The Great American Grid — Rebooted

Photograph: Visitor7 / Wikimedia Commons (CC BY-SA 3.0).

Inside the American power grid — buckling under AI, on fire in the West, and finally being rewired by the people who broke it.

At seven in the evening on July 10, 2024, a bolt of lightning struck a 230-kilovolt transmission line in Fairfax County, Virginia. It was an unremarkable evening thunderstorm. The line, owned by Dominion Energy, was equipped for precisely this contingency: a lightning arrestor designed to shunt the surge harmlessly to ground. The arrestor failed. A permanent electrical fault locked the line out of service. By the standards of a utility that maintains thousands of miles of high-voltage wire in a region known for summer storms, this was a Tuesday.

What happened next was not.

Across the surrounding countryside — the asphalt-and-fiber landscape outside Ashburn, in Loudoun County, where rows of windowless data centers now stretch toward the horizon — sixty hyperscale facilities were watching the grid the way a sniper watches a tree line. Inside each one, banks of uninterruptible power supplies were running an algorithm with a simple, brittle logic: if you see three voltage dips in a single minute, the grid is failing. Switch to diesel. Save the servers.

The grid did not fail. But the protection systems on Dominion’s faulted line tried, six times in eighty-two seconds, to automatically reclose the breaker — a routine recovery procedure called auto-reclosing. Each attempt produced a small voltage flicker, roughly ten percent, well within tolerance for any toaster, refrigerator, or human being. Inside the data centers, sixty UPS systems counted the flickers, decided the apocalypse had arrived, and unplugged themselves from the grid in unison.

Roughly 1.5 gigawatts of demand vanished in under a minute and a half — the rough equivalent of every household in Boston going dark at once, except in reverse: the load had disappeared, and the generators were still spinning, pushing power into wires with nowhere for it to go. PJM Interconnection, the regional grid operator responsible for sixty-seven million people across thirteen states, watched its system frequency leap to 60.047 Hertz, well outside the band that keeps transformers from melting. Operators in PJM’s control rooms threw the throttles down, cutting hundreds of megawatts from natural gas plants in Pennsylvania and a nuclear unit in Virginia. The grid held. Barely.

The North American Electric Reliability Corporation later called the incident a near miss and stood up a Large Loads Task Force. In a way, that response was beside the point. The story of July 10 was not that the grid failed but that the grid almost failed in an entirely new way — failed in a direction it was never designed to fail in, brought there not by a hurricane or a heat dome but by the defensive instincts of buildings full of computers. For a century, American electrical engineering had assumed that load was diverse, distributed, and slow. In Loudoun County, on a Wednesday in summer, load became concentrated, synchronized, and digital. The largest machine ever built by humanity had encountered a new species of stress, and it discovered, in real time, that its reflexes were wrong.

This is the story of how the American power grid — the staid, regulated, century-old machine that won the twentieth century — is being torn open and rebuilt to survive the twenty-first. It is a story of physics and finance, of carbon fiber and grain-oriented steel, of a 1968 court decision about a plastic cup glued to a telephone, and of how a generation of startups, regulators, and tech executives are racing to keep the lights on for a civilization that has, almost without noticing, decided to run itself on electricity.

I. The Demand Tsunami

Photograph: Victor Grigas / Wikimedia Commons (CC BY-SA 3.0).

For more than a decade, American electricity demand was nearly flat. LED bulbs, better refrigerators, and Energy Star washing machines quietly canceled out the country’s growing appetite for gadgets. Utility planners spoke about kilowatt-hours the way librarians speak about new shelves: slow, predictable, manageable.

Then the forecasts broke. The five-year outlook for nationwide peak load growth, as compiled by the consultancy Grid Strategies, rose more than sixfold in three years — from a projected 24 gigawatts of additions in 2022 to 166 gigawatts in 2025. The U.S. Energy Information Administration now expects total electricity use to climb roughly a third by 2030, and a 2025 National Electrical Manufacturers Association projection puts annual consumption growth at 55 percent by 2050, with most of the surge front-loaded into this decade.

There are three reasons. They are not equal, but they are concurrent.

The first is the data center. The Lawrence Berkeley National Laboratory’s most recent assessment, published in December 2024, estimated that U.S. data centers consumed 176 terawatt-hours in 2023 — about 4.4 percent of national electricity — and projected that figure could climb to between 325 and 580 TWh by 2028, capturing as much as 12 percent of all American power. The International Energy Agency reports that data-center electricity use grew 17 percent globally in 2025, against an overall demand growth rate of just 3 percent. Capital expenditures by the five largest U.S. technology companies — Microsoft, Google, Amazon, Meta, and Apple — passed $400 billion in 2025, much of it routed into server farms whose appetite for electrons has become the single largest variable in American grid planning.

Unlike the residential and industrial loads grid planners cut their teeth on, AI workloads do not idle. A modern training cluster does not sleep at night, does not cool off on weekends, does not flicker with weather. It runs near peak utilization continuously, in densities that would have astonished a 1980s utility engineer — tens of megawatts per building, hundreds per campus, gigawatts per region. And it tends to cluster. Northern Virginia, where MAE-East began routing transcontinental internet traffic in the 1990s, now hosts the most concentrated data-center market on Earth. (The widely repeated figure that “70 percent of global internet traffic” passes through the region — promoted by Amazon and Virginia tourism officials alike — is, as Cardinal News and others have documented, more folklore than measurement; Virginia’s actual share of global data-center capacity is closer to 13 percent. But concentration matters more than the marketing number, and concentration is the problem.)

The second driver is the car. To support the roughly 33 million electric vehicles the National Renewable Energy Laboratory projects on American roads by 2030, the country will need on the order of 28 million charging ports — most of them slow Level 1 and Level 2 chargers in suburban garages, the rest fast DC chargers along highways. EV charging is theoretically tractable; cars sit still most of the day, and software can stagger when they suck. In practice, unmanaged charging stacks neatly on top of the existing residential evening peak, when families come home, turn on ovens, and start streaming. The strain falls on the lowest, most fragile rung of the grid: the neighborhood distribution transformer, a refrigerator-sized lump of steel and oil on a pole that was sized in 1978 to power a few air conditioners.

The third is the weather. The same warming that drives wildfires also drives air conditioners; the IEA expects cooling alone to account for more than a fifth of electricity demand growth in fast-electrifying nations like India between 2026 and 2030. Heat domes in Texas, atmospheric rivers in California, and polar vortices over the Midwest now routinely set new records, and electric heat pumps — the climate-friendly replacement for gas furnaces — pull hardest from the grid exactly when transmission lines, suffering their own cold-snap brittleness, are least able to deliver.

The three drivers stack. They are not separate forecasting problems. A heat wave in Phoenix raises HVAC load at the same moment that a Phoenix data center is training a model and a Phoenix commuter has plugged in her car. The grid sees the sum, not the parts.

Energy Demand Drivers

II. The Infrastructure Bottleneck

Photograph: Nixdorf / Wikimedia Commons (CC BY-SA 3.0).

If demand were the only problem, the grid could be brute-forced with capital. The deeper trouble is on the supply side, and it begins with an unglamorous object the size of a small house.

A large power transformer — LPT, in the jargon — is the workhorse of the high-voltage grid. There is no electricity in modern life without one: every long-distance transmission line ends in a transformer that drops voltage so the current can be safely fed into a substation, then a neighborhood, then a wall. They weigh hundreds of tons. They are made one at a time, to bespoke specifications, in factories whose total global output is well short of what the world now wants.

Industry surveys compiled by Wood Mackenzie and reported across 2025 put the average delivery time for a standard power transformer at roughly 128 weeks — nearly two and a half years. Generator step-up transformers, the larger units that connect new power plants to the transmission system, are running closer to 144 weeks, with some specialty orders stretching past four. Prices have risen accordingly: roughly 80 percent since 2019 for large power transformers, more than 20 percent since 2023 for smaller pole-top distribution units. A growing share of the roughly $1.1 trillion that American investor-owned utilities plan to spend between 2025 and 2029 is, in effect, being absorbed by equipment inflation rather than translated into actual capacity.

The reasons are structural and embarrassing. Every utility in America buys transformers to a slightly different spec — different impedance ratings, different bushing layouts, different cooling configurations — and every order arrives at the factory as effectively a one-off. There is no Model T transformer. Building a new domestic LPT plant costs an estimated $450 to 500 million per gigavolt-ampere of capacity, several times the comparable Asian figure.

The bottleneck has a bottleneck. The magnetic core of every high-voltage transformer is made from grain-oriented electrical steel, or GOES, a fussy iron-silicon alloy that took decades to perfect. In the entire United States, only one company produces it: Cleveland-Cliffs, at mills in Butler, Pennsylvania, and Zanesville, Ohio. Those mills cover, by industry estimates, roughly a fifth of domestic demand. The rest is imported. By 2024, the U.S. depended on foreign supply for the great majority of the steel inside its large power transformers — a situation the Department of Commerce has formally flagged as a national-security vulnerability.

Domestic capacity is, belatedly, being built. In February 2026, GE Vernova closed its acquisition of the remaining fifty percent of Prolec GE, its longtime joint venture with the Mexican industrial group Xignux, for $5.275 billion — folding one of the largest North American transformer manufacturers wholly under American corporate control. Cleveland-Cliffs is converting an idled tinplate mill in Weirton, West Virginia, into a $150 million distribution-transformer plant. The Department of Energy’s electrical-steel rule, finalized in 2024, preserved GOES as the dominant transformer-core material. But factories take years to come online, skilled labor is scarce, and the queue is growing faster than the throughput. For the next half-decade, the binding constraint on the American grid will not be money. It will be steel.

III. The Wildfire Crisis

While the grid struggles to grow, the grid it already has is decaying — and, in the most arid corners of the West, catching fire.

The average age of high-voltage transmission infrastructure in the United States is approaching four decades. Wood, steel, and insulators originally specified for a cooler, wetter climate now run hot, dry, and wind-whipped. Between 1992 and 2020, electrical transmission and distribution lines were determined to have ignited more than 32,000 wildfires across the country. Most were small. A handful rewrote the financial logic of the entire utility industry.

On the morning of November 8, 2018, a worn hook on a 1921-vintage transmission tower on Pacific Gas and Electric’s Caribou-Palermo line — an asset the utility had purchased used from another company nearly a century earlier — failed in high wind. The arcing conductor ignited brush in the Feather River Canyon. The Camp Fire killed 85 people, mostly elderly residents of the town of Paradise, and destroyed roughly 18,000 structures. PG&E pleaded guilty in 2020 to 84 counts of involuntary manslaughter — prosecutors could not prove causation in one death — making the case the deadliest corporate prosecution in American history. The company entered bankruptcy in 2019 and ultimately committed roughly $13.5 billion across settlements covering Camp and other recent fires. An internal engineer had requested $800,000 in 2007 to replace the section of line that broke; the project was approved at a quarter of that amount, then quietly killed.

NASA Earth Observatory image by Joshua Stevens, using Landsat data from the U.S. Geological Survey (public domain).

The Camp Fire was not an outlier; it was a tutorial. Hawaiian Electric is restructuring under the weight of litigation from the 2023 Lahaina fire, which killed at least 102 people on Maui. Xcel Energy faces hundreds of millions in claims from the 2021 Marshall Fire in Colorado. PacifiCorp has already been ordered to pay tens of millions in damages from the 2020 Labor Day fires in Oregon, with class actions threatening to add billions more. California, Oregon, and several other Western states apply some form of strict liability — most notoriously California’s doctrine of inverse condemnation, under which a utility whose equipment ignites a fire is liable for damages regardless of negligence. For an investor-owned utility, that is closer to existential risk than to insurance arithmetic.

Table 2: Utility fire liabilities

The response, where it has been most aggressive, has been to put the wires underground. PG&E’s program, announced in 2021, originally aimed to bury 10,000 miles of distribution line; under pressure from California regulators it has been trimmed, and the utility now expects to harden roughly 11,000 miles by 2037, with about 5,000 of those undergrounded. By late 2025, after four years of construction, PG&E had completed 1,000 miles, at a cost that has declined from $4 million per mile to around $3.1 million. Undergrounding works — it nearly eliminates ignition risk on the affected segments — but at a pace and price that strain the imagination. High-voltage transmission undergrounding, which is engineered very differently from distribution, runs $15 million to $35 million per mile and is rarely undertaken at scale.

The bill, inevitably, falls to customers. In 2025, American electric and gas utilities filed for roughly $31 billion in rate increases, more than double the 2024 figure. Residential electricity prices have risen about 40 percent since 2021, well ahead of general inflation. The utilities are not getting richer in any obvious sense; they are running, very hard, to stand still.

IV. The Generation Renaissance

The conventional grid response to load growth — build more dispatchable, carbon-free baseload — has, until very recently, looked impossible. Nuclear was expensive and politically frozen. Geothermal was geographically lucky. Fusion was forty years away, as it had been for forty years. None of that was going to feed an Nvidia cluster.

The companies that need the electrons are now writing the checks.

The Resurrection at Three Mile Island

Photograph: U.S. Department of Energy (public domain).

On September 20, 2024, Constellation Energy and Microsoft announced a twenty-year fixed-price power purchase agreement to restart Unit 1 of the Three Mile Island nuclear plant in Pennsylvania. Unit 1 is not the reactor of American nuclear lore — that was Unit 2, which suffered a partial meltdown in 1979 and is being decommissioned. Unit 1 sat adjacent to that disaster, ran safely for forty more years, and was shuttered in 2019 for the unromantic reason that natural gas had become cheaper than it was.

Constellation will spend approximately $1.6 billion overhauling the plant — turbine, generator, main power transformer, cooling and control systems — and aims to bring it back online as the Crane Clean Energy Center as soon as 2027, with a current target of producing roughly 7 million megawatt-hours of carbon-free electricity per year over a 20-year contract with Microsoft. The Department of Energy closed a $1 billion loan to the project in November 2025. Microsoft does not collocate a data center at the site; the PPA simply matches Microsoft’s regional consumption with the plant’s output. Constellation, in its investor materials, expects the deal to lift its long-run earnings growth meaningfully.

It is hard to overstate what a strange deal this is by historical standards. A software company, founded in 1975 to sell BASIC interpreters for hobbyist computers, is now underwriting the resurrection of a 1970s nuclear reactor on an island in the Susquehanna River to power an artificial-intelligence buildout that did not exist five years ago. The transaction is rational, in the way late-stage capitalism is sometimes rational. It is also a statement: the hyperscalers have decided that the utility procurement cycle is too slow for them, and they will buy their baseload directly.

The Deep Earth

Geothermal energy has always been the renewable that nobody talked about, for the simple reason that you could only get it in places where hot water happened to seep close to the surface — Iceland, Kenya, a few corners of California and Nevada. Enhanced Geothermal Systems, or EGS, propose to remove that constraint by drilling miles into hot dry rock, fracturing it the way oil-and-gas crews fracture shale, and circulating water through the engineered fractures to harvest heat.

The Houston-based startup Fervo Energy is the company that has made the idea real. In 2023, Fervo’s Project Red in Nevada became, by independent measurement, the most productive EGS pilot in history, sustaining controlled flow through 3,250-foot horizontal laterals at 191 degrees Celsius and feeding the grid round-the-clock carbon-free power. The company has since moved to a larger play: Cape Station, in southwest Utah, designed for an initial 500 megawatts of capacity. In June 2025, Fervo announced that its Sugarloaf appraisal well there had reached a true vertical depth of 15,765 feet with a projected bottom-hole temperature of 520 degrees Fahrenheit, drilled in just sixteen days — a 79 percent reduction against the Department of Energy’s baseline for ultra-deep geothermal wells. The independent reserves consultancy DeGolyer & MacNaughton, in a report Fervo commissioned, estimated thermal recovery factors in the range of 50 to 60 percent, against perhaps a third of that for conventional geothermal.

Translation: the oil-patch tooling of horizontal drilling, hydraulic fracturing, and distributed acoustic sensing — refined over two decades by the same shale industry that environmentalists spent two decades trying to shut down — turns out to be exactly the right toolbox for harvesting heat from the basement of the continent. Hundreds of gigawatts of theoretical geothermal capacity exists, by the U.S. Geological Survey’s reckoning, in rock between 10,000 and 20,000 feet down. Fervo is the first company to make it cheaply accessible. It will not be the last.

The Fusion Bet

At the speculative end of the spectrum sits Helion Energy, the Everett, Washington startup that in May 2023 signed the world’s first commercial power purchase agreement for fusion electricity, with Microsoft as offtaker and Constellation as power marketer. Helion has promised to begin delivering at least 50 megawatts by 2028 — a timeline shorter than most fusion researchers think is plausible by decades. Helion’s design is unconventional; it uses a deuterium-helium-3 reaction and a magneto-inertial confinement scheme that, if it works, can produce electricity directly rather than boiling water to spin a turbine. Whether it works is the question. Plasma scientists outside the company range from cautiously interested to politely scathing. The Fusion Industry Association’s 2024 survey found that most fusion companies expect commercial plants between 2030 and 2035; few expect Helion’s 2028.

The fact that Microsoft signed the contract anyway is the data point. It does not mean fusion is around the corner. It means the hyperscalers have decided that their long-run electricity problem is severe enough that even a long-shot bet on a 2028 fusion plant is worth taking. Helion broke ground on its first facility in 2025. The world will see.

V. Rewiring the Arteries

Photograph: Stefan Andrej Shambora / Wikimedia Commons (CC BY 2.0).

New generation is necessary but not sufficient. The electrons have to get from where they are made to where they are needed, and the wires that carry them are the slowest piece of the entire problem.

Building a new high-voltage transmission line in the United States is, in 2026, an exercise in patience that frequently outlasts a presidential administration. Five to ten years of environmental review, right-of-way acquisition, and local litigation are typical. The national interconnection queue — the line of new generation and storage projects waiting to plug into the grid — held more than 10,300 active proposals representing roughly 2,060 gigawatts of capacity as of the most recent counts, with an average wait time north of four years.

The country does not have a decade. So a small set of technologies is being deployed to wring more capacity out of the wires already in the air.

Grid Enhancing Technologies

For most of the past century, transmission lines have been rated by a single, conservative number — a “static” line rating — calculated under worst-case weather assumptions: hot summer, no breeze. The wire is treated as though every day were the worst day. In reality, on a brisk autumn evening, the same wire can safely carry substantially more power.

Dynamic Line Rating, or DLR, replaces the conservative number with a real one. Heimdall Power, a Norwegian startup with a growing American footprint, clamps small sensors directly onto high-voltage conductors — the company calls them “Neurons,” utility crews call them “magic balls” — and streams real-time temperature, current, sag, and vibration data back to grid operators. LineVision, an American competitor, mounts non-contact lidar sensors on the towers themselves. Both let the operator know what the wire can actually do, minute by minute.

In March 2024, Great River Energy, the Minnesota power cooperative, announced what was then the largest DLR deployment in the United States — 52 of Heimdall’s Neurons across roughly 175 miles of transmission line, following a pilot project that had shown an average 42.8 percent capacity gain on a single key line. By the time Great River reported one year of operational data, in 2025, peak-period capacity gains had reached as high as 63 percent on some lines, with millions of dollars in avoided congestion costs. The Federal Energy Regulatory Commission estimates that transmission congestion costs U.S. consumers between $12 billion and $21 billion a year. DLR is a software and hardware overlay that begins paying back the day it is installed, with no new steel in the ground.

The Reconductoring Revolution

When software optimization reaches its physical limits, the wires themselves can be replaced. Traditional transmission conductors are some variant of aluminum strands wrapped around a steel core, a design largely unchanged since the early twentieth century. Steel sags when heated, which is why the static rating exists in the first place. A handful of advanced-conductor startups — TS Conductor in the U.S. is the most aggressive — have replaced the steel core with a composite core of aluminum-encapsulated carbon fiber. The resulting wire is lighter, stronger, and barely expands when hot. It carries up to twice the current of the cable it replaces.

The strategic value lies in what reconductoring does not require. The new conductor strings onto the existing towers, in the existing right-of-way, using existing crews and existing fittings. There is no greenfield permitting, no new land acquisition, no decade-long environmental impact statement. Modeling published in 2024 by GridLab and the Energy Institute at Haas estimated that aggressive reconductoring could meet roughly 80 percent of the interzonal transmission expansion needed for a 90-percent-clean U.S. grid by 2035. The per-project savings, after accounting for the civil engineering costs avoided, run 30 to 40 percent against a traditional new build.

Table 3: Grid enhancements

VI. The Regulatory Hammer

Technology can’t solve a problem that the rules forbid solving. For decades, the rules forbade it.

In July 2023, the Federal Energy Regulatory Commission issued Order №2023, the most significant overhaul of interconnection procedures in twenty years. The old “first-come, first-served” system had been overrun with speculative projects; developers filed applications they had no intention of building, clogging the queue and forcing engineering restudies every time one withdrew. Order 2023 imposed a “first-ready, first-served” cluster study process, with substantial financial commitments, site-control requirements, and withdrawal penalties for projects that drop out. The intent is brutal but rational: clear the ghosts, study the serious projects in regional batches, and impose deadlines on the transmission providers themselves.

Less than a year later, in May 2024, FERC issued Order №1920, which reaches further. It requires regional transmission organizations to plan on a twenty-year horizon, to model multiple scenarios including extreme weather and AI-driven load growth, and — critically — to explicitly consider Grid Enhancing Technologies, advanced conductors, and other low-cost capacity options before they propose billion-dollar new transmission lines. For incumbent utilities that have spent decades treating software as a threat and steel as a profit center, Order 1920 is a slow-acting solvent.

VII. The Telecom Parallel

The remaining question — the question that hangs over every conversation in this industry — is why, given that dynamic line ratings, advanced conductors, and behind-the-meter batteries are already commercial, demonstrably cheaper, and now mandated for evaluation by federal regulators, utilities have been so reluctant to deploy them.

The answer is not technical. It is fiduciary.

Since the early twentieth century, American electric utilities have operated as regulated natural monopolies under what is called Cost-of-Service Regulation. A utility’s allowed revenue is split into two pots. Operating expenses — fuel, vegetation management, third-party software subscriptions — pass through to ratepayers at cost. The utility earns nothing on them. Capital expenditures — substations, towers, transformers, the physical kit the utility owns — are added to a regulated “rate base,” and the utility is permitted to charge customers for depreciation plus an allowed return on equity, typically several hundred basis points above its actual cost of capital. The arithmetic is simple. To grow shareholder dividends, build more stuff.

Economists call the resulting distortion the Averch-Johnson effect, after the two scholars who described it in 1962. Practitioners call it capex bias. A utility executive considering a $10 million software overlay that doubles the effective capacity of an existing line is, in effect, contemplating an action that destroys the engineering justification for a $1 billion new line her shareholders would have collected a regulated return on for the next forty years. The software is the right answer. The shareholders are not paid for right answers. They are paid for assets.

This is the structural reason grid modernization has spent two decades in what the industry, with admirable candor, calls “pilot purgatory.”

There is a precedent. In the middle decades of the twentieth century, AT&T — the Bell System — owned the long-distance lines and the telephones plugged into them and, as a matter of policy, prohibited customers from attaching anything else. The company argued, with the same combination of safety claim and proprietary interest that incumbent utilities now deploy, that allowing foreign attachments would threaten network reliability.

It was cracked, eventually, by lawsuits over very small objects. In 1956, the D.C. Circuit’s Hush-A-Phone decision allowed users to clip a plastic sound-dampening cup over the mouthpiece of a Bell handset. In 1968, the FCC’s Carterfone ruling went further, permitting an acoustic coupler that connected mobile two-way radios to the public telephone network. The principle the FCC articulated — that third-party equipment is lawful so long as it does no actual, demonstrable harm to the network — eventually dissolved AT&T’s hardware monopoly entirely. Answering machines, fax machines, modems, and ultimately the consumer internet itself flowed through the door Carterfone propped open.

Photograph: Museo Nazionale della Scienza e della Tecnologia Leonardo da Vinci, Milano / Wikimedia Commons (CC BY-SA 4.0).

The electrical grid is approaching a Carterfone moment of its own. Hyperscale buyers like Microsoft and Amazon are already routing around the utility procurement cycle by signing direct PPAs with nuclear operators and geothermal startups; they have enough capital to underwrite gigawatt-scale projects without asking a public utility commission’s permission. Behind the meter, customer-owned solar, batteries, and virtual power plants are doing something analogous at smaller scale. And in state capitals, a slow conversion to Performance-Based Regulation — under which utilities are paid for outcomes (carbon reduction, interconnection speed, line capacity unlocked) rather than capital deployed — is beginning to take hold. The Averch-Johnson effect is not a law of physics. It is a regulatory choice, and choices can be unmade.

Epilogue

There is a way of telling the story of the American grid that is mostly bad news: aging hardware, supply-chain captivity, monopoly capture, manslaughter convictions, $31 billion in rate hikes, a near-miss in Virginia that few outside the industry have heard about. All of it is real.

There is another way to tell it, equally true. The same year the lights almost went out in Loudoun, a Norwegian startup put sensors on Minnesota power lines and unlocked nearly half again as much capacity in wires that were already there. A Houston company drilled a 15,000-foot well in Utah in sixteen days. Constellation announced it would resurrect a nuclear reactor on the Susquehanna. Microsoft bet on fusion. FERC, an agency most Americans could not name, issued the most consequential transmission rules in a generation. And in Pennsylvania, a steel mill that has made grain-oriented electrical steel since the 1920s — the steel inside almost every transformer that keeps the country running — geared up to make more.

The grid is the largest machine ever built. It is also, in the end, a political artifact: a set of choices about who pays, who builds, who profits, and who is held responsible when the thing burns. The choices made over the next ten years will determine whether the data centers come online, whether the climate targets are reached, whether the cost of electricity becomes a quietly regressive tax on everything else. The hardware is ready. The wires can carry more current. The reactor can be restarted. The rock is hot enough.

What remains is whether the country can rewrite its rules as quickly as it rewrote its computers. The lightning struck on a Tuesday in July. The grid held. It would be a mistake to assume it always will.

A note on sources and fact-checking

Quantitative claims in this piece were verified against primary and secondary reporting through May 2026. Key figures:

• The July 10, 2024 PJM incident: NERC incident report and follow-on reporting (Belfer Center, Data Center Dynamics, Yale Clean Energy Forum, *gridstatus.io* analysis). Frequency excursion to 60.047 Hz and ~1.5 GW of synchronized load drop are confirmed across multiple sources.
• The widely-cited “70% of global internet traffic” claim about Northern Virginia is contested. Cardinal News and analysts citing JLARC and Cushman & Wakefield data put the region’s share of global data-center capacity closer to 13%. The piece notes this explicitly.
• Camp Fire: 85 deaths, 84 manslaughter pleas (causation unprovable in one death), ~18,000 structures, Caribou-Palermo line, ~1921 vintage. PG&E settlements totaling $13.5 billion covered multiple Northern California fires, not Camp alone.
• Three Mile Island / Crane Clean Energy Center: Sept. 20, 2024 announcement; $1.6 billion restart; 835 MW; 20-year PPA; restart target now accelerated to 2027 (originally 2028); $1 billion DOE loan closed Nov. 2025.
• Fervo Sugarloaf well: 15,765 ft TVD, projected 520°F bottomhole, 16 drilling days, 79% reduction vs. DOE baseline (June 2025 announcement); Cape Station Phase 1 designed for 500 MW.
• Helion: 50 MW or greater target, May 2023 PPA with Microsoft, 2028 target, Constellation as marketer.
• Heimdall Power / Great River Energy: 52 Neurons across ~175 miles, 42.8% pilot gain; subsequent first-year reporting indicates peak-period gains up to ~63%.
• GE Vernova / Prolec GE: $5.275 billion for the remaining 50% stake (joint venture since 1995), closed February 2, 2026.
• Cleveland-Cliffs is the sole U.S. producer of GOES (via the former AK Steel subsidiary, at Butler, PA and Zanesville, OH).
• PG&E undergrounding: originally proposed at 10,000 miles of distribution (not transmission); now scaled to ~5,000 miles undergrounded plus ~6,000 miles of other hardening through 2037; 1,000 miles completed as of October 2025; per-mile cost ~$3.1M, down from $4M.
• Interconnection queue figures are from the Lawrence Berkeley National Laboratory’s annual queue reports; FERC Orders 2023 (July 2023) and 1920 (May 2024) are public documents; Averch-Johnson (1962) is in the American Economic Review; Carterfone (1968) and Hush-A-Phone (1956) are in the FCC and D.C. Circuit records, respectively.