Introduction: The Second Transformer
In late June 2026, a gathering in South Boston, Virginia, marked the beginning of something much larger than another American factory.
The location was far removed from Silicon Valley’s research laboratories, Seattle’s cloud campuses, Manhattan’s investment banks, and Washington’s debates over artificial-intelligence regulation. There were no humanoid robots walking across a demonstration stage, no new reasoning model answering difficult scientific questions, and no rows of Nvidia accelerators flashing beneath liquid-cooled server racks. Instead, the center of attention was an enormous industrial machine whose basic operating principle has been understood for more than a century: the electrical transformer.
On June 29, 2026, Hitachi Energy broke ground on what it describes as the largest facility in the United States dedicated to producing large power transformers. The planned $457 million expansion of its South Boston campus—a site that has built transformers since 1968, when it began life as a Westinghouse plant—is expected to create 825 new jobs and nearly double a workforce of roughly 850 people, forming the cornerstone of Hitachi Energy’s broader commitment of more than $1 billion to U.S. grid-equipment manufacturing.[1] Virginia’s governor, a United States senator, and a member of Congress stood beside company executives at the ceremony, a tableau that would once have been reserved for a semiconductor fabrication plant or an automobile factory. The announcement was not presented as an artificial-intelligence breakthrough. Yet the factory may ultimately prove as important to America’s AI ambitions as many celebrated investments in chip fabrication or frontier-model development.
“Power transformers are not a niche product.”
— Governor Abigail Spanberger of Virginia, at the South Boston groundbreaking [2]
“…built by our skilled Virginia workforce, helping power communities across the country.”
— Greg Callahan, Senior Vice President, Transformers North America, Hitachi Energy [3]
The ceremony represented a collision between two very different technological clocks.
The first clock is the clock of artificial intelligence. It moves in weeks and months. A model is released, evaluated, improved, replaced, and sometimes forgotten before a traditional power project has completed its permitting review. AI companies announce new accelerator clusters, reasoning systems, autonomous agents, robotic platforms, and multibillion-dollar data-center campuses with extraordinary speed. Investors respond almost immediately. Markets reprice companies overnight. A product that did not exist at the beginning of the year can become the foundation of an industry before the year is over.
The second clock is the clock of electrical infrastructure. It moves in years.
A large power transformer must be designed for a particular voltage, capacity, utility specification, physical location, and operating environment. Its core must be formed from specialized electrical steel. Its conductors must be wound with great precision. Its insulation must be dried and protected. Its components must be assembled, tested, certified, transported, installed, and integrated into a substation whose other equipment may also be delayed. The completed machine can weigh hundreds of tons and may require specialized railcars, reinforced bridges, police escorts, route studies, cranes, and carefully coordinated installation teams. The artificial-intelligence industry speaks about scaling in units of compute. The electrical-equipment industry must scale through steel, copper, factories, test bays, engineering expertise, heavy transportation, and skilled human labor. That distinction is becoming one of the most consequential—and least appreciated—features of the global AI economy.
Consider a hypothetical AI campus planned outside a fast-growing American city. The developer has secured land. The state has approved a generous package of tax incentives. A cloud provider has committed to become the principal tenant. The semiconductor order has been placed, and thousands of advanced accelerators have been allocated. Fiber routes have been mapped. A utility has agreed, in principle, to deliver several hundred megawatts of power. The public announcement describes jobs, investment, national competitiveness, and technological leadership. The governor calls the project evidence that the state is winning the future. Local officials imagine new tax revenue. Construction begins on the first buildings.
Then the project team confronts the second transformer.
The required substation equipment cannot be delivered on the original schedule. The large transformer has a manufacturing lead time measured not in months but in years. High-voltage circuit breakers are also delayed. Switchgear availability has tightened. The utility must decide whether to reserve equipment before the interconnection study is complete. The developer must determine whether to pay a large deposit to secure a manufacturing position for a facility that may still encounter regulatory or community opposition. Alternative suppliers offer earlier delivery, but some are located in jurisdictions that raise national-security, cybersecurity, trade, or political concerns. The campus has land but not electricity. It has chips but not conversion capacity. It possesses computational ambition without electrical deliverability.
This is no longer an unusual scenario. Artificial-intelligence data centers are intensifying an equipment shortage that began with pandemic-era disruptions, aging infrastructure, electrification, manufacturing growth, extreme weather, and rising utility demand. By July 2026, Reuters reported that lead times for generator step-up transformers had surpassed 160 weeks by the first quarter of the year, compared with an average of 143 weeks in 2024, while high-voltage circuit-breaker lead times had climbed to 125 weeks in the second half of 2025, versus 77 weeks in 2023. Utilities were purchasing components three to five years ahead, refurbishing aging units, requesting advance payments from large customers, and diversifying sourcing across multiple countries. Transformer prices were expected to rise a further four to ten percent within a year.[5] Long-term supply agreements offer partial relief, but as the National Rural Electric Cooperative Association’s senior vice president for government relations observed:
“They don’t solve everything, particularly for smaller utilities that don’t have the scale.”
— Louis Finkel, National Rural Electric Cooperative Association [5]
The U.S. Department of Energy has described the magnitude of the problem in equally stark terms. Demand for distribution transformers increased by roughly 41 percent between 2019 and the most recent available data. Order times that had once ranged from three to six months expanded to one or two years—and sometimes longer. Large transformers required for substations and power plants could take three to four years. The country also operates with tens of thousands of distinct transformer configurations, limiting interchangeability and making it harder for utilities to share critical equipment during emergencies.[9]
These delays reveal an important weakness in the dominant narrative surrounding artificial intelligence. The AI race is commonly portrayed as a competition for the most advanced semiconductor, the largest model, the greatest concentration of engineering talent, or the deepest pool of investment capital. Within that narrative, Nvidia, AMD, TSMC, Google, Meta, Amazon, Microsoft, OpenAI, Anthropic, xAI, and other technology companies occupy the central positions. The principal strategic objects are GPUs, high-bandwidth memory, advanced packaging systems, model weights, proprietary datasets, cloud contracts, and algorithmic expertise. All of these assets matter. But none can operate at industrial scale without a functioning electrical system beneath them.
An accelerator does not consume electricity directly from a distant nuclear reactor, gas plant, solar farm, wind project, or hydroelectric facility. Between generation and computation stands an extensive architecture of transmission lines, substations, transformers, breakers, conductors, power electronics, backup systems, and distribution equipment. Electricity must be generated, transported, converted, controlled, stabilized, and delivered at the precise voltages required by the data center and ultimately by the server rack. Even the internal power architecture of the AI data center is beginning to change: the rapid growth of accelerator density is forcing engineers to reconsider traditional power-delivery systems and investigate medium-voltage distribution, direct-current architectures, advanced conversion technologies, and solid-state transformers.[18] The transformer problem therefore exists both outside the campus, where utilities must connect enormous new loads, and inside it, where unprecedented computational density is reshaping the movement of electricity from the substation to the processor.
Artificial intelligence is consequently built upon two transformers. The first is the mathematical transformer: the model architecture that identifies relationships across words, images, sounds, biological structures, software code, and other forms of data. It converts sequences into representations, representations into predictions, and predictions into increasingly sophisticated forms of machine-generated reasoning. The second is the electrical transformer: the physical machine that converts voltage, enables electricity to move across long distances, connects generation to transmission, links transmission to distribution, and makes power usable by factories, buildings, data centers, and individual machines.
The first transformer converts tokens into intelligence. The second converts voltage into usable power. The first receives global attention. The second determines whether the first can operate.
This relationship is particularly clear within the Five-Layer AI Economy. Energy forms the first layer, followed by chips, data centers, models, and applications or agents. The layers are often discussed as though technological and economic value rises steadily upward, from physical infrastructure toward increasingly abstract forms of intelligence. Yet the transformer shortage demonstrates that the lowest layer can impose an absolute limit on every layer above it. A shortage of models can be addressed through research. A shortage of applications may generate entrepreneurial opportunity. A shortage of accelerators may encourage new semiconductor entrants, alternative architectures, or more efficient inference. But a missing large power transformer can leave billions of dollars of completed or partially completed infrastructure unable to connect to the grid. Software cannot download a replacement. Financial capital cannot instantly manufacture one. A political announcement cannot compress years of specialized industrial production into a few weeks.
The constraint is not merely technical. It is geopolitical. Large transformers and their components are produced through international networks involving the United States, China, South Korea, Japan, Mexico, Canada, and European industrial economies. Domestic factories may rely on foreign electrical steel, bushings, conductors, cooling systems, insulation materials, manufacturing tools, or specialized expertise. Utilities confronting long domestic lead times may turn to overseas manufacturers. In some cases, foreign suppliers can offer lower prices or earlier delivery. Those commercial advantages must then be evaluated against cybersecurity risks, trade tensions, domestic-content rules, transportation challenges, and the possibility that a future conflict could interrupt replacement parts or technical support.
The United States government has begun to recognize this vulnerability as a national-security issue. On April 20, 2026, the White House issued Presidential Determination No. 2026-10 under Section 303 of the Defense Production Act of 1950, finding that America’s capacity to design, produce, and deploy large-scale grid infrastructure—including transformers, high-voltage transmission components, advanced conductors, power electronics, substations, and grid-supporting manufacturing equipment—is “dangerously limited.” The determination declared grid infrastructure and its upstream supply chains, expressly naming high-voltage circuit breakers, protective relay systems, capacitor banks, electrical core steel, and related raw materials and manufacturing tools, to be essential to national defense, and it authorized purchases, purchase commitments, and financial support to expand domestic production.[6][7] Legal analysts immediately observed that the determinations could reshape financing, procurement, and industrial strategy across the American energy sector.[8]
That action marked a conceptual shift. Grid equipment was no longer simply a utility-procurement concern. It had become part of the national industrial base.
This paper calls the emerging system Transformer Diplomacy. Transformer Diplomacy is the strategic use of manufacturing capacity, critical materials, procurement finance, technical standards, foreign investment, trade policy, industrial alliances, and emergency reserves to secure the grid hardware required for artificial intelligence, semiconductor fabrication, advanced manufacturing, military readiness, and national economic growth.
The word diplomacy is essential because no major economy can solve the problem through domestic production alone—not quickly, and perhaps not completely. The United States needs new factories, but it also needs trusted foreign suppliers. It needs domestic electrical-steel production, but it may also need allied material agreements. It needs national-security screening, but it cannot treat every imported component as equally dangerous. It needs standardized equipment that can be exchanged during emergencies, but utilities have historically relied on thousands of customized specifications. It must accelerate data-center construction while ensuring that wealthy technology companies do not reserve so much manufacturing capacity that smaller utilities, rural cooperatives, housing projects, hospitals, and public infrastructure are pushed to the back of the queue.
Transformer Diplomacy is therefore not synonymous with transformer manufacturing. It is the political economy of electrical conversion. It asks which countries control the factories. It asks who produces the steel, copper, insulation, breakers, bushings, electronics, and manufacturing tools. It asks which buyers can reserve production years in advance and which communities must wait. It asks whether a transformer manufactured abroad can be trusted, repaired, monitored, and replaced during an international crisis. It asks whether allies should establish shared inventories and mutual-assistance agreements. It asks whether the government should prioritize certain projects when equipment is scarce. And it asks whether the physical infrastructure of artificial intelligence should be governed with the same strategic seriousness already applied to advanced semiconductors.
The paper develops these questions through Five Circuits of Transformer Diplomacy: Production Capacity, Material Security, Procurement Power, Standards Power, and Alliance Capacity. Together, these circuits explain why transformer availability is not determined by a single factory or material. It emerges from the interaction of industrial capability, raw-material access, corporate purchasing power, engineering rules, government policy, and international relationships.
The central argument is straightforward: the countries that lead the next phase of artificial intelligence will not simply be those that design the best models or purchase the most advanced chips. They will be those capable of converting political ambition, financial capital, industrial materials, and electrical power into reliable computational capacity.
The new factory in South Boston represents one answer to that challenge. It promises additional production, skilled employment, and a stronger domestic industrial base, and it belongs to a broader wave in which original-equipment manufacturers have committed roughly $1.8 billion to North American capacity expansions since 2023, capped by GE Vernova’s completed acquisition of Prolec GE in February 2026 to consolidate regional production.[4] Yet one factory cannot resolve decades of underinvestment, thousands of incompatible equipment specifications, foreign-component dependence, limited workforce capacity, transportation bottlenecks, or the accelerating competition among utilities, hyperscalers, semiconductor plants, manufacturers, housing developments, and public institutions. The significance of the factory lies not in the claim that it will solve the transformer shortage. Its significance lies in what its construction admits.
The artificial-intelligence race has moved beyond the semiconductor fab and the data-center campus. It has reached the substation, the steel mill, the winding line, the testing bay, the rail network, and the utility procurement office. The future of intelligence now depends upon the machinery of electricity. And the second transformer is no longer invisible.
Section 1: From GPU Scarcity to Grid-Hardware Scarcity
For the first several years of the generative artificial-intelligence boom, scarcity had a recognizable shape. It was a graphics processing unit mounted inside a server.
Technology companies measured strategic advantage by the number of advanced accelerators they could acquire, install, and connect. Frontier-model developers competed for Nvidia systems. Cloud providers reserved semiconductor production, advanced packaging capacity, high-bandwidth memory, networking equipment, and entire manufacturing runs. Startups described their financial needs in terms of GPU access. Governments designed export controls around processor performance and interconnection speed. Investors attempted to determine which companies possessed enough compute to train the next generation of models. The accelerator became both a machine and a symbol. It represented computational capacity, national technological power, corporate ambition, and access to the future.
That framing was understandable. A frontier model cannot be trained without enormous amounts of computation, and advanced accelerators remain among the most sophisticated products ever manufactured. Yet the industry’s focus on the chip created a misleading impression: that once accelerators became available, artificial-intelligence capacity could be created simply by installing more of them. The physical reality is more complicated. A GPU is not an independent unit of intelligence. It is one component inside a rack-scale computational system. That rack is one component inside a data hall. The data hall belongs to a campus containing cooling plants, electrical rooms, backup systems, substations, transmission connections, water infrastructure, fiber routes, security systems, and industrial control equipment. The campus, in turn, depends upon power plants, regional transmission networks, fuel supplies, utility planning, equipment manufacturers, construction labor, and government approvals.
The bottleneck therefore does not disappear when the semiconductor arrives. It moves. As accelerator supply expands and AI systems grow more power-intensive, the limiting factor migrates downward—from chips to racks, from racks to buildings, from buildings to substations, and from substations into the wider electrical grid. This is the transition from GPU scarcity to grid-hardware scarcity, and it has become the organizing insight of a growing academic and policy literature published between 2020 and 2026.
“The scarcest resource in AI isn’t chips or talent — it’s grid capacity.”
— Santiago Gallino, Professor of Operations, Information and Decisions, The Wharton School, University of Pennsylvania [35]
1.1 The Bottleneck Beneath the Chip
The newest AI systems are no longer collections of conventional servers arranged inside conventional data centers. They are increasingly designed as integrated computational machines whose processors, memory, networking, cooling, and power systems operate at rack or cluster scale. Nvidia’s Blackwell generation helped establish the rack as a primary unit of AI-system integration. Its succeeding Vera Rubin platform extends that approach through multiple purpose-built rack-scale systems intended for large agentic, reasoning, inference, and simulation workloads.[16] Nvidia has also begun preparing power architectures capable of supporting information-technology racks approaching or exceeding one megawatt—orders of magnitude denser than the racks installed in traditional enterprise data centers.
That increase in computational density changes the infrastructure surrounding the processor. A traditional data center could distribute electricity through architectures designed for kilowatt-scale racks. A future AI factory may need to deliver vastly greater amounts of power into a much smaller physical footprint. Higher densities increase the importance of power conversion, voltage management, liquid cooling, redundant distribution, fault protection, and the ability to move electricity efficiently from the utility connection to the individual processor. Nvidia’s proposed 800-volt direct-current architecture illustrates the direction of travel: the company argues that conventional low-voltage rack designs become increasingly difficult to scale as systems move toward one-megawatt racks, partly because of the volume of copper, conversion equipment, and physical space required, and that higher-voltage direct-current architectures can reduce conversion losses and enable much denser AI systems beginning later in the decade.[17] The 2026 engineering literature on data-center power delivery reaches similar conclusions, identifying medium-voltage distribution and solid-state transformers as candidate technologies for the AI factory’s internal grid.[18]
The result is an apparent paradox. Semiconductor innovation can increase the amount of computation performed per unit of energy. At the same time, the commercial success of those systems can cause companies to deploy so many processors that total electricity consumption continues to rise. Efficiency improves at the chip or workload level, while aggregate power demand expands at the campus, regional, and national levels. The International Energy Agency’s 2026 analysis captures this dynamic precisely: power consumption per AI task is declining at a rate the agency describes as unprecedented in energy history, and yet total data-center electricity consumption is still set to double by 2030, with consumption at AI-focused facilities poised to triple.[14]
A single large data center can consequently resemble an industrial city more than a conventional office complex. The Electric Power Research Institute estimates that a new facility with demand between 100 and 1,000 megawatts requires electricity comparable to approximately 80,000 to 800,000 average homes, and it emphasizes the mismatch between development timelines: data centers may be proposed and constructed within a few years, while the transmission, generation, and utility infrastructure required to serve them often takes much longer to plan, permit, finance, manufacture, and build.[15] Lawrence Berkeley National Laboratory’s congressionally mandated assessment for the Department of Energy found that data centers consumed about 4.4 percent of total U.S. electricity in 2023 and projected consumption of between 6.7 and 12 percent of total U.S. electricity by 2028, with total usage climbing from 176 terawatt-hours in 2023 toward 325 to 580 terawatt-hours.[10][11]
“These are the largest single points of consumption of electricity in history.”
— Jesse Jenkins, Associate Professor of Energy Systems Engineering, Princeton University [36]
The chip-development cycle and the electrical-development cycle have become structurally misaligned. AI companies plan according to product roadmaps. Utilities plan according to load forecasts, rate cases, interconnection studies, equipment availability, transmission approvals, and reliability obligations. A semiconductor architecture may be replaced in two years. A transmission asset may remain in service for half a century. A model developer may revise its compute forecast several times within a quarter. A transformer manufacturer must reserve materials, factory space, labor, testing capacity, and transportation years before a completed machine is energized. The faster the upper layers of the AI economy move, the more pressure they place upon the slower layers below them.
1.2 The Electrical Chain of Artificial Intelligence
To understand where scarcity now binds, it helps to walk the electrical chain from the power plant to the processor, because each link corresponds to a distinct class of industrial equipment with its own manufacturers, materials, and lead times. Generator step-up transformers stand at the exit of every power plant, raising generation voltage to transmission voltage; they are among the most severely delayed items in the entire chain, with lead times exceeding 160 weeks by early 2026.[5] Large power transformers move electricity between transmission voltages and form the backbone of the bulk grid; the Department of Energy estimates delivery in three to four years for many units.[9] Substation transformers step transmission voltage down for regional delivery and are the machines that data-center campuses most often wait upon. Distribution transformers perform the final conversion for neighborhoods, commercial buildings, and individual facilities; demand for them rose roughly 41 percent from 2019 while order times stretched from months to years.[9] Switchgear provides the controlled switching, protection, and isolation that make high-power circuits safe and operable; manufacturers such as GE Vernova are expanding air-insulated switchgear output from 9,000 to 10,500 units this year and still cannot keep pace.[24] High-voltage circuit breakers interrupt fault currents at transmission scale, and their lead times climbed to 125 weeks in late 2025 from 77 weeks in 2023.[5] Finally, turbines and on-site generation have become part of the chain by necessity, as developers unable to obtain timely grid connections attempt to bring their own power—only to discover that heavy gas turbines are contracted years ahead, with GE Vernova alone holding 116 gigawatts of gas-turbine backlog and slot-reservation agreements as of mid-2026.[23]
| Equipment Class | Function in the AI Electrical Chain | Lead-Time Signal (2023 → 2026) |
| Generator step-up transformers | Raise power-plant output to transmission voltage | Avg. 143 weeks (2024) → 160+ weeks (Q1 2026) [5] |
| Large power transformers | Bulk transmission backbone; substation interties | ~1 year historically → 3–4 years (DOE) [9] |
| Distribution transformers | Final voltage conversion for buildings and campuses | 3–6 months → 1–2 years or longer; demand +41% since 2019 [9] |
| High-voltage circuit breakers | Fault interruption and grid protection | 77 weeks (2023) → 125 weeks (H2 2025) [5] |
| Switchgear (MV/LV) | Switching, protection, isolation inside substations and data halls | Multi-year backlogs; output expansions fully absorbed [24] |
| Heavy gas turbines | On-site or grid generation for large loads | Slot reservations through the late 2020s; 116 GW contracted at one OEM [23] |
Table 1. The electrical chain of artificial intelligence: equipment classes and lead-time deterioration, 2023–2026.
1.3 Why Grid Hardware Cannot Scale Like Semiconductors
The semiconductor industry solved its scarcity problem, at least partially, through a combination of enormous capital investment, government subsidy, and an underlying manufacturing model built for replication: identical wafers, identical process steps, and factories designed to stamp out billions of near-identical devices. Grid hardware obeys different physics and different economics. A large power transformer is closer to a ship than to a chip. Each unit is substantially bespoke, engineered to a specific substation, voltage class, impedance requirement, cooling scheme, and transport constraint. Its production requires grain-oriented electrical steel that only a handful of mills on Earth can supply, copper conductor wound by technicians whose skills take years to develop, insulation systems that must be dried in vacuum ovens for days or weeks, and factory test bays capable of applying transmission-level voltages—facilities so expensive and specialized that they themselves constitute a bottleneck. None of these steps compresses under capital pressure alone. Money can build a new factory in three years; it cannot conjure a certified test engineer, a qualified steel grade, or a fifty-year reputation for units that do not fail in service.
This is why the shortage has persisted even as manufacturers earn record profits and announce record expansions. The constraint is not willingness to invest. It is the long-cycle nature of the industry itself, layered atop several simultaneous demand booms.
1.4 A Shortage Created by Several Booms at Once
Artificial intelligence did not create the transformer shortage. It arrived in the middle of one and made it structural. At least six demand streams now pull from the same factories at the same time. First, replacement demand: a large share of the American transformer fleet was installed in the postwar decades and is approaching or exceeding its design life. Second, electrification: electric vehicles, heat pumps, and industrial electrification add load at every voltage level. Third, renewable interconnection: every solar farm, wind project, and battery plant requires its own step-up transformers and switchgear. Fourth, extreme weather and grid hardening: hurricanes, wildfires, and winter storms destroy equipment and motivate utilities to hold larger spares inventories. Fifth, reindustrialization: semiconductor fabs, battery plants, and reshored factories are themselves enormous electrical loads. And sixth—the accelerant—AI data centers, which arrive in gigawatt increments and compress into two or three years the connection timelines that utilities once spread across a decade.
The North American Electric Reliability Corporation’s 2025 Long-Term Reliability Assessment, released in January 2026, quantified the collision: summer peak demand across the bulk power system is now forecast to grow by 224 gigawatts over the next ten years—a 69 percent increase over the prior year’s forecast—with winter peaks rising even faster, and thirteen of twenty-three assessment areas facing elevated or high resource-adequacy risk within five years. NERC explicitly identified supply-chain delays for grid equipment among the threats to infrastructure pace.[19][21]
“The system is changing faster than the infrastructure needed to support it.”
— John Moura, Director of Reliability Assessments, North American Electric Reliability Corporation [20]
Not every analyst accepts the most alarming projections. Grid Strategies, reviewing the NERC assessment for a coalition of environmental organizations, argued that the load forecasts may double-count speculative data-center applications and underestimate resources in advanced development—and, notably, that data-center growth itself may be limited by shortages of chips, power transformers, and other key inputs.[22] The disagreement is instructive: even the skeptics of the demand forecast concede that transformer scarcity is real enough to constrain the very boom being forecast.
1.5 From Chip Allocation to Factory-Slot Allocation
In 2023, the scarce commodity in artificial intelligence was an allocation of H100 accelerators. By 2026, an equally consequential scarce commodity is a slot on a transformer factory’s production schedule. Manufacturers have formalized the shift. GE Vernova books gas-turbine “slot reservation agreements” years ahead of production.[23] HD Hyundai Electric operates an explicit slot-reservation system in which customers pre-book manufacturing positions and pay premium prices for guaranteed delivery windows, while the company pursues what it calls a selective order policy—declining lower-margin work entirely.[31] The language of the AI industry—allocation, reservation, priority access—has migrated wholesale into heavy electrical manufacturing. What is being allocated is no longer silicon. It is industrial time.
1.6 The New Meaning of “Speed to Power”
Real-estate brokers once marketed data-center sites by fiber connectivity and tax treatment. They now market “speed to power”: the interval between commitment and energization. A site with an executed interconnection agreement, an energized substation, and headroom above current load commands a premium that no other attribute can match, because it is the only schedule advantage that money can still buy once the equipment queue has formed. Sites without a credible power strategy increasingly do not get built at all: of the roughly twelve gigawatts of 2026 U.S. data-center capacity announced across some 140 projects, industry trackers found only about five gigawatts actually under construction, with much of the remainder stalled in significant part by high-voltage transformer, switchgear, and interconnection constraints. Electrical equipment represents less than a tenth of a data center’s total cost—and, at present, essentially all of its schedule risk.
1.7 Grid-Hardware Scarcity as the Next Strategic Constraint
The transition from GPU scarcity to grid-hardware scarcity carries three strategic implications that the remainder of this paper develops. First, the constraint has moved from a sector the United States dominates in design (semiconductors) to a sector in which it holds a visible manufacturing deficit (heavy electrical equipment), transforming a commercial procurement problem into a question of industrial statecraft. Second, the constraint has moved from a globally traded, easily transported good to one that weighs hundreds of tons, resists substitution, and embeds fifty-year relationships between buyer and builder—making alliances, standards, and trust matter as much as price. Third, the constraint has moved from private balance sheets into public infrastructure: when a hyperscaler outbids a rural cooperative for a transformer, the consequences fall on communities that never participated in the AI economy at all. These are the questions of geography, materials, procurement, and governance to which the following sections turn.
Section 2: The Geography of Electrical Conversion
Every era of industrial competition produces its own strategic map. The oil age drew attention to the Persian Gulf, the Strait of Hormuz, and the refinery belt of the American Gulf Coast. The semiconductor age drew attention to Hsinchu, Seoul, Eindhoven, and Phoenix. The age of artificial intelligence is now drawing a third map—quieter, less photographed, and in many ways older—composed of transformer works, electrical-steel mills, switchgear plants, high-voltage test laboratories, and the heavy-transport corridors that connect them. This section reads that map country by country, because the ability to convert electricity into computation is now distributed across a small number of industrial geographies whose politics, alliances, and vulnerabilities differ profoundly.
The financial results reported through the first half of 2026 make the geography legible. The companies that build conversion equipment are experiencing the strongest demand environment in their modern history, and where their order books grow tells us where the AI economy’s electrical foundation is actually being purchased.
2.1 The United States: The Largest Opportunity and the Most Visible Deficit
The United States presents the paradox at the heart of Transformer Diplomacy: it is simultaneously the world’s largest and fastest-growing market for grid conversion equipment and a country that allowed much of its capacity to manufacture that equipment to erode over four decades. American data centers are projected by the IEA to account for nearly half of all U.S. electricity-demand growth between now and 2030, by which point the country will consume more electricity for data processing than for the production of aluminum, steel, cement, chemicals, and all other energy-intensive goods combined.[13] Yet the majority of large power transformers energized on American soil are imported, the domestic workforce of winding technicians and test engineers is aging, and a single mill in Butler, Pennsylvania, produces the country’s entire domestic supply of grain-oriented electrical steel.[32]
The market response has been dramatic. Eaton, the Dublin-domiciled but operationally American power-management giant, reported record first-quarter 2026 results in which Electrical Americas data-center orders rose approximately 240 percent year over year, total Electrical-sector backlog grew 48 percent, and management raised full-year organic growth guidance to roughly ten percent.[27][28] The company now tracks 32 gigawatts of U.S. data-center capacity under construction—about 70 percent of it tied to AI—against a total pipeline it estimates at up to 228 gigawatts, the equivalent of roughly twelve years of construction at 2025 build rates.[28]
“Strong demand across our markets drove solid first quarter performance.”
— Paulo Ruiz, Chief Executive Officer, Eaton Corporation [27]
GE Vernova, the closest thing the United States possesses to a national electrical champion, reported second-quarter 2026 orders of $24.2 billion—up 88 percent organically—with total backlog reaching a record $176 billion and Electrification-segment equipment backlog up 69 percent year over year to roughly $41 billion, swollen by more than $5 billion of first-half data-center orders. The company completed its acquisition of Prolec GE in February 2026, consolidating North American transformer production, and raised its full-year free-cash-flow guidance by fully $5 billion.[23] Its chief executive framed the moment in language that would have seemed fantastical for a heavy-electrical company a decade ago:
“…the early stages of a multi-decade growth opportunity.”
— Scott Strazik, Chief Executive Officer, GE Vernova, Q2 2026 earnings call [24]
And the federal government has now placed its thumb on the scale. The April 2026 Defense Production Act determination authorizes the Department of Energy to make direct purchases, purchase commitments, and financial commitments to expand domestic grid-equipment manufacturing, with fiscal-year 2026 DPA funds available and standard procedural requirements waived to expedite implementation.[6][7][8] Combined with the Hitachi Energy expansion in Virginia[1], announced Korean and American plant investments, and the tariff wall surrounding imported steel, the United States is attempting something it has not seriously tried since the Cold War: the deliberate reconstruction of a heavy-electrical industrial base, on a timeline set by its technology sector.
2.2 China: Scale, State Coordination, and the Captive Market Advantage
China occupies the opposite corner of the map. Its transformer industry—anchored by giants such as TBEA, Sieyuan, and China XD—was built to serve the largest and fastest-growing electricity system ever constructed, including the world’s only commercial fleet of ultra-high-voltage transmission corridors operating at 800 and 1,100 kilovolts. That captive domestic market gives Chinese manufacturers continuous production volume, deep supplier ecosystems, and state-coordinated access to materials. By early 2026, reports indicated that major Chinese transformer manufacturers were fully booked through 2027, absorbing not only domestic demand but also orders from developing economies and from Western buyers desperate enough to accept the associated risks.
For the United States and its allies, Chinese equipment presents the sharpest version of the trust dilemma that runs through this paper. Chinese transformers are frequently cheaper and sometimes available years earlier than allied alternatives. But a large power transformer is not a passive object; modern units ship with sensors, monitoring electronics, and communication interfaces, and they anchor substations that constitute the physical nervous system of a national economy. American policy has already answered part of the question—executive actions dating to 2020 restricted bulk-power-system equipment from foreign adversaries, and the 2026 DPA determination is explicitly framed around reducing reliance on potentially coercive suppliers.[6] The unresolved question is what to do about the rest of the world, where Chinese conversion equipment is rapidly becoming the default infrastructure of electrification, carrying with it long-term service relationships, spare-parts dependencies, and standards influence. Transformer Diplomacy, like semiconductor diplomacy before it, is partly a contest over whose hardware the developing world builds its grids upon.
2.3 South Korea: The Allied Industrial Bridge
If the United States is the demand pole and China the scale pole, South Korea has become the indispensable bridge. Korea’s three major power-equipment makers—HD Hyundai Electric, Hyosung Heavy Industries, and LS Electric—entered 2026 with a combined order backlog exceeding 32 trillion won (roughly $23 billion), an all-time record driven overwhelmingly by North American demand.[29] Hyosung’s power-equipment division alone booked 4.17 trillion won of new orders in the first quarter and became the first Korean maker to surpass a 15-trillion-won backlog, underpinned by its 765-kilovolt ultra-high-voltage transformer technology—the voltage class that has become, in the words of Korean analysts, a “selling point” for attracting American data centers.[30] HD Hyundai Electric posted a 24.9 percent operating margin in the first quarter, with 73 percent of new orders coming from North America and export sales representing 81 percent of revenue; its backlog reached $7.9 billion, and it signed a $119 million ultra-high-voltage transformer contract on the floor of a Chicago trade show.[29][31]
“…orders continue to grow thanks to the expansion of AI data centers.”
— HD Hyundai Electric official, first-quarter 2026 earnings disclosure [31]
Crucially, the Korean firms are not merely exporting; they are planting factories inside the American perimeter. HD Hyundai Electric is investing $200 million in a new transformer plant in Alabama, LS Electric is expanding its Utah facility, and Hyosung plans to raise capacity at its Memphis plant by more than 50 percent by 2028.[29] This is Transformer Diplomacy functioning as designed: an allied industrial power converting its manufacturing depth into physical presence within its partner’s market, hedging tariff risk while deepening the alliance’s collective conversion capacity. It is also a reminder of how concentrated the top of the industry remains—globally, only a handful of companies can manufacture transformers at the highest voltage classes, and two of them are Korean.
2.4 Japan: Materials Sovereignty and Precision Engineering
Japan’s role in the transformer economy is easily misread because its most important contributions are upstream and therefore invisible in trade statistics about finished machines. Nippon Steel and JFE rank among the world’s few producers of the highest grades of grain-oriented electrical steel—the laser-scribed, domain-refined material that determines a transformer core’s efficiency—while Japanese firms supply bushings, tap-changer components, insulation systems, and the precision manufacturing tools used in transformer works worldwide. Hitachi, through Hitachi Energy (the former ABB power-grids business it acquired), is simultaneously one of the world’s top finished-transformer manufacturers, and its more-than-$1-billion American expansion program—of which South Boston is the cornerstone—illustrates how Japanese industrial capital is being redeployed inside the allied perimeter.[1][4] Japan’s position resembles the one it holds in semiconductors, where it dominates photoresists and specialty gases rather than leading-edge logic: a materials-sovereignty power whose cooperation is a precondition for everyone else’s factories.
2.5 Europe: Engineering Networks and the High-Voltage Industrial Base
Europe hosts the deepest concentration of high-voltage engineering expertise outside East Asia. Siemens Energy’s Grid Technologies division has become one of the clearest financial barometers of the global conversion boom: in the first quarter of fiscal 2026 the company reported record group orders of €17.6 billion and a record backlog of €146 billion, with Grid Technologies orders up 21.8 percent and its divisional backlog reaching €45 billion, lifted by triple-digit-million-euro U.S. data-center contracts.[25][26] By the second quarter, group orders had reached a new all-time high of €17.7 billion, backlog had climbed to €154 billion, and the company raised Grid Technologies’ revenue-growth guidance to 25–27 percent with margins of 18–20 percent—extraordinary figures for heavy electrical manufacturing.[26]
“We have made a very strong start to the financial year.”
— Christian Bruch, Chief Executive Officer, Siemens Energy [25]
European strength extends beyond one company: ABB in switchgear and power electronics, Schneider Electric in data-center electrical architecture, and the HVDC and cable champions of the North Sea corridor all feed the same buildout. Europe’s vulnerability is different from America’s—less a manufacturing deficit than a velocity deficit, with transmission permitting measured in a decade or more and energy costs pressuring the competitiveness of its own mills. But as a source of allied conversion capacity, engineering standards, and testing infrastructure, Europe is a pillar of any credible transformer compact.
2.6 Canada and Mexico: The North American Manufacturing Perimeter
The practical unit of American electrical-industrial planning is not the United States but North America. Mexico hosts major transformer manufacturing—most prominently the Prolec operations that GE Vernova moved to full ownership in February 2026[4]—along with wire, harness, and component ecosystems that feed U.S. plants. Canada supplies transformers, hydro-linked expertise, and materials, and shares with the United States an integrated grid whose reliability institutions, including NERC itself, are continental by charter.[19] Tariff politics periodically strain the perimeter, but the industrial logic is durable: for equipment that weighs hundreds of tons, adjacency is a supply-chain strategy in itself, and the USMCA region increasingly functions as a single conversion-capacity bloc negotiating, collectively if implicitly, with Asian and European suppliers.
2.7 India, Türkiye, and the Emerging Production Belt
A third tier of the map is forming across India, Türkiye, Brazil, and Southeast Asia—economies with growing domestic electricity demand, credible engineering bases, and governments eager to climb the electrical value chain. India in particular combines an enormous captive market with export ambitions, and global OEMs, including Siemens Energy, have directed brownfield investments toward Indian, Croatian, Austrian, Saudi, and Brazilian capacity as part of multibillion-euro expansion programs.[26] For the allied world, the emerging belt offers both an opportunity—additional trusted-adjacent capacity that can relieve pressure on the core—and a contest, because China is courting the same geographies with finance, speed, and standards. Where the emerging producers ultimately anchor their supply chains will shape the second decade of Transformer Diplomacy more than any single Western factory.
2.8 From Supply Chain to Alliance Map
Read together, the geography resolves into a recognizable strategic structure. The United States supplies demand, capital, and political urgency; Korea and Japan supply manufacturing depth and materials; Europe supplies engineering networks and high-voltage expertise; Mexico and Canada supply adjacency; the emerging belt supplies future capacity; and China supplies the competitive pressure that makes coordination necessary. No single node is sufficient. The table below summarizes the map that the doctrine in Section 6 will attempt to organize.
| Geography | Primary Role in Conversion Capacity | Signature Strength (2026 evidence) | Principal Vulnerability |
| United States | Largest demand center; rebuilding producer | DPA determination; Hitachi $457M Virginia plant; GEV $176B backlog [1][6][23] | Four-decade manufacturing and workforce deficit; single GOES mill [32] |
| China | Scale producer; UHV pioneer | Manufacturers fully booked through 2027; captive UHV market | Excluded from allied bulk-power systems on security grounds [6] |
| South Korea | Allied industrial bridge | KRW 32T+ combined backlog; 765 kV capability; U.S. plants in AL/UT/TN [29][30] | Tariff exposure until U.S. plants ramp (2028) |
| Japan | Materials sovereignty; top-tier OEM | Premium GOES grades; Hitachi Energy $1B+ U.S. program [1] | Domestic demography; energy import dependence |
| Europe | High-voltage engineering base | Siemens Energy €154B backlog; GT margins 18–20% [26] | Permitting velocity; energy costs |
| Canada / Mexico | North American perimeter | Prolec GE consolidation; adjacency for 400-ton cargo [4] | Periodic trade-policy friction |
| India / Türkiye / Brazil | Emerging production belt | OEM brownfield investments; growing export base [26] | Contested between allied and Chinese ecosystems |
Table 2. The geography of electrical conversion: roles, strengths, and vulnerabilities across the transformer alliance map.
Section 3: The Material Anatomy of Transformer Power
Strategic analysis of technology often stops at the level of the finished machine, as though a transformer were a single object that a country either can or cannot produce. In reality, a large power transformer is an assembly of a dozen specialized subsystems, each with its own materials, suppliers, manufacturing tools, and choke points, and the sovereignty of the finished machine is only as strong as the weakest of them. The nationality of the factory does not necessarily determine the nationality of the machine. A transformer assembled in Virginia may stand upon Japanese core steel, Chilean copper, European bushings, German tap changers, and monitoring electronics fabricated in Asia. This section opens the tank and reads the machine’s anatomy as a map of material dependence, because Transformer Diplomacy conducted at the level of finished goods alone will fail at the level of components.
3.1 The Core: Where Metallurgy Becomes Electrical Efficiency
At the heart of every transformer lies a laminated core of grain-oriented electrical steel, or GOES—a silicon-iron alloy whose crystal grains have been aligned through a sequence of cold rolling and annealing so precise that it remained a guarded metallurgical art for decades after its American invention in the 1930s. The alignment creates a preferential path for magnetic flux, slashing the energy dissipated as heat every time the field reverses, sixty times per second, for the machine’s entire fifty-year life. Core steel quality therefore determines a transformer’s no-load losses, its acoustic signature, its size, and a meaningful share of its lifetime cost. The highest grades—thin-gauge, laser-scribed, domain-refined—are produced by only a handful of mills worldwide, concentrated in Japan, South Korea, China, and Europe.
The United States retains exactly one domestic producer: Cleveland-Cliffs, whose Butler Works in Pennsylvania is the country’s sole source of GOES and of the high-permeability grades used in power-transformer cores.[32] The strategic weight of that single mill has been formally recognized twice within a year: the Defense Logistics Agency awarded Cleveland-Cliffs a five-year, $400 million sole-source contract in September 2025 to supply up to 53,000 short tons of GOES for government stockpiling[33], and the April 2026 presidential determination expressly named “electrical core steel” among the materials essential to national defense.[6] Cleveland-Cliffs has meanwhile moved downstream, converting its idled Weirton, West Virginia, tinplate facility into a $150 million distribution-transformer plant designed to pull additional GOES demand through Butler.[34]
“I can’t think of a better business move than the production of electrical transformers.”
— Lourenco Goncalves, Chairman, President and CEO, Cleveland-Cliffs [34]
3.2 The Electrical-Steel Paradox
Here the material story acquires a paradoxical structure that policy has yet to resolve. Tariffs of 50 percent on imported steel—extended in 2025 to cover electrical-steel cores and laminations—protect the Butler mill and make domestic transformer assembly more attractive. But domestic GOES capacity cannot yet supply the full requirement of the very transformer factories the tariffs are meant to encourage, particularly in the premium thin-gauge grades where import reliance has been deepest. Every new American transformer plant announced on a 2027–2028 timeline will need qualified electrical steel from somewhere, and steel qualification for transformer cores is itself a multi-year process. Meanwhile, the same rolling and melting infrastructure that produces GOES also produces non-oriented electrical steel for electric-vehicle motors, so the electrification of transport competes directly with the electrification of the grid for finite mill capacity. The result is a genuine industrial-policy dilemma: protection without sufficient domestic capacity raises costs for the buildout; open imports without protection guarantee the domestic capacity never gets built. Section 6 proposes treating this as an allied-materials problem rather than a purely national one.
3.3 Copper and Aluminum: The Conductive Arteries
Around the core are wound the conductors—continuously transposed copper cable in large units, precision-wrapped, brazed, and braced against electromagnetic forces that can reach many tons during a fault. Copper links the transformer economy to the global commodities cycle and to the same mines and smelters that feed electric vehicles, wiring, and the data centers themselves; the IEA’s Energy and AI analysis provided the first systematic estimates of data-center demand for critical minerals and flagged the concentration of their supply chains.[12] Aluminum serves as substitute and complement—lighter and cheaper per unit of conductivity, standard in many distribution transformers, and increasingly considered wherever copper economics bind. The conductor choice is a quiet example of design-level diplomacy: material availability now shapes engineering standards, not merely the reverse.
3.4 Insulation, Fluids, and the Slow Chemistry of Reliability
A transformer is, in one sense, a carefully organized argument between voltage and distance, and insulation is the medium of that argument. Kraft paper and pressboard—specialized cellulose products from a small number of qualified mills—wrap every conductor and fill every duct, and their vacuum-drying cycles are among the least compressible steps in manufacturing: moisture measured in parts must be driven from tons of cellulose before oil ever enters the tank. The insulating and cooling fluid itself—mineral oil traditionally, natural and synthetic esters increasingly where fire safety and environmental exposure matter—adds another qualified-supplier dependency. None of these materials is exotic. All of them are certified, and certification, not chemistry, is the choke point: a new supplier of transformer-grade pressboard needs years of testing history before a conservative industry will stake a fifty-year machine on it.
3.5 Bushings, Tap Changers, and the Components That Stop the Line
Two component families deserve particular attention because their failure to arrive stops final assembly entirely. High-voltage bushings—the engineered ceramic-and-condenser passages through which conductors exit the grounded tank—are produced by a concentrated set of specialist firms, and bushing shortages have repeatedly idled otherwise complete transformers. On-load tap changers—the mechanical intelligence inside the machine that adjusts voltage ratio under load—are dominated by an even smaller set of manufacturers, led by German specialists whose products sit inside transformers of every nationality. Sensors, protective relays, and monitoring electronics add a digital material layer with its own semiconductor dependencies and its own cybersecurity implications, since these are the components that speak to the outside world. The 2026 presidential determination’s inclusion of “power control electronics” and “protective relay systems” reflects a growing official understanding that the transformer’s vulnerability is distributed across its bill of materials.[6][7]
3.6 The Manufacturing Tools and the Testing-Bay Constraint
Beneath the materials lies a further stratum: the tools. Core-cutting lines, winding machines, vacuum-drying autoclaves, and vapor-phase ovens are themselves specialized capital goods with their own concentrated suppliers and their own lead times, which is why a greenfield transformer factory takes three to four years to reach production even when every permit is granted on day one. And at the end of every line stands the test bay—the high-voltage laboratory where each unit must survive impulse tests simulating lightning, applied-voltage tests at multiples of operating stress, and heat runs lasting days. Test bays capable of transmission-class voltages cost tens of millions of dollars, exist in limited numbers, and cannot be rented casually; they are a bottleneck within the bottleneck, and any national expansion strategy that funds winding lines without funding test capacity will discover the omission at the worst possible moment.
3.7 The Bottleneck Cascade
The anatomy yields a systemic insight: transformer scarcity is not one shortage but a cascade of potential shortages, any of which can become binding as the others are relieved. Relieve the factory-slot constraint and the GOES constraint binds; relieve GOES and bushings bind; relieve bushings and test bays bind; relieve test bays and heavy transport binds—the specialized Schnabel railcars that carry 400-ton units exist in the low hundreds worldwide. Policy that targets whichever bottleneck is currently most visible will perpetually chase the cascade. Policy that maps the full material passport of the machine, and invests across it, can break the cascade’s logic. The table below sketches that passport.
| Component / Input | Function | Supply Concentration | Dependency Character |
| Grain-oriented electrical steel (GOES) | Core; determines losses and efficiency | A handful of mills globally; one U.S. producer (Butler Works) [32][33] | Specialized; long qualification; allied-manageable |
| Copper conductor / CTC cable | Windings; current carrying | Global commodity, concentrated smelting [12] | Abundant but price-cyclical |
| Kraft paper / pressboard | Solid insulation | Few qualified specialty mills | Certification-locked |
| Insulating fluids (mineral oil, esters) | Cooling and insulation | Moderate; qualified grades | Replaceable with lead time |
| HV bushings | Insulated passage through tank | Concentrated specialist firms | Design-locked; line-stopping |
| On-load tap changers | Voltage regulation under load | Very concentrated (German-led) | Design-locked; trusted-allied |
| Relays, sensors, monitoring electronics | Protection and digital visibility | Global electronics chains | Cyber-sensitive; short-cycle |
| Winding machines, vacuum ovens, core lines | Manufacturing tools | Concentrated toolmakers (EU/Japan) | Long-cycle capital goods |
| HV test bays | Certification of every unit | Limited installed base; $10M+ each | Bottleneck within the bottleneck |
| Schnabel railcars / heavy transport | Delivery of 200–500 ton units | Low hundreds of cars worldwide | Shared scarce logistics asset |
Table 3. The Transformer Material Passport: components, concentration, and the character of each dependency.
3.8 Material Security and the Limits of Autarky
The passport clarifies why complete self-sufficiency is neither achievable on relevant timelines nor, properly understood, the correct objective. Dependencies differ in kind. Abundant, globally traded inputs like copper require hedging, not reshoring. Certification-locked inputs like pressboard require early qualification of second sources. Design-locked components like tap changers require trusted-allied relationships, because redesigning around them takes longer than the crisis one is preparing for. Coercive dependencies—inputs controlled by a strategic competitor willing to weaponize them—are the only category demanding elimination at nearly any cost. A doctrine that treats all dependence as equally dangerous will spend scarce capital reshoring the wrong things; a doctrine that distinguishes trusted, diversified, concentrated, and coercive dependency can direct each policy tool to the dependency it actually fits. That distinction becomes the analytical spine of the findings in Section 7 and the pillars in Section 6.
Section 4: Procurement Hierarchy and Infrastructure Crowding
Scarcity does not merely raise prices. It creates hierarchies. When a good cannot be produced fast enough for everyone who wants it, the decisive question stops being “what does it cost?” and becomes “who stands where in the queue?” For most industrial products, markets answer that question invisibly and tolerably. For grid conversion equipment, the answer is neither invisible nor tolerable, because the queue for transformers is simultaneously the queue for economic development, housing, hospital expansion, storm recovery, and national reindustrialization. This section examines the procurement hierarchy that has formed around the world’s transformer factories, and the phenomenon this paper calls infrastructure crowding: the displacement of essential public electrification by the concentrated purchasing power of the AI buildout.
4.1 The Factory Schedule as a Strategic Asset
A transformer factory’s production schedule for the next four years is now among the more strategically consequential documents in the industrial world, and manufacturers manage it accordingly. Order backlogs have become the industry’s central financial metric: $176 billion at GE Vernova, €154 billion at Siemens Energy, more than 32 trillion won across Korea’s big three, 48 percent year-over-year Electrical-sector backlog growth at Eaton.[23][26][29][28] Slot-reservation systems, advance payments, price-escalation clauses, and selective-order policies have shifted bargaining power decisively toward producers for the first time in a generation. HD Hyundai Electric’s explicitly “selective order policy focused on profitability”—declining lower-margin work while customers pre-pay for guaranteed slots—is not an aberration but the new normal of the industry.[31] In such a market, access is not purchased at the margin; it is negotiated years ahead, with balance-sheet strength as the entry ticket.
4.2 Who Stands at the Front of the Queue?
The resulting hierarchy is observable across utility filings, earnings calls, and procurement disputes, and it runs roughly as follows. Hyperscalers and frontier-AI developers stand first: they can pay cash deposits years ahead, sign take-or-pay commitments, absorb price escalation without regulatory approval, and—when the grid queue is too slow—buy turbines and build their own generation. Large investor-owned utilities stand second: they enjoy scale, framework agreements, and creditworthiness, though every dollar of advance payment must eventually survive a rate case. Semiconductor fabs and advanced manufacturers stand third, often armed with governmental backing that partially offsets their smaller order volumes. Independent data-center and power developers stand fourth, speculating on equipment as they speculate on land, sometimes reserving slots for projects that will never be built. Municipal utilities and rural cooperatives stand fifth, lacking both scale and balance sheet—precisely the buyers about whom the National Rural Electric Cooperative Association warned.[5] And housing, hospitals, schools, and public infrastructure stand last, invisible in the queue because their transformers are ordered late, in small numbers, by the least-empowered purchasers, even though their connection delays translate most directly into human cost.
| Queue Position | Buyer Class | Sources of Procurement Power | Characteristic Vulnerability |
| 1 | Hyperscalers / frontier-AI developers | Cash deposits; take-or-pay; own-generation option; multi-GW framework deals | Political backlash; stranded reservations if demand forecasts miss |
| 2 | Large investor-owned utilities | Scale; framework agreements; credit strength | Rate-case recovery risk on advance payments |
| 3 | Semiconductor fabs / advanced manufacturing | National-priority status; state incentives | Compete with AI for identical equipment classes |
| 4 | Independent developers | Speed; speculative reservations | Financing fragility; queue attrition |
| 5 | Municipal utilities / rural cooperatives | Pooled purchasing (where it exists) | No scale; storm-restoration exposure [5] |
| 6 | Housing, hospitals, public infrastructure | Political voice (episodic) | Last in line; delays borne by communities |
Table 4. The procurement hierarchy for grid conversion equipment, 2026.
4.3 The Ability to Buy Time
What the front of the queue is actually purchasing is time—and the ability to buy time has become a form of power that operates before markets, before regulators, and largely outside public view. A hyperscaler that reserves four years of substation equipment is not merely procuring hardware; it is pre-allocating a region’s electrical future. This is procurement power in the strict sense: the capacity to convert financial strength into temporal priority over a shared industrial resource. The December 2025 PJM capacity auction made the stakes vivid, when the nation’s largest grid operator fell nearly 6,600 megawatts short of its reserve target for the first time in its history, with 94 percent of projected load growth attributable to AI-driven data centers—an outcome that pushed capacity prices to record highs and triggered bipartisan political backlash.[35]
4.4 Binding Versus Speculative Demand, and the Stranded-Equipment Problem
The hierarchy is complicated by a peculiarity of the current boom: nobody is entirely sure how much of the demand is real. Interconnection queues contain far more proposed gigawatts than will ever be constructed, as developers file duplicative applications across multiple utilities to preserve optionality. Grid Strategies’ review of the NERC assessment argued that data-center load forecasts likely double-count such phantom projects.[22] Speculative reservations impose real costs on everyone behind them in line; and if the demand forecasts prove overbuilt, the system will face the mirror-image problem of stranded equipment—transformers manufactured against reservations that evaporate, in configurations too customized to redeploy easily. Both failure modes point to the same remedies: financial commitment tests that separate binding from speculative demand, transferability of reserved equipment wherever technically possible, and regulatory visibility into who holds which slots.
4.5 Should Transformer Reservations Be Disclosed?
That last remedy raises a genuinely novel governance question. Securities law requires disclosure of material financial positions; energy law requires disclosure of generation adequacy; but no regime currently requires anyone to disclose control over long-lead electrical equipment, even though such control now shapes regional development as surely as spectrum licenses or water rights. A disclosure regime need not be punitive—it could be as simple as confidential reporting to FERC or state commissions of reservations above a capacity threshold, aggregated and published in anonymized form. The argument against is commercial sensitivity; the argument for is that the public cannot govern an allocation it cannot see. This paper takes the position, developed as Pillar Three of the doctrine, that visibility is the minimum condition of legitimacy for private allocation of a quasi-public industrial resource.
4.6 The Conflict Between AI and Reindustrialization
Infrastructure crowding acquires its sharpest edge in the collision between two policies the United States is pursuing simultaneously: winning the AI race and reshoring manufacturing. Both policies are transformer-intensive. A semiconductor fab, a battery plant, and an AI campus draw upon the same substation transformers, the same switchgear, the same crews, and the same factory slots. When they compete, the AI buyer usually wins on speed and balance sheet—which means the country’s technology strategy can quietly cannibalize its industrial strategy through the medium of the equipment queue. The IMF’s modeling of AI-driven electricity demand adds a distributional warning: under constrained infrastructure expansion, U.S. electricity prices could rise 8.6 percent, spreading the cost of the crowding to every ratepayer.[38][39] No market mechanism internalizes these cross-policy externalities; only deliberate allocation doctrine can.
4.7 A Doctrine for Fair Infrastructure Allocation
The elements of such a doctrine follow from the analysis. First, the customer creating the infrastructure need should bear its direct risk—through deposits, take-or-pay contracts, and infrastructure charges that insulate other ratepayers, in the spirit of the Ratepayer Protection Pledge that seven major AI companies signed in March 2026. Second, speculative applications should not hold grid or equipment capacity indefinitely—readiness milestones and forfeitable commitments must separate real projects from options. Third, regulators should receive visibility into major long-lead reservations, so that allocation is at least observable. Fourth, essential and smaller buyers need collective procurement mechanisms—pooled purchasing, standardized designs, and credit support that convert their aggregate demand into queue standing. Fifth, unused equipment must be transferable whenever technically possible, so that a canceled campus’s transformer becomes a cooperative’s storm spare rather than a stranded asset. These five principles reappear, institutionalized, in Section 6.
Section 5: The Governors’ Grid
American energy policy is written in Washington but delivered in state capitals, and the transformer era has made governors into infrastructure diplomats whether they sought the role or not. It is governors who stand at factory groundbreakings, as Virginia’s did in South Boston[1][2]; governors who negotiate with hyperscalers over siting, water, and tax treatment; governors who answer to ratepayers when capacity prices spike; and governors whose economic-development promises collide with interconnection queues they do not control. Because states differ in market structure, resource endowment, and political economy, they are running a natural experiment in how to govern the collision between AI load growth and conversion-equipment scarcity. This section reads seven state models as seven distinct answers, and then proposes the compact that could knit them together.
5.1 Seven States, Seven Models
Virginia governs saturation. Northern Virginia hosts the largest data-center concentration on Earth, and the state’s task is no longer attraction but management: transmission buildout, ratepayer protection, land-use conflict, and—with the Hitachi expansion—the deliberate cultivation of the manufacturing base that serves the load it already hosts.[1][2] Texas operates the pay-your-way grid: ERCOT’s energy-only market and large-load interconnection reforms ask data centers to bear cost and curtailment risk directly, trading regulatory protection for speed. Pennsylvania practices conditional industrial support, leveraging PJM’s capacity crisis and its own manufacturing assets—including the Butler Works GOES mill[32]—to demand that AI investment arrive coupled to grid contribution and jobs. Michigan pursues restorative capacity, using data-center demand to justify grid modernization and generation life-extension that its legacy industrial base needs anyway. California attempts flexible growth under a constrained system, pushing load flexibility, storage, and efficiency because its permitting and land constraints preclude brute-force expansion; the Princeton–Camus–encoord line of research showing that modest data-center flexibility can unlock large amounts of interconnection capacity is, in effect, California’s thesis generalized.[40] Arizona embodies semiconductor sovereignty meeting desert constraints: fabs and AI campuses compete for the same water, power, and equipment, forcing explicit prioritization earlier than elsewhere. Indiana has become the laboratory of bring-your-own-power, welcoming campuses that arrive with their own generation and thereby shifting the bottleneck from the transmission queue to the turbine queue.
| State | Governing Model | Core Mechanism | Transformer-Diplomacy Significance |
| Virginia | Governing saturation | Transmission buildout + manufacturing attraction | Hosts largest DC cluster and the nation’s largest LPT plant [1] |
| Texas | Pay-your-way grid | Energy-only market; large-load cost assignment | Fastest connections; demand bears its own risk |
| Pennsylvania | Conditional industrial support | Capacity-market leverage; PJM politics | Home of sole U.S. GOES mill [32]; PJM shortfall epicenter [35] |
| Michigan | Restorative capacity | DC demand funds modernization | Legacy grid renewed on AI’s balance sheet |
| California | Flexible growth | Load flexibility; storage; efficiency | Tests whether flexibility substitutes for hardware [40] |
| Arizona | Semiconductor sovereignty vs. desert limits | Explicit fab/DC prioritization | First open AI-vs-reindustrialization allocation |
| Indiana | Bring-your-own-power | On-site generation mandates/incentives | Moves bottleneck from wires to turbines [23] |
Table 5. Seven states, seven models of governing AI load growth under equipment scarcity.
5.2 The Governors’ Transformer Compact
The seven models share a weakness: each state negotiates alone against counterparties—hyperscalers and manufacturers—who operate nationally. A Governors’ Transformer Compact, organized through the National Governors Association or a standing interstate agreement, could convert fragmentation into leverage around six commitments. Full-cost responsibility: large loads pay the full incremental cost of the conversion capacity they consume, in every member state, eliminating the race to subsidize. Credible load forecasting: member states share and reconcile data-center pipelines, deflating the phantom demand that distorts planning.[22] Long-lead equipment transparency: states exchange information on major reservations and public-fleet spares. Community resource protection: common floors for water, land, and ratepayer safeguards, so that concessions cannot be arbitraged across borders. Grid contribution: standardized expectations that campuses contribute flexibility, storage, or generation, in the manner the flexibility literature shows is cheap for developers and valuable for grids.[40] Interstate equipment cooperation: mutual-assistance agreements extending the storm-restoration tradition to transformer spares, so that a unit idled in one state can energize a hospital in another. None of these commitments requires federal legislation. All of them convert the states from price-takers in the equipment queue into a coordinated demand bloc—Transformer Diplomacy practiced domestically.
5.3 The Political Geography of Transformer Diplomacy
The state experiments also reveal the political coalition available to whoever chooses to lead on this issue nationally. Transformer manufacturing employs welders, winders, and technicians in exactly the regions—Southern Virginia, western Pennsylvania, West Virginia, Alabama, Tennessee—where industrial employment carries the greatest political salience.[1][29][34] Ratepayer protection resonates in every PJM state that watched capacity prices spike.[35] Reliability anxiety spans the NERC risk map without partisan pattern.[19][21] And the AI industry itself, having signed ratepayer pledges and funded grid upgrades, has acknowledged that its social license depends on not being seen to crowd out communities. Few industrial-policy questions offer such naturally bipartisan raw material; the doctrine that follows is written to be executable by either party.
Section 6: Building a Transformer-Diplomacy Doctrine
Nations write doctrines when an activity becomes too consequential to leave to improvisation. The United States has a nuclear doctrine, a cyber doctrine, an export-control doctrine for advanced semiconductors—and, as of this writing, no doctrine at all for the electrical conversion capacity upon which its technological, industrial, and military ambitions now jointly depend. The April 2026 Defense Production Act determination supplied authority and urgency; it did not supply architecture.[6][7][8] This section proposes that architecture. Its premise is the sentence that has organized this entire paper: a country does not possess computational capacity merely because it owns chips, models, land, or generation—it possesses computational capacity when it can reliably convert electricity into operating intelligence.
A doctrine is necessary for three reasons. First, the time constants of the problem exceed political cycles: a transformer ordered by one administration is energized under another, and only doctrine survives transitions. Second, the problem spans jurisdictions—federal, state, utility, corporate, allied—that no single instrument reaches; doctrine assigns responsibilities across them. Third, scarcity invites improvisation, and improvisation under scarcity systematically favors the powerful; doctrine is how a society decides its allocation principles before the emergency, rather than during it. The doctrine proposed here consists of six pillars, mapped onto the Five Circuits introduced earlier, followed by an implementation architecture, a financing plan, metrics, and—because intellectual honesty requires it—a register of the ways the doctrine could fail.
| Circuit of Transformer Diplomacy | Corresponding Doctrine Pillar(s) |
| Production Capacity | Pillar One: National Conversion Reserve · Pillar Five: Transformer Workforce Corps |
| Material Security | Pillar One (Layer Four) · Pillar Two: Allied Transformer Compact |
| Procurement Power | Pillar Three: Long-Lead Disclosure and Allocation · Pillar Four: Small-Utility Facility |
| Standards Power | Pillar Six: Grid-Hardware Passport and Interoperability Standard |
| Alliance Capacity | Pillar Two: Allied Transformer Compact |
Table 6. Mapping the Five Circuits onto the six pillars of the doctrine.
6.1 Pillar One: The National Conversion Reserve
The United States maintains a Strategic Petroleum Reserve because it once concluded that oil scarcity was a national-security condition, not a market inconvenience. The same conclusion now applies to conversion equipment, and the answer is a National Conversion Reserve organized in four layers. Layer One—deployable emergency equipment: a federally owned fleet of mobile substations, transportable transformers, and standardized spares, pre-positioned regionally, able to restore critical loads within days of storm, attack, or catastrophic failure; existing utility spare-sharing programs are the seed, but they were sized for weather, not for an era in which the replacement queue is three years long.[5][9] Layer Two—regional strategic spares: co-funded stocks of the most common large-transformer configurations, held against the highest-consequence substations, with the DLA’s $400 million GOES stockpiling contract demonstrating that the government has already accepted the stockpile logic one level upstream.[33] Layer Three—virtual manufacturing capacity: standing DPA purchase commitments and capacity-reservation contracts that pay manufacturers to maintain surge capability—idle winding lines, extra test-bay shifts, qualified second-source components—that the market alone would never carry; this is the layer the April determination makes immediately financeable.[6][8] Layer Four—component and material reserves: bushings, tap changers, core steel, and pressboard held against the cascade dynamics documented in Section 3, because a reserve of finished transformers that cannot be repaired is a museum. Governance matters as much as inventory: the Reserve requires transparent release criteria, utility co-investment to prevent moral hazard, and technology-refresh rules so that stored equipment does not obsolesce into scrap.
6.2 Pillar Two: The Allied Transformer Compact
Section 2 demonstrated that conversion capacity is an alliance property: American demand, Korean and Japanese manufacturing depth, European engineering, North American adjacency.[26][29] The Allied Transformer Compact formalizes what tariffs and ad-hoc deals currently improvise, through five instruments. Allied production mapping: a shared, classified-where-necessary inventory of member manufacturing capacity, test bays, and material sources, so the alliance knows what it collectively possesses before a crisis reveals what it lacks. Mutual procurement commitments: reciprocal market access and volume guarantees that give allied manufacturers the demand certainty to expand—the Korean plants rising in Alabama, Utah, and Tennessee show the model working through investment; the Compact would secure it through agreement.[29] Shared emergency inventories: extending Layer One of the Reserve across borders, in the tradition of IEA oil-stock obligations. Coordinated export security: common screening of conversion equipment sold into third markets, so that alliance members do not undercut one another’s security rules. Joint industrial investment: co-financed expansion of the choke points no single member will fund alone—premium GOES capacity, test laboratories, heavy-transport assets. Domestic production and allied production are complements, not rivals: the Compact’s purpose is to ensure that “Buy American” is nested inside “Build Allied,” rather than set against it.
6.3 Pillar Three: Long-Lead Equipment Disclosure and Allocation
Pillar Three institutionalizes Section 4’s answer to infrastructure crowding. Its disclosure half requires confidential reporting, to FERC and state commissions, of transformer and switchgear reservations above defined capacity thresholds, published in aggregated form—making the queue visible without exposing commercial terms. Its allocation half replaces first-come-first-served, during declared scarcity, with readiness-based allocation: equipment flows to projects that have demonstrated site control, financing, interconnection progress, and—for public-interest loads—essential-service status. Anti-hoarding rules attach forfeitable deposits to reservations and time-limit unexercised slots; a secondary market for qualified equipment lets canceled projects’ units transfer, at regulated terms, to buyers further back in the queue, converting the stranded-equipment problem into a liquidity mechanism. The pillar deliberately mirrors how the industry already behaves—slot reservations, deposits, selective ordering[31][23]—but bends those private mechanisms toward public visibility and fairness.
6.4 Pillar Four: The Small-Utility Purchasing and Credit Facility
The Reuters reporting that opened this paper carried a warning inside its statistics: long-term supply agreements “don’t solve everything, particularly for smaller utilities that don’t have the scale.”[5] Pillar Four answers with a federally chartered facility offering four services to municipal utilities, rural cooperatives, and small public power. Pooled purchasing aggregates their orders into framework contracts large enough to command factory slots. Credit support guarantees the advance payments that manufacturers now demand but small balance sheets cannot carry. Standardized designs—a limited menu of pre-engineered distribution and small-power transformers—let pooled orders run down factory lines at volume, attacking simultaneously the cost problem and the tens-of-thousands-of-configurations interchangeability problem the DOE has documented.[9] Shared technical expertise supplies the specification, inspection, and testing capabilities that small utilities cannot individually maintain. A priority carve-out for essential community infrastructure—hospitals, water systems, storm restoration—ensures that the facility serves the queue positions the market serves last.
6.5 Pillar Five: The Transformer Workforce Corps
Every constraint documented in this paper eventually resolves into people: winders whose craft takes five years to mature, test engineers certified on equipment few universities possess, transformer designers whose apprenticeship is measured in careers. Hitachi Energy’s own executives, at the South Boston groundbreaking, identified training and even local housing as binding constraints on their 825-job expansion.[2][3] Pillar Five builds the human layer through regional training hubs co-located with the new plants in Virginia, Alabama, Tennessee, and West Virginia[1][29][34]; apprenticeships and stackable credentials that make grid manufacturing a visible, portable career; engineering and materials research funding—university programs in magnetics, insulation science, and power electronics that rebuild the academic base beneath the industry, of the kind MIT’s Energy Initiative and Princeton’s ZERO Lab have begun modeling for the data-center age[37][40]; and a National Grid Service track that ties tuition support to service years in utilities, manufacturers, or the Reserve itself. Machines can be stockpiled; expertise can only be grown, and its growing season is a decade.
6.6 Pillar Six: The Grid-Hardware Passport and Interoperability Standard
The final pillar addresses the quietest form of power in the system: standards. A Grid-Hardware Passport would accompany every major unit sold into member markets, documenting provenance of core steel, bushings, tap changers, and electronics; firmware and communication interfaces; and service-network commitments—making the material anatomy of Section 3 legible to buyers and regulators, and making “cybersecurity by industrial lifespan” auditable across a machine’s fifty-year service life. In parallel, an interoperability standard—a rationalized set of preferred ratings, impedances, and physical interfaces for new public procurement—would attack the configuration chaos that currently prevents utilities from sharing equipment in emergencies.[9] Standardization has limits, and the doctrine should respect them: transmission-class units will always carry site-specific engineering, and standards frozen too early can entomb obsolete designs. The objective is not uniformity but exchangeability—a grid whose critical organs, like the alliance’s ammunition, fit more than one weapon.
6.7 Implementing the Doctrine
Responsibilities distribute across five actors. Federal: fund and govern the Reserve; negotiate the Compact; operate disclosure and the small-utility facility; direct DPA instruments toward the mapped choke points.[6][8] State: enact the Governors’ Compact of Section 5; align siting and rate policy with full-cost responsibility. Utility: adopt standardized designs; participate in spare-sharing; report reservations. Hyperscaler and large-customer: honor ratepayer-protection and readiness commitments; contribute flexibility and, where appropriate, co-fund community equipment—the price of the social license their queue position consumes. Manufacturer: accept surge-capacity contracts, passport transparency, and expanded training in exchange for the demand certainty the doctrine uniquely provides—certainty the industry has said it wants, in an environment its own executives describe as a multi-decade opportunity.[24][25]
6.8 Financing Transformer Diplomacy
The doctrine is financed from six sources, most already in motion. Defense Production Act support: Title III purchases, purchase commitments, and loan instruments activated by the April 2026 determinations.[6][8] Federal credit: loan guarantees for domestic and allied-in-America capacity, on the model of the DLA steel contract.[33] Large-load infrastructure charges: tariffed contributions from the customers whose demand drives the buildout, consistent with the pledges the AI industry has already signed. State economic-development funds, redirected from pure incentive competition toward compact-consistent grid contribution—Virginia’s performance-based grant to Hitachi Energy is the template.[2] Utility cost recovery for prudently incurred reserve and standardization investments. Allied co-investment through the Compact for the shared choke points. The sums are large in absolute terms and small relative to the stakes: the entire multi-year cost of the doctrine is comparable to a single quarter of hyperscaler capital expenditure.
6.9 Measuring Success
A doctrine without metrics is a speech. Success should be measured against a public dashboard: median lead times by equipment class (target: distribution transformers back under twelve months, large power transformers under twenty-four, against the 160-week baselines of 2026[5][9]); domestic-plus-allied share of installed large transformers; number of qualified GOES sources serving the allied market (target: at least three, against today’s single domestic mill[32]); reserve depth in days-to-restore for the 100 most critical substations; small-utility order fulfillment times relative to hyperscaler times (target: convergence); and workforce pipeline counts against the announced plants’ hiring schedules.[1][29]
6.10 Strategic Limits and Necessary Cautions
Honesty about failure modes is part of the doctrine. The demand forecast may be wrong: if AI electricity demand disappoints—a possibility serious analysts defend[22]—the Reserve must be sized to no-regret levels justified by weather and aging-fleet replacement alone. Industrial policy can protect inefficiency: surge contracts need competitive discipline and sunset review. Reserves can become obsolete: technology-refresh rules and the solid-state transition[24][18] must be designed in, not bolted on. Standards can freeze innovation: the interoperability menu must version, like software. Security rules can delay infrastructure: passport requirements must be administrable, or they become another queue. Public support can subsidize private speculation: every public dollar should attach to the readiness tests of Pillar Three. The doctrine’s one-sentence summary absorbs all six cautions: build what must be domestic, secure what should be allied, standardize what must be shared, reserve what cannot be replaced quickly, and require the largest beneficiaries of the AI grid to pay for the conversion capacity they consume.
Section 7: What Have We Learned?
A long argument earns its conclusion only by compressing honestly. Five findings survive the compression.
7.1 Finding One: Compute Capacity Is Inseparable from Conversion Capacity
The evidence assembled here—160-week lead times[5], NERC’s 224-gigawatt demand revision[19][21], the PJM shortfall[35], the IEA’s finding that transformer wait times doubled in three years while transmission takes four to eight years to build[13]—converges on a single proposition: the binding constraint on national AI capacity has moved beneath the chip. A country does not possess AI capacity merely because it owns chips. It possesses AI capacity when it can convert electricity into dependable computation. The IEA’s executive director compressed the same finding into six words:
“…there is no AI without energy…”
— Fatih Birol, Executive Director, International Energy Agency [14]
7.2 Finding Two: A Domestic Factory Is Not the Same as a Domestic Supply Chain
Section 3’s material passport showed that the nationality of assembly and the nationality of capability diverge: a Virginia-built transformer can remain dependent on foreign core steel, bushings, tap changers, and electronics. Domestic assembly strengthens resilience, but true conversion security depends upon the materials, components, tools, knowledge, and service networks behind the factory—which is precisely why the presidential determination named electrical core steel and manufacturing tools alongside finished transformers[6], and why the Pentagon contracted the Butler mill’s output years ahead.[32][33]
7.3 Finding Three: Procurement Power Can Determine Regional AI Development
Section 4 documented the queue and its hierarchy: slot reservations, selective ordering, advance payments, and the systematic disadvantage of small utilities and public infrastructure.[5][31] In a constrained infrastructure economy, the ability to reserve industrial time can become as important as the ability to invent technology—and unlike invention, reserved time is a zero-sum asset whose allocation society is entitled to see and shape.
7.4 Finding Four: Standardization Is a Form of National Resilience
The DOE’s finding of tens of thousands of transformer configurations[9] converts a procurement inefficiency into a security exposure: equipment that cannot be exchanged cannot be shared in a crisis. Standardization does not eliminate innovation; it creates the common interfaces that allow innovation, emergency response, and industrial scale to coexist—as it has in every other strategic industry, from shipping containers to ammunition.
7.5 Finding Five: Trusted Industrial Alliances Are More Realistic Than Complete Autarky
Section 2’s geography and Section 3’s dependency taxonomy—trusted, diversified, concentrated, coercive—support the paper’s final analytical claim. Korea’s record backlogs serving American demand[29][30], Japan’s materials position, Europe’s engineering depth[26], and North American adjacency[4] collectively constitute a conversion capacity no member possesses alone. The strongest transformer strategy is neither unrestricted dependence nor complete isolation. It is an alliance system designed around redundancy, transparency, technical trust, and shared industrial capacity.
7.6 The Five Findings Together, and the Deeper Lesson
Taken together, the findings describe a single structural truth about the Five-Layer AI Economy: value may be created at the top of the stack, but possibility is created at the bottom. The deeper lesson reaches beyond transformers. Every general-purpose technology in history—steam, electricity, computing—eventually collided with the slower physical systems it depended upon, and the societies that mastered each era were those that learned to govern the collision rather than merely celebrate the technology. Artificial intelligence has now had that collision, and the point of impact is a machine invented in the 1880s. How the United States and its allies respond—with doctrine or with improvisation, with alliance or with autarky, with visible allocation or with quiet crowding—will shape not only who builds AI, but what kind of industrial society builds it.
Conclusion: The Machine Between the Power Plant and the Model
This paper began at a groundbreaking in Southern Virginia and has traveled through factory schedules in Ulsan and Nuremberg, a steel mill in Butler, Pennsylvania, the risk maps of NERC, the order books of the world’s electrical manufacturers, and the procurement offices where the AI economy’s real queue is kept. It ends where it began: with a machine that does not think, learn, or generate a single token, and without which no machine can.
The return of industrial time is the era’s defining sensation. For thirty years, the technology industry lived inside software time, where scaling meant replication and the marginal unit was nearly free. The transformer has reintroduced the older physics: the marginal unit is four years away, weighs four hundred tons, and is spoken for. Every actor in the AI economy is currently relearning, at different speeds and different costs, that industrial time cannot be venture-funded into compression—it can only be planned for, shared, and, at the margin, purchased from whoever planned earlier.
From supply chain to strategic system: what was once a procurement category has become an arena in which manufacturing capacity, materials, finance, standards, and alliances interact—the Five Circuits this paper has traced. The United States has, in the space of a single year, stood up the legal machinery of that recognition: a Defense Production Act determination naming grid equipment essential to national defense[6][7], a Pentagon steel stockpile[33], and a wave of allied factory investment on American soil.[1][29] What it has not yet done is bind the machinery into doctrine. That is the unfinished work this paper has attempted to draft.
The new meaning of AI sovereignty follows directly. Sovereignty in artificial intelligence will not be measured only in models trained or chips fabricated, but in the quieter question of whether a nation can energize what it builds—whether its ambition survives contact with its substations. By that measure, the sovereign nations of the AI era will be those that treated conversion capacity as a strategic resource while their competitors still treated it as a line item.
The responsibilities divide cleanly. The AI builders—whose capital expenditures now rival interstate-highway programs[36] and whose demand has repriced the entire electrical industry[23][26][27]—owe the system full-cost responsibility, credible forecasts, flexibility, and restraint in the queue. Governments owe the system doctrine: reserves, compacts, disclosure, workforce, and standards, executed with the humility the cautions of Section 6 demand. Governors, standing between the two, hold the compact that could align fifty separate negotiations into one. And the manufacturers—enjoying the strongest market in their history[24][25][29]—owe the system the transparency and surge capacity that public support justly prices in.
The transformer, in the end, is the machine that measures commitment. Anyone can announce a model, a campus, or a corridor; announcements are free. A transformer order is a four-year, nine-figure, non-refundable statement about what a society actually intends to build. Read the order books—$176 billion here, €154 billion there, thirty-two trillion won across the Sea of Japan[23][26][29]—and you are reading the most honest forecast of the AI economy that exists: a forecast written not in press releases but in reserved steel, scheduled test bays, and committed industrial time.
The future beneath intelligence is being assembled now, in South Boston and Ulsan, Butler and Monterrey, Memphis and Nuremberg, by welders and winders whose names will appear in no model card. The first transformer will keep converting tokens into intelligence, faster every year. Whether it gets the chance will be decided by the second transformer—and by whether the nations that want the future were willing to build the machine that carries it.
The second transformer is no longer invisible. It is now the measure of who is serious.
Footnotes and Endnotes
[1] Hitachi Energy, “Hitachi Energy breaks ground on the nation’s largest facility for the production of large power transformers in South Boston, Virginia,” Press Release, June 29, 2026. https://www.hitachienergy.com/us/en/news-and-events/press-releases/2026/06/hitachi-energy-breaks-ground-on-the-nation-s-largest-facility-for-the-production-of-large-power-transformers-in-south-boston-virginia
[2] Engineering News-Record, “Hitachi Breaks Ground on $457M Va. Large Power Transformer Plant Expansion” (remarks of Gov. Abigail Spanberger), July 2026. https://www.enr.com/articles/63252-hitachi-breaks-ground-on-457m-va-large-power-transformer-plant-expansion
[3] Daily Energy Insider, “Hitachi breaks ground on new large power transformer production facility in Virginia” (remarks of Greg Callahan), June–July 2026. https://dailyenergyinsider.com/news/52883-hitachi-breaks-ground-on-new-large-power-transformer-production-facility-in-virginia/
[4] Construction Review Online, “Hitachi Energy Breaks Ground on $457 Million Virginia Transformer Plant” (OEM North American expansion commitments; GE Vernova–Prolec GE consolidation, February 2026), June 2026. https://constructionreviewonline.com/hitachi-energy-breaks-ground-on-457-million-virginia-transformer-plant/
[5] Kavya Balaraman, “US power companies scramble to secure equipment as surging data center demand strains supplies,” Reuters, July 9, 2026 (lead-time data attributed to C. Boucher, Wood Mackenzie; quotation of Louis Finkel, NRECA). https://www.reuters.com/business/energy/us-power-companies-scramble-secure-equipment-surging-data-center-demand-strains-2026-07-09/
[6] The White House, “Presidential Determination Pursuant to Section 303 of the Defense Production Act of 1950, as Amended, on Grid Infrastructure, Equipment, and Supply Chain Capacity” (Presidential Determination No. 2026-10), April 20, 2026. https://www.whitehouse.gov/presidential-actions/2026/04/presidential-determination-pursuant-to-section-303-of-the-defense-production-act-of-1950-as-amended-on-grid-infrastructure-equipment-and-supply-chain-capacity/
[7] Federal Register, Presidential Determination No. 2026-10 of April 20, 2026, published April 23, 2026 (91 FR 21931). https://www.federalregister.gov/documents/2026/04/23/2026-08013/presidential-determination-pursuant-to-section-303-of-the-defense-production-act-of-1950-as-amended
[8] Foley & Lardner LLP, “Defense Production Act Determinations Could Reshape Domestic Energy and Infrastructure Strategy,” May 2026. https://www.foley.com/p/102mqpt/defense-production-act-determinations-could-reshape-domestic-energy-and-infrastru/
[9] U.S. Department of Energy, Office of Electricity, “Distribution Transformer Webinar” (demand growth, lead times, and configuration diversity). https://www.energy.gov/oe/distribution-transformer-webinar-text-alternative
[10] U.S. Department of Energy, “DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers” (Lawrence Berkeley National Laboratory, 2024 Report on U.S. Data Center Energy Use), December 2024. https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers
[11] Lawrence Berkeley National Laboratory News Center, “Berkeley Lab Report Evaluates Increase in Electricity Demand from Data Centers” (lead researcher Arman Shehabi), January 15, 2025. https://newscenter.lbl.gov/2025/01/15/berkeley-lab-report-evaluates-increase-in-electricity-demand-from-data-centers/
[12] International Energy Agency, “AI is set to drive surging electricity demand from data centres…” announcing the special report Energy and AI (statement of Executive Director Fatih Birol), April 2025. https://www.iea.org/news/ai-is-set-to-drive-surging-electricity-demand-from-data-centres-while-offering-the-potential-to-transform-how-the-energy-sector-works
[13] International Energy Agency, Energy and AI — Executive Summary (transmission timelines of four to eight years; transformer and cable wait times doubled in three years; U.S. data-center demand growth). https://www.iea.org/reports/energy-and-ai/executive-summary
[14] International Energy Agency, “Data centre electricity use surged in 2025, even with tightening bottlenecks driving a scramble for solutions,” announcing Key Questions on Energy and AI (statement of Fatih Birol; supply-chain tightening for gas turbines and transformers), 2026. https://www.iea.org/news/data-centre-electricity-use-surged-in-2025-even-with-tightening-bottlenecks-driving-a-scramble-for-solutions
[15] Electric Power Research Institute (EPRI), Powering Intelligence: Analyzing Artificial Intelligence and Data Center Energy Consumption — Executive Summary. https://powering-intelligence.epri.com/executive-summary.html
[16] NVIDIA Developer Blog, “NVIDIA Vera Rubin Pod: Seven Chips, Five Rack-Scale Systems, One AI Supercomputer.” https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/
[17] NVIDIA Developer Blog, “NVIDIA 800 V HVDC Architecture Will Power the Next Generation of AI Factories.” https://developer.nvidia.com/blog/nvidia-800-v-hvdc-architecture-will-power-the-next-generation-of-ai-factories/
[18] arXiv:2606.25095, survey of AI data-center power-delivery architectures, medium-voltage distribution, and solid-state transformers, 2026. https://arxiv.org/abs/2606.25095
[19] North American Electric Reliability Corporation (NERC), 2025 Long-Term Reliability Assessment, released January 2026. https://www.nerc.com/globalassets/our-work/assessments/nerc_ltra_2025.pdf
[20] Robert Walton, “NERC forecasts peak demand to rise 24% on new data center loads,” Utility Dive (quotation of John Moura, NERC), January 30, 2026. https://www.utilitydive.com/news/nerc-10-year-peak-demand-forecast-jumps-24-on-new-data-center-loads/810955/
[21] Sonal Patel, “NERC Warns Long-Term Grid Reliability Risks Mounting from Surging Demand, Lagging Resources,” POWER Magazine, February 2026. https://www.powermag.com/nerc-warns-long-term-grid-reliability-risks-mounting-from-surging-demand-lagging-resources/
[22] Robert Walton, “NERC overstates reliability risks in long-term assessment: Grid Strategies,” Utility Dive (Grid Strategies review by A. Brooks, M. Goggin, J. Wilson), March 2026. https://www.utilitydive.com/news/nerc-overstates-reliability-risks-ltra-grid-strategies/814292/
[23] GE Vernova Inc., Form 8-K, Second Quarter 2026 Financial Results (orders $24.2B, +88%; backlog $176B; Electrification backlog +69%; 116 GW gas-turbine contracts), July 2026. https://www.sec.gov/Archives/edgar/data/0001996810/000199681026000147/gev2q2026form8-k.pdf
[24] Benzinga, “Full Transcript: GE Vernova Q2 2026 Earnings Call” (remarks of CEO Scott Strazik; switchgear capacity 9,000→10,500 units; solid-state transformer prototypes), July 2026. https://www.benzinga.com/news/26/07/60604921/full-transcript-ge-vernova-q2-2026-earnings-call
[25] Siemens Energy, “Earnings Release Q1 FY 2026” (record orders €17.6B; backlog €146B; statement of CEO Christian Bruch), February 2026. https://www.siemens-energy.com/global/en/home/press-releases/earnings-release-q1-fy-2026.html
[26] Siemens Energy, “Earnings Release Q2 FY 2026” (all-time-high orders €17.7B; backlog €154B; Grid Technologies guidance raised to 25–27% growth, 18–20% margin), May 2026. https://www.siemens-energy.com/us/en/home/press-releases/earnings-release-q2-fy-2026.html
[27] Eaton Corporation plc, “Eaton Reports Record First Quarter 2026 Results” (statement of CEO Paulo Ruiz; Electrical Americas 12-month orders +42%; Electrical backlog +48%), May 2026. https://www.eaton.com/us/en-us/company/news-insights/news-releases/2026/eaton-reports-record-first-quarter-2026-results.html
[28] Eaton Corporation plc, First Quarter 2026 Earnings Presentation (data-center orders +~240% YoY; U.S. capacity under construction and pipeline estimates), May 2026. https://www.eaton.com/content/dam/eaton/company/investor-relations/quarterly-earnings/filings/2026/q1/q1-2026-analyst-presentation.pdf
[29] Ellie Kim, “HD Hyundai Electric, Hyosung Heavy Industries, LS Electric See Record KRW 32 Trillion Order Backlog on AI Data Center Boom,” Alpha Biz, May 2026 (Alabama, Utah, Memphis capacity expansions). https://m.alphabiz.co.kr/news/amp.html?ncode=1065573071517120
[30] Seoul Economic Daily, “Korean Power Equipment Makers Clinch Record Orders on US Grid, AI Boom” (Hyosung KRW 15.1T backlog; 765 kV technology), May 1, 2026. https://en.sedaily.com/news/2026/05/01/korean-power-equipment-makers-clinch-record-orders-on-us
[31] Seoul Economic Daily, “HD Hyundai Electric Q1 Operating Profit Jumps 18.4% on North American Transformer Orders” (company statement; selective order policy; backlog $7.888B), April 28, 2026. https://en.sedaily.com/business/2026/04/28/hd-hyundai-electric-q1-operating-profit-jumps-184-percent
[32] Cleveland-Cliffs Inc., “Butler Works” (sole U.S. producer of grain-oriented electrical steel and high-permeability transformer grades). https://www.clevelandcliffs.com/operations/steelmaking/butler-works
[33] Defence Blog, “Pentagon locks in critical transformer steel supply” (Defense Logistics Agency five-year, $400M sole-source GOES contract with Cleveland-Cliffs, awarded September 9, 2025), July 2026. https://defence-blog.com/pentagon-locks-in-critical-transformer-steel-supply/
[34] Manufacturing Dive, “Cleveland-Cliffs moves ahead with $150M electric transformer plant” (remarks of Chairman, President and CEO Lourenco Goncalves; Weirton, W.Va. conversion). https://www.manufacturingdive.com/news/cleveland-cliffs-confirms-150-million-electric-transformer-weirton-plant/722787/
[35] Santiago Gallino, “AI’s Supply Chain Problem,” Knowledge at Wharton, The Wharton School of the University of Pennsylvania (PJM reserve shortfall of ~6,600 MW; 94% of load growth from AI data centers), May 2026. https://knowledge.wharton.upenn.edu/article/ais-supply-chain-problem/
[36] The Atlantic (syndicated), “Inside the Dirty, Dystopian World of AI Data Centers” (quotation of Jesse Jenkins, Princeton University; hyperscaler capital expenditures exceeding $600B since November 2022), March 2026. https://dnyuz.com/2026/03/13/inside-the-dirty-dystopian-world-of-ai-data-centers/
[37] Nancy W. Stauffer, “The multi-faceted challenge of powering AI,” MIT Energy Initiative, Massachusetts Institute of Technology, January 2025. https://energy.mit.edu/news/the-multi-faceted-challenge-of-powering-ai/
[38] Christian Bogmans, Patricia Gomez-Gonzalez, Ganchimeg Ganpurev, Giovanni Melina, Andrea Pescatori, and Sneha D. Thube, “Power Hungry: How AI Will Drive Energy Demand,” IMF Working Paper 2025/081, International Monetary Fund, April 2025. https://www.imf.org/en/Publications/WP/Issues/2025/04/21/Power-Hungry-How-AI-Will-Drive-Energy-Demand-566304
[39] International Monetary Fund Blog, “AI Needs More Abundant Power Supplies to Keep Driving Economic Growth,” May 13, 2025. https://www.imf.org/en/Blogs/Articles/2025/05/13/ai-needs-more-abundant-power-supplies-to-keep-driving-economic-growth
[40] David Roberts, “For data centers, a little flexibility goes a long way,” Volts (with Jesse Jenkins, Princeton ZERO Lab, and Astrid Atkinson, Camus Energy; Princeton–encoord–Camus flexibility study), March 25, 2026. https://www.volts.wtf/p/for-data-centers-a-little-flexibility
