governor-abbott-directs-tceq-to-halt-data-center-permits
[9] Houston Public Media, “Gov. Abbott orders TCEQ to pause environmental permits for data centers until audit is complete,” September 21, 2026. https://www.houstonpublicmedia.org/articles/news/energy-environment/2026/09/21/562389/gov-abbott-orders-tceq-to-pause-environmental-permits-for-ai-crytocurrency-data-centers-until-audit-is-complete/
[10] Office of Governor Gavin Newsom, “Governor Newsom signs most comprehensive data center laws in the nation, providing communities more control on water, electricity, and land use,” September 21, 2026. https://www.gov.ca.gov/2026/09/21/governor-newsom-signs-most-comprehensive-data-center-laws-in-the-nation-providing-communities-more-control-on-water-electricity-and-land-use/
[11] WABC-TV, “Gov. Hochul signs executive order pausing new AI data centers in New York,” July 14, 2026. https://abc7ny.com/post/new-york-gov-kathy-hocul-delays-ai-data-centers-energy-environmental-concerns/19505160/
[12] CNBC, “TSMC is accelerating Arizona factory build-out to capitalize on AI ‘megatrend,’ CFO says” (interview with Wendell Huang), July 20, 2026. https://www.cnbc.com/2026/07/20/tsmc-arizona-fab-capacity-ai-chip-demand.html
[13] NVIDIA Corporation, “NVIDIA Guarantees SB Energy’s PORTS-Pike Technology Campus in Ohio to Exclusively Host NVIDIA AI Compute,” press release, August 17, 2026. https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Guarantees-SB-Energys-PORTS-Pike-Technology-Campus-in-Ohio-to-Exclusively-Host-NVIDIA-AI-Compute/default.aspx
[14] Amazon, “Amazon announces $15 billion investment in Northern Indiana data centers,” About Amazon, November 2025. https://aboutamazon.com/news/company-news/amazon-15-billion-indiana-data-centers
[15] GeekWire, “Amazon pledges $1B to data center communities, warns that local opposition threatens U.S. AI lead” (quoting AWS CEO Matt Garman), October 2, 2026. https://www.geekwire.com/2026/amazon-pledges-1b-to-data-center-communities-warns-that-local-opposition-threatens-u-s-ai-lead/
[16] WXYZ / Yahoo News, “6 data center companies sign new pledge after Gov. Whitmer releases plan calling for them to cover their own costs,” July 2026. https://www.yahoo.com/news/politics/articles/6-data-center-companies-sign-205736107.html
[17] Statista, “Big Tech’s AI Spending to Reach $760 Billion in 2026,” July 31, 2026. https://www.statista.com/chart/35046/capital-expenditure-of-meta-alphabet-amazon-and-microsoft/
[18] PJM Interconnection, “PJM Reliability Backstop Proposal Outlines Steps To Secure New Supply and Maintain Reliability,” Inside Lines, July 31, 2026. https://insidelines.pjm.com/pjm-reliability-backstop-proposal-outlines-steps-to-secure-new-supply-and-maintain-reliability/
[19] Utility Dive, “FERC chair slams PJM ‘mess’ as country’s largest grid operator delays power auction,” September 30, 2026. https://www.utilitydive.com/news/pjm-delays-backstop-procurement-ferc-data-center/831751/
[20] Goldman Sachs Research, “How AI Debt Is Reshaping Credit Markets,” Goldman Sachs Exchanges (Amanda Lynam, Allison Nathan), August 2026. https://www.goldmansachs.com/insights/goldman-sachs-exchanges/how-ai-debt-is-reshaping-the-credit-market
[21] Inside AI News, “AI Data-Center Debt Hits $500 Billion as Lenders Demand Tougher Terms,” September 8, 2026. https://insideai.news/news/ai-in-business/ai-data-center-debt/9934/
[22] NVIDIA Corporation, “NVIDIA Announces Financial Results for Second Quarter Fiscal 2027” (quoting Jensen Huang), August 26, 2026. https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027
[23] The National Law Review, “Texas Halts Data Center Permits Pending ERCOT Grid & Water Audits,” September 2026. https://natlawreview.com/article/texas-governor-directs-halt-data-center-permitting-state-agency-audits
[24] Office of Governor Josh Shapiro, “Governor Shapiro’s Remarks as Prepared for Delivery at the Signing Ceremony for an Executive Order on Data Center Development in Pennsylvania,” August 18, 2026. https://www.pa.gov/governor/newsroom/2026-press-releases/gov–shapiro-s-remarks-at-signing-ceremony-for-an-exec–order-on
[25] Blue Ridge Leader, “Spanberger Unveils Sweeping Data Center Accountability Plan for Virginia,” September 2026. https://blueridgeleader.com/spanberger-unveils-sweeping-data-center-accountability-plan-for-virginia/
[26] Utility Dive, “Behind the New York data center pause is legislation that could impact existing facilities,” July 28, 2026. https://www.utilitydive.com/news/behind-the-new-york-data-center-pause-is-legislation-that-could-impact-exis/826369/
[27] The Philadelphia Inquirer, “Gov. Josh Shapiro signs executive order restricting data center development in Pennsylvania,” August 18, 2026. https://www.inquirer.com/politics/pennsylvania/josh-shapiro-data-center-order-20260818.html
[28] Lawyer Monthly, “Gavin Newsom Signs Seven California Data Centre Laws Covering Power, Water and Land Use,” September 2026. https://www.lawyer-monthly.com/2026/09/gavin-newsom-signs-seven-california-data-centre-laws-covering-power-water-and-land-use/
[29] Gadget Review, “Amazon Ends Data Center Secrecy, Pledges $1 Billion to Communities,” October 5, 2026. https://www.gadgetreview.com/amazon-ends-data-center-secrecy-pledges-1-billion-to-communities
[30] Data Center Watch, “Q1 2026: Data Center Watch Report,” June 2026. https://www.datacenterwatch.org/q1-2026
[31] Bloomberg, “New Data Centers Worth $68 Billion Disrupted in US, Data Show,” September 21, 2026. https://www.bloomberg.com/news/articles/2026-09-21/new-data-centers-worth-68-billion-disrupted-in-us-data-show
[32] Trending Topics, “US Opposition Stalls 120 Data Center Projects Worth Nearly €170 Billion” (summarizing Data Center Watch Q2 2026), September 24, 2026. https://www.trendingtopics.eu/us-data-center-opposition-q2-2026/
[33] SB Energy, Inc., Form S-1/A Registration Statement (PORTS-Pike leases and NVIDIA residual value guaranty), U.S. Securities and Exchange Commission, 2026. https://www.sec.gov/Archives/edgar/data/0002133037/000162828026062846/sbenergy-sx1a2.htm
[34] Converge Digest, “AWS Activates Project Rainier Data Center Campus,” 2026. https://convergedigest.com/?p=106770
[35] Data Center Dynamics, “Amazon announces $15bn data center & AI investment plan for Northern Indiana,” November 2025. https://www.datacenterdynamics.com/en/news/amazon-announces-15bn-data-center-ai-investment-plan-for-northern-indiana/
[36] Yield Theory, “Hyperscaler Capex 2026: $720B–$745B Guidance by Company,” September 30, 2026. https://www.yieldtheory.app/research/hyperscaler-ai-capex-tracker-2026
[37] ValueAdd VC, “Meta AI Capex 2026: Guidance Narrows to $130–145B After Q2,” September 2026. https://valueaddvc.com/blog/meta-145b-ai-capex-2026-why-zuckerberg-raised-guidance-twice
[38] NVIDIA Corporation, Form 10-Q for the quarter ended July 26, 2026, U.S. Securities and Exchange Commission. https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000075/nvda-20260726.htm
[39] NVIDIA Corporation, “NVIDIA Announces Financial Results for First Quarter Fiscal 2027,” Form 8-K, May 2026. https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000051/q1fy27pr.htm
[40] Yahoo Finance, “AI-Related Debt Sales Are Near $500 Billion. Meta and CoreWeave Show Why the Same Boom Creates Two Different Risks” (citing Reuters Breakingviews), September 2026. https://finance.yahoo.com/markets/stocks/articles/ai-related-debt-sales-near-190243211.html
[41] U.S. Department of Energy / Lawrence Berkeley National Laboratory, “DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers” (2024 Report on U.S. Data Center Energy Use, Arman Shehabi et al.), December 2024. https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers
[42] Data Center Dynamics, “DOE: Data centers consumed 4.4% of US power in 2023, could hit 12% by 2028,” December 2024. https://www.datacenterdynamics.com/en/news/doe-data-centers-consumed-44-of-us-power-in-2023-could-hit-12-by-2028
[43] International Energy Agency, “AI is set to drive surging electricity demand from data centres while offering the potential to transform how the energy sector works” (Energy and AI report; quoting Fatih Birol), April 10, 2025. https://www.iea.org/news/ai-is-set-to-drive-surging-electricity-demand-from-data-centres-while-offering-the-potential-to-transform-how-the-energy-sector-works
[44] Christian Bogmans, Patricia Gomez-Gonzalez, Giovanni Melina and Sneha Thube, “AI Needs More Abundant Power Supplies to Keep Driving Economic Growth,” IMF Blog, International Monetary Fund, May 13, 2025. https://www.imf.org/en/blogs/articles/2025/05/13/ai-needs-more-abundant-power-supplies-to-keep-driving-economic-growth
[45] DayOne Data Centers Ltd., Form F-1 Registration Statement (citing Structure Research), U.S. Securities and Exchange Commission, 2026. https://www.sec.gov/Archives/edgar/data/0002118192/000119312526414188/d223752df1.htm
[46] Ohio Manufacturers’ Association, “OMA Says ‘Enough’ on Data Center Forecasts Nobody Can Prove” (Jonathan Koomey keynote, Ohio Manufacturers’ Energy Conference), August 27, 2026. https://www.ohiomfg.com/wp-content/uploads/2026/08/OMA-Says-Enough-on-Data-Center-Forecasts-Nobody-Can-Prove.pdf
[47] Ohio Manufacturers’ Association, “Koomey Tests Forecasts, OMA Puts Regulators on Watch,” August 28, 2026. https://www.ohiomfg.com/our-communities/koomey-tests-forecasts-oma-puts-regulators-on-watch/
[48] Tyler H. Norris, Tim Profeta, Dalia Patiño-Echeverri and Adam Cowie-Haskell, “Rethinking Load Growth: Assessing the Potential for Integration of Large Flexible Loads in US Power Systems,” Nicholas Institute for Energy, Environment & Sustainability, Duke University, February 2025. https://nicholasinstitute.duke.edu/publications/rethinking-load-growth
[49] Latitude Media, “The US grid may have over 100 GW of load to spare” (interview with Tyler Norris), February 2025. https://www.latitudemedia.com/news/the-us-grid-may-have-over-100-gw-of-load-to-spare/
[50] Nicholas Institute, Duke University, “Three Key Takeaways: Rethinking Load Growth in U.S. Power Systems” (quoting Tyler Norris), February 11, 2025. https://nicholasinstitute.duke.edu/articles/three-key-takeaways-rethinking-load-growth-us-power-systems
[51] Dave Friedman, “Inside Texas’s AI Data Center Queue,” Buy the Rumor; Sell the News (Substack), 2025–2026. https://davefriedman.substack.com/p/inside-texass-ai-data-center-queue
[52] Daron Acemoglu, “What Do We Know About the Economics of AI?” (summarizing “The Simple Macroeconomics of AI”), MIT Department of Economics, 2024; see also BBC World Business Report interview, February 5, 2026. https://economics.mit.edu/news/daron-acemoglu-what-do-we-know-about-economics-ai
[53] Federal Reserve Bank of Richmond, “Will AI Investments Pay Off?” Econ Focus, First/Second Quarter 2026 (quoting Erik Brynjolfsson). https://www.richmondfed.org/publications/research/econ_focus/2026/q1-q2_feature2
[54] The Decoder, “Stanford’s Brynjolfsson sees AI boosting US productivity, but he also co-founded an AI consulting firm,” 2026. https://the-decoder.com/stanfords-brynjolfsson-sees-ai-boosting-us-productivity-but-he-also-co-founded-an-ai-consulting-firm/
[55] Consumer Reports, “AI Data Centers: Big Tech’s Impact on Electric Bills, Water, and More,” March 2026. https://www.consumerreports.org/data-centers/ai-data-centers-impact-on-electric-bills-water-and-more-a1040338678/
[56] USA Today (via AOL), “Surging electricity rates put data centers on 2026 ballot,” 2026. https://www.aol.com/articles/surging-electricity-rates-put-data-100820887.html
[57] Bay Journal, “Virginia governor creates data center accountability policy,” September 2026. https://www.bayjournal.com/news/growth_conservation/virginia-governor-creates-data-center-accountability-policy/article_e40a72da-53b4-4cdd-b995-f20210676ae5.html
[58] Office of Governor Abigail Spanberger, “Virginia’s Data Center Accountability Framework” (PDF), September 2026. https://www.governor.virginia.gov/media/governorvirginiagov/governor-of-virginia/pdf/Data-Center-Accountability-Framework.pdf
[59] NBC News, “Gov. Josh Shapiro takes a hard line against ‘predatory’ data center developers in Pennsylvania,” August 18, 2026. https://www.nbcnews.com/politics/2028-election/gov-josh-shapiro-executive-order-data-centers-pennsylvania-rcna593177
[60] Greenberg Traurig LLP, “Gov. Shapiro Issues Executive Order 2026-05 Imposing New Requirements on Pennsylvania Data Center Projects,” August 2026. https://www.gtlaw.com/en/insights/2026/8/gov-shapiro-issues-executive-order-2026-05-imposing-new-requirements-on-pennsylvania-data-center-projects
[61] Mike Helbing, “Data center projects in Pennsylvania face stronger local scrutiny after Shapiro’s executive order,” Penn State Institute of Energy and the Environment, September 1, 2026. https://iee.psu.edu/news/data-center-projects-pennsylvania-face-stronger-local-scrutiny-after-shapiros-executive-order
[62] Briefs.co, “Abbott halts Texas data center permits until ERCOT waitlist audit wraps,” September 2026. https://www.briefs.co/news/abbott-halts-texas-data-center-permits-until-ercot-waitlist/
[63] Aterio, “TSMC’s $265 Billion Arizona Gigafab: Three Fabs You Can See, Seven on Paper,” July 2026. https://www.aterio.io/blog/tsmc-phoenix-gigafab-satellite-construction-arizona
[64] The National Law Review, “California Governor Signs Bills Regulating Data Center Industry,” September 2026. https://natlawreview.com/article/california-signs-sweeping-data-center-energy-laws
[65] Reuters (via U.S. News & World Report), “California Governor Signs Broad Data Center Oversight Bill Package,” September 21, 2026. https://www.usnews.com/news/politics/articles/2026-09-21/california-governor-signs-broad-data-center-oversight-bill-package
[66] Foley & Lardner LLP, “New York Just Pressed Pause on Large Data Center Permitting,” August 2026. https://www.foley.com/insights/publications/2026/08/new-york-just-pressed-pause-on-large-data-center-permitting/
[67] U.S. PIRG, “Where have Americans hit pause on new data centers?” August 10, 2026. https://pirg.org/articles/where-have-americans-hit-pause-on-new-data-centers/
[68] City of Phoenix, “TSMC Announces Additional $100 Billion Investment In Arizona,” July 16, 2026. https://www.phoenix.gov/newsroom/ced-news/tsmc-announces-additional–100-billion-investment-in-arizona.html
[69] AZ Big Media, “TSMC will invest an additional $100 billion into Arizona operations” (quoting C.C. Wei), August 27, 2026. https://azbigmedia.com/business/tsmc-will-invest-an-additional-100-billion-into-arizona-operations/
[70] OpenAI, “OpenAI joins PORTS-Pike project,” August 2026. https://openai.com/index/openai-joins-ports-pike-project/
[71] Data Center Frontier, “PORTS-Pike Takes Shape as an 8-GW AI Infrastructure Model,” August 2026. https://www.datacenterfrontier.com/hyperscale/article/55398883/ports-pike-takes-shape-as-an-8-gw-ai-infrastructure-model
[72] Yahoo Finance, “Nvidia Backs 8-GW Ohio AI Campus for OpenAI With $1.5 Billion Investment,” August 2026. https://finance.yahoo.com/technology/ai/articles/nvidia-backs-8-gw-ohio-013000089.html
[73] Office of Governor Gretchen Whitmer, “Gov. Whitmer Launches Michigan Affordable and Responsible Growth Action Plan, Calls on Data Center Companies to Sign Pledge,” July 15, 2026. https://www.michigan.gov/whitmer/news/press-releases/2026/07/15/gov-whitmer-data
[74] E&E News by POLITICO, “Whitmer pushes data center pledge — not a moratorium,” July 15, 2026. https://www.eenews.net/articles/whitmer-pushes-data-center-pledge-not-a-moratorium/
[75] Daily Energy Insider, “Michigan governor unveils action plan to make data centers pay their own way” (quoting Gov. Whitmer), July 2026. https://dailyenergyinsider.com/news/53068-michigan-governor-unveils-action-plan-to-make-data-centers-pay-their-own-way/
[76] Utility Dive, “Palisades becomes first decommissioned US nuclear plant to reach ‘operations’ status,” August 2025. https://www.utilitydive.com/news/palisades-nuclear-plant-holtec-nrc-operations/758845/
[77] Canary Media, “America’s first nuclear plant restart may be near the finish line,” July 2026. https://www.canarymedia.com/articles/nuclear/americas-first-nuclear-plant-restart
[78] Nuclear Engineering International, “Palisades restart delayed,” September 28, 2026. https://www.neimagazine.com/news/palisades-restart-delayed/
[79] PolitiFact, “How much have data centers increased electricity prices?” (quoting Kenneth Gillingham, Yale, and Ari Peskoe, Harvard), June 12, 2026. https://politifact.com/factchecks/2026/jun/12/elizabeth-warren/data-centers-rising-electricity-costs/
[80] Carnegie Mellon University, “Data Center Growth Could Increase Electricity Bills 8% Nationally and as Much as 25% in Some Regional Markets” (Open Energy Outlook Initiative, CMU and NC State), June 2025. https://www.cmu.edu/work-that-matters/energy-innovation/data-center-growth-could-increase-electricity-bills-8
[81] NPR, “AI data centers: How much are ratepayers on the hook for?” (quoting Ari Peskoe, Harvard; Lucy Qiu, University of Maryland; PJM Independent Market Monitor), October 1, 2026. https://www.npr.org/2026/10/01/nx-s1-5984949/data-centers-ai-ratepayers-congress
[82] Newsweek, “Congress Targets Data Centers: What It Could Mean for Home Prices, Bills” (quoting Hannah Wiseman, Penn State Dickinson Law), September 29, 2026. https://www.newsweek.com/congress-targets-data-centers-what-mean-home-prices-bills-12500410
[83] EPRI and Watershed researchers, “Have Data Centers Raised Your Electric Bill? Causal Evidence from the United States,” arXiv working paper 2606.19777, June 2026. https://arxiv.org/html/2606.19777v1
[84] Eliza Martin and Ari Peskoe, “Extracting Profits from the Public: How Utility Ratepayers Are Paying for Big Tech’s Power,” Harvard Electricity Law Initiative, Harvard Law School, March 2025. https://eelp.law.harvard.edu/extracting-profits-from-the-public-how-utility-ratepayers-are-paying-for-big-techs-power/
[85] Utility Dive, “Utilities may subsidize data center growth by shifting costs to ratepayers: Harvard report” (quoting Martin and Peskoe), March 2025. https://www.utilitydive.com/news/utilities-subsidize-data-center-growth-ratepayer-cost-shif/742001/
[86] Michigan Advance, “Power for data centers could come at ‘staggering’ cost to consumers” (quoting Ari Peskoe), March 6, 2025. https://michiganadvance.com/2025/03/06/power-for-data-centers-could-come-at-staggering-cost-to-consumers/
[87] Utility Dive, “PJM files backstop auction plan at FERC to meet capacity shortfall,” August 2026. https://www.utilitydive.com/news/pjm-backstop-capacity-auction-ferc-data-centers/826792/
[88] Inside the Datacenter, “The Backstop That Stalled: PJM’s One-Time Auction for Data Center Power Runs Into FERC,” October 2026. https://insidethedatacenter.com/articles/pjm-backstop-auction/
[89] Daily Caller News Foundation, “Feds Freeze Data Center Power Plan For Grid Serving 67 Million — Over Who Pays” (quoting Joseph Bowring, PJM Independent Market Monitor), October 5, 2026. https://dailycaller.com/2026/10/05/ferc-pjm-backstop-procurement-data-centers-cost-allocation-maryland-electric-bills/
[90] The White House, “Proclamation 11014 of March 4, 2026 — Ratepayer Protection Pledge,” Federal Register, Vol. 91, No. 45, March 9, 2026. https://www.federalregister.gov/documents/full_text/html/2026/03/09/2026-04645.html
[91] Straight Arrow News, “Utilities are now pitching data centers as a way to cut electric costs,” August 9, 2026. https://san.com/cc/utilities-are-now-pitching-data-centers-as-a-way-to-cut-electric-costs/
[92] Politico / E&E News (via Yahoo), “Trump wants to move data centers — Congress is far from settling on a strategy” (quoting Dan Diorio, Data Center Coalition), 2026. https://www.yahoo.com/news/articles/trump-wants-move-data-centers-174716512.html
[93] HostingJournalist, “AWS Pledges Over $1 Billion to Support U.S. Data Center Communities,” October 3, 2026. https://hostingjournalist.com/news/aws-pledges-over-1-billion-to-support-u-s-data-center-communities
[94] R&D World, “Big tech turns to debt to fund the AI buildout,” October 3, 2026. https://www.rdworldonline.com/big-tech-turns-to-debt-to-fund-the-ai-buildout/
[95] NBC News, “Senators introduce first bipartisan effort to curb utility bill hikes related to data centers” (GRID Act, Sens. Hawley and Blumenthal), 2026. https://www.nbcnews.com/politics/congress/senators-introduce-first-bipartisan-effort-curb-utility-bill-hikes-rel-rcna258577
