The Machine Beneath the Machine

Right now, somewhere in North America, an operator at PJM Interconnection—the organization that coordinates electricity delivery across 13 states and 65 million people—is watching a screen that displays the frequency of alternating current flowing through the eastern grid. That frequency is supposed to be 60 hertz. Not approximately 60 hertz. Not close to 60 hertz. Exactly 60 hertz. The gap between the acceptable band and cascading blackouts is plus or minus 0.036 hertz—a tolerance smaller than most people imagine exists in any physical system, let alone one spanning from New Jersey to Illinois.

This is the fundamental thing almost nobody understands about the electrical grid: it has no warehouse. Unlike water, which can sit in a reservoir, or natural gas, which can be stored in underground caverns, electricity must be generated at the precise moment it is consumed. Every second of every day, generation must exactly match demand. If generation exceeds demand, frequency rises. If demand exceeds generation, frequency drops. Either direction, pushed far enough, triggers automatic shutdowns that can cascade across regions. The grid is not a battery. It is a continuous, real-time balancing act performed at continental scale.

This system was engineered for a world of slow, predictable growth. For most of the twentieth century, electricity demand in the United States increased at a gentle, linear rate—roughly 1–2% per year, driven by population growth and the steady electrification of daily life. Utilities planned decades ahead, and those plans almost always proved adequate. Between 2010 and 2020, demand actually declined by about 1%, according to the Department of Energy. Grid operators could be forgiven for assuming the age of dramatic demand growth was behind them.

They were wrong. And the thing that proved them wrong is, in part, the same technology producing the words you are reading right now.

The Physics of a System That Cannot Wait

To understand why artificial intelligence is straining the grid, you first have to understand how the grid actually works—not as an abstraction, but as a physical system governed by physics that do not negotiate.

An electrical grid has three layers. Generation is the first: power plants—coal, natural gas, nuclear, solar, wind—converting some other form of energy into electrical current. Transmission is the second: high-voltage power lines that carry electricity across long distances, stepping voltage up to hundreds of thousands of volts to minimize energy loss over the journey. Distribution is the third: the local infrastructure—substations, transformers, neighborhood power lines—that steps voltage back down and delivers it to your home, your office, or a data center in Northern Virginia.

Each layer depends on a class of equipment that most people have never thought about: transformers. These are the devices that convert electricity between voltage levels. A large power transformer at a substation can weigh 400,000 pounds, contain thousands of gallons of cooling oil, and cost several million dollars. They are, in a very real sense, the critical nodes of the grid. Every watt of electricity that moves from a power plant to your wall outlet passes through multiple transformers along the way.

Here is what makes the grid unlike nearly every other engineered system: it has almost no capacity to store its own product. When you flip a light switch, a generator somewhere must instantaneously produce the corresponding electricity. When you turn the light off, that generation must instantaneously decrease. This balancing happens through a set of automated controls and market mechanisms that coordinate thousands of generators across an interconnected system—a system designed, built, and refined over seven decades for a particular pattern of demand. That pattern was linear, dispersed, and predictable.

AI demand is none of those things.

When Exponential Demand Meets a Linear System

Consider a single data point that makes the abstract concrete: a query to an AI chatbot consumes roughly ten times more electricity than a conventional search engine query. Estimates vary—Epoch AI's 2025 analysis found that a GPT-4o query uses about 0.3 watt-hours, compared to roughly 0.03 watt-hours for a modern Google search—but the order-of-magnitude difference is consistent across independent analyses. A single query is trivial. Multiply it by the billions of AI queries processed daily, and the aggregate power draw becomes the energy consumption of a mid-sized country.

The numbers at the system level are staggering. According to the Lawrence Berkeley National Laboratory's 2024 United States Data Center Energy Usage Report (Shehabi et al., December 2024), U.S. data centers consumed approximately 176 terawatt-hours of electricity in 2023—about 4.4% of the nation's total annual electricity consumption. The same report projects that figure could reach 325 to 580 terawatt-hours by 2028, representing 6.7–12% of national consumption. The Electric Power Research Institute, in a March 2026 analysis cited by Harvard's Salata Institute, estimates data centers could consume between 9 and 17% of U.S. power by 2030—a trajectory far steeper than the same group projected just two years earlier.

But the raw numbers, alarming as they are, obscure the more dangerous problem: geography. AI data centers are not dispersed evenly across the country. They cluster. Northern Virginia alone hosts the largest concentration of data centers on the planet, constituting roughly 13% of all reported data center capacity globally, according to a Virginia Joint Legislative Audit and Review Commission report. Amazon, Microsoft, Google, and Meta collectively spent over $200 billion on capital expenditures in 2024—a 62% year-over-year increase—and a significant share of that spending is concentrated in a handful of regions: Northern Virginia, central Ohio, the Dallas–Fort Worth corridor, and a few others.

This geographic concentration creates demand hotspots that overwhelm local infrastructure. A single hyperscale campus in Loudoun or Prince William County, Virginia, can draw 30–60 megawatts of continuous power—orders of magnitude more than the industrial facilities the region's substations were originally designed to serve. When dozens of such campuses sit on the same transmission backbone, the strain is not additive. It is systemic.

The Night the Grid Learned What Concentration Means

On the evening of July 10, 2024, a thunderstorm rolled through Northern Virginia. A lightning arrestor—a piece of protective equipment on a 230-kilovolt transmission line—failed. The failure caused a series of brief voltage disturbances, each lasting only milliseconds. For residential customers, these micro-fluctuations would have been imperceptible. But data centers are not residential customers.

Data center servers are exquisitely sensitive to voltage deviations. Their uninterruptible power supply systems are programmed to react in milliseconds—faster than any human operator can intervene. When the voltage disturbances hit, approximately 60 data centers in the region automatically disconnected from the grid and switched to internal backup power. Simultaneously. According to a NERC incident report documented by Data Center Dynamics in March 2025 and subsequently analyzed by the Harvard Belfer Center (February 2026), the result was a sudden, unplanned loss of nearly 1,500 megawatts of demand—equivalent to roughly a third of all households in Virginia vanishing from the grid in an instant.

In grid physics, losing generation is the emergency most operators train for. Losing demand this abruptly is a different kind of crisis. With generation still running but load suddenly gone, frequency surged to 60.047 hertz—well above the target band, according to NERC's incident report, as confirmed by Grid Status's April 2025 analysis of the event data. Dominion Energy operators scrambled to remove capacitor banks from the local transmission network. The data centers, meanwhile, remained off-grid for hours; while switching to backup power is automatic, reconnecting to the grid requires manual intervention.

The Harvard Belfer Center's February 2026 report on AI and the U.S. electric grid called this incident a watershed moment. It demonstrated that the same features that make data centers resilient—instantaneous failover to backup power—can, when deployed at scale and in geographic concentration, become a source of systemic grid instability. The protection systems across dozens of facilities were programmed nearly identically. When one sensed a fault, they all reacted together. Individual reliability became collective fragility.

Why You Cannot Just Build More

The instinctive response to surging demand is straightforward: build more. More power plants, more transmission lines, more substations. But the grid's expansion is constrained by a chain of bottlenecks, each one longer than the last, and the most critical is the one most people have never heard of.

Start with transformers. Large power transformers—the kind that sit at substations and handle the conversion between transmission and distribution voltages—are among the most complex pieces of industrial equipment in existence. They are custom-engineered for specific applications. Their cores require grain-oriented electrical steel, a specialty material produced by exactly one domestic manufacturer in the United States, according to a Department of Energy supply chain assessment. The United States imports roughly 80% of its power transformers.

Lead times for large power transformers now average 128 weeks—approximately two and a half years—according to a 2025 survey by Wood Mackenzie. Generator step-up transformers average 144 weeks. Demand for these units has surged 116% and 274%, respectively, since 2019. The supply deficit for power transformers stands at roughly 30%. Prices have climbed 77% in the same period. More than half of the approximately 40 million distribution transformers currently in service in the United States are over 33 years old, approaching or exceeding their designed lifespan.

The transformer bottleneck alone would be sufficient to slow grid expansion. But it is only the first link in the chain. Transmission lines—the high-voltage infrastructure needed to move power from where it is generated to where data centers need it—require a decade or more of planning, permitting, environmental review, and construction. An interconnection queue study found that gigawatts of ready-to-build generation projects are waiting years for grid connection. You can construct a solar farm in months. The transmission line to connect it to the grid may take ten years.

And beneath all of this sits a more fundamental constraint: approximately 70% of the U.S. grid's infrastructure is approaching the end of its designed life cycle, according to industry analysts. Most of the grid was built between the 1950s and 1970s. The system does not merely need expansion to accommodate new demand. It needs replacement of the aging infrastructure that serves existing demand—at the same time.

The Accelerating Machine

The demand problem is compounding because the hardware driving it is evolving faster than the infrastructure meant to support it. The average power draw per server rack in a data center has undergone a transformation that would have been difficult to predict even three years ago.

In 2021, a typical data center rack consumed roughly 8 kilowatts. By 2024, that average had climbed to approximately 17 kilowatts. NVIDIA's GB200 NVL72 system—a rack-scale AI computing platform containing 72 Blackwell GPUs—consumes 120 to 140 kilowatts per rack. That is a single rack drawing as much power as approximately 40 American homes. Most existing data centers were designed to handle 15–30 kilowatts per rack. Facilities capable of supporting 120-kilowatt racks essentially require new construction from the ground up. And NVIDIA's own roadmap suggests the next generation—the Rubin platform, expected in 2027—could push rack-level power consumption into the 400 to 600 kilowatt range, with some industry analysts projecting even higher densities for successor platforms.

This is exponential growth hitting a system designed for linearity. Each new generation of AI hardware demands more power per unit of compute, deployed at greater density, in facilities concentrated in the same geographic clusters. The grid was not engineered for this demand profile. Nothing was.

Who Pays for the Machine

In Granville, Ohio—a small town outside Columbus—a retired engineer named Ken Apacki keeps a spreadsheet. Every month for the past five years, he has recorded every charge on his electric bill. In July 2020, he and his wife Carol paid between 11 and 12 cents per kilowatt-hour for electricity. By 2025, they were paying 19 cents—a 60% increase, as NPR Planet Money reporter Keith Romer documented in a January 2, 2026 segment. Carol suspects the price increase is connected to the 130 data centers that have appeared in central Ohio.

She is largely correct, though the mechanism is more structural than most people realize. An electric bill has three components: distribution (local power lines to your home), transmission (high-voltage lines carrying power across distances), and generation (the power plants themselves). Data center construction drives up costs across all three. Local utilities must expand distribution infrastructure to deliver power to data centers, and those costs are socialized across all customers. Transmission companies spend billions building out the grid for data centers, and those costs filter down to residential ratepayers. And the surge in demand pushes generation prices upward for everyone in the wholesale market, because supply cannot ramp fast enough to match the new load.

The structural irony runs deeper. In many states, data centers receive substantial tax incentives and discounted electricity rates designed to attract their investment. Virginia's data center industry has benefited from sales and use tax exemptions on servers, cooling equipment, and backup energy hardware. Meanwhile, Dominion Energy proposed its first base-rate increase since 1992 in February 2025, adding approximately $8.51 per month to a typical household's bill in 2026, according to the Harvard Belfer Center report. The pattern is clear: the entities generating the demand receive incentives to consume, while the costs of accommodating that demand are distributed across all ratepayers. A retired couple in Ohio is subsidizing the infrastructure that powers trillion-dollar companies.

Cathy Kunkel, an analyst at the Institute for Energy Economics and Financial Analysis, put it plainly in the NPR report: ordinary people will almost inevitably end up subsidizing the wealthiest industry in the world—unless the rules governing cost allocation are reformed to require data centers to bear a proportional share of the infrastructure costs they generate. PJM's capacity market prices have spiked nearly tenfold, driven in part by this demand surge, and that region faces a projected reliability shortfall of 6 gigawatts by 2027. That gap is equivalent to six large natural gas plants that do not yet exist.

The Other Consumption

Electricity is not the only resource the grid's new tenants consume at industrial scale. Data centers require enormous quantities of water for cooling—pumping cold water through pipes to absorb heat generated by thousands of servers, then venting the heated water as steam in evaporative cooling systems. The numbers are rarely discussed in the same breath as energy consumption, but they are comparably striking.

In 2023, Google's data centers consumed approximately 6.1 billion gallons of water globally, according to the company's own environmental report—and a single Google facility in Council Bluffs, Iowa, accounted for 1 billion of those gallons in 2024, enough to supply all of Iowa's residential water demand for five days. The Lawrence Berkeley National Laboratory estimated that U.S. data centers consumed 17 billion gallons of water directly through cooling in 2023, with projections indicating that figure could double or quadruple by 2028.

Some of the fastest-growing data center markets sit in regions already experiencing water stress. Maricopa County, Arizona, is facing extreme drought conditions while simultaneously becoming a hub for hyperscale data center development. Northern Virginia's data centers collectively consumed close to 2 billion gallons of water in 2023, a 63% increase from 2019. The irony sharpens: several of the companies driving this water consumption had previously pledged to be "water positive" by 2030—compensating for their usage by returning high-quality water to local systems. Those pledges are looking increasingly difficult to fulfill as consumption scales with AI demand.

The Invisible Cost

There is a particular kind of engineering elegance in a system designed to be invisible. The electrical grid's genius is that you never think about it. You flip a switch, and the lights come on. You type a query, and an answer appears. The distance between the action and the infrastructure that enables it—the generating stations, the transformers, the transmission lines, the cooling systems, the water, the retired couple's rising electric bill—is obscured by design.

That invisibility was always the point. The grid was engineered so that the complexity of real-time continental-scale power balancing would be entirely abstracted away from the end user. But abstraction is not the same as absence. The physical costs persist whether you see them or not. Every query carries a weight in watts, a volume in water, a line item on someone's electric bill. Transformers built for linear growth are being asked to handle exponential load. Substations engineered for factories are absorbing demand densities that behave more like small cities switching on and off at once. None of this was an oversight. The system worked as designed; the design simply did not anticipate software as its largest industrial customer.

None of the constraints described here are mysteries. Transformer lead times are scheduled. Transmission permitting is scheduled. The end-of-life replacement curve for 70% of the existing grid is scheduled. The only variable that has moved faster than expected is the demand, and it has moved in a direction the system was not built to absorb. That leaves two questions that the physics cannot answer and the engineering will not decide. Who pays for the rebuild. And who gets their power first when the rebuild is not yet finished.