The Grid Will Have Its Say

In a 2025 report titled AI’s Power Requirements Under Exponential Growth, the RAND Corporation examined what continued scaling of frontier artificial intelligence systems would mean in physical terms. Their conclusions were expressed plainly, in gigawatts.

“Training could demand up to 1 GW in a single location by 2028 and 8 GW – equivalent to eight nuclear reactors – by 2030, assuming current training compute scaling trends.”

For those not initiated in the mathematics of energy, as I wasn’t before writing this piece, let me slow that down. Eight gigawatts is eight billion watts of continuous electrical draw. It is roughly the amount of electricity being consumed by a city the size of Chicago at peak demand; 3:00 p.m. on a Wednesday in August, temperatures pressing past 100 degrees, millions of air conditioners running at full blast across apartments, office towers, hospitals, and schools; commuter trains pulling power from overhead lines; factories humming; water treatment plants cycling; traffic lights blinking across hundreds of intersections; a workforce of millions typing, welding, transporting, trading, cooling, manufacturing, and coordinating. That surge, the moment when the grid strains hardest to keep voltage stable and frequency steady, is measured in gigawatts. RAND is suggesting that a single AI training campus could soon require that much power on its own.

But that figure is only referring to training, the episodic, compute intensive runs that build frontier models. Inference, the daily operation of those models once deployed across search engines, enterprise software, logistics systems, financial markets, and military platforms, creates a more diffuse but permanent layer of electrical demand. The IEA and Lawrence Berkeley National Laboratory project that AI and data center growth will add between 200 and 500 terawatt-hours of annual U.S. demand by 2030 alone, a surge layered on top of a grid that was already being asked to grow.

To give that scale some grounding, 200 terawatt-hours is roughly the annual electricity consumption of New York State. Five hundred terawatt-hours approaches the annual electricity consumption of South Korea, one of the most technologically advanced industrialized economies in the world. The United States has not absorbed a demand expansion of that magnitude in modern grid history without decades of coordinated generation and transportation buildout. For much of the past two decades, national electricity consumption was effectively flat. Utilities planned for incremental growth, renewable integration, and steady electrification. They did not plan for nation-state scale loads appearing in clusters over five-year windows.

Artificial intelligence adoption curves may scale exponentially according to investor decks and research papers. The grid does not scale in the same way. It expands through turbine manufacturing schedules, transformer lead times, environmental reviews, right-of-way disputes, interconnection studies, and reserve margin calculations that assume something closer to linear growth. We are therefore living inside a divergence. Point A is the electrical system we actually have; synchronized, high voltage, thermodynamically constrained, and built for incremental expansion. Point B is the electrical system implicitly assumed by frontier AI developers: elastic, rapidly scalable, capable of delivering hundreds of terawatt-hours of new demand without delay. The gap between Point A and Point B is not ideological. It is physical. And it is that gap, not model capability, that will ultimately shape the trajectory of artificial intelligence adoption.

Power, Voltage, Heat, Balance

Before we go any further, we need three terms: power, voltage, and current.

Power is what we care about. It is the thing data centers consume. It runs servers, spins cooling systems, moves pumps, lights buildings. When we say a facility requires one gigawatt, we are talking about power. That is the prize.

Power is the product of two things: voltage and current.

Voltage is electrical pressure. Current is the flow of electrical charge. Increase either and power increases. Increase both and it increases faster.

So if we want more power, why don’t we just crank everything up? Because electricity comes with two constraints. The first is heat. When current flows through a conductor, it generates heat, and not in a gentle way. Heat increases with the square of current. Double the current and you roughly quadruple the heat. Triple it and you get nine times as much. Push too much current through a transmission line, transformer, or busbar and equipment degrades or fails.

The grid’s longstanding solution to heat has been to move power at very high voltage. Higher voltage allows larger amounts of power to travel with comparatively lower current. Lower current means less heating. That is why transmission lines operate at hundreds of thousands of volts and why substations exist to “step” voltage up and down. Thermal management.

The second constraint is balance. Electricity, at grid scale, cannot sit in a tank until its needed. It is produced and consumed at the same moment. Every second, generation must match demand exactly. Too little supply and protective systems begin to trip equipment offline. Too much and the system destabilizes in the opposite direction. Either imbalance has the capacity to cause cascading failure. The grid operates inside that narrow equilibrium continuously. It is one of the most complex coordination systems ever built; generation rising and falling across vast distances to meet demand in the same instant it appears.

Data centers do not care about the elegance of this arrangement. They care about receiving enormous amounts of power. The equipment delivering that power, however, is constrained by heat and by the need to maintain balance across the system. So when we talk about adding gigawatts of load, we are not just talking about demand curves. We are talking about thicker conductors, larger transformers, upgraded substations, expanded generation, additional reserves, and transmission corridors capable of handling both thermal stress and real-time volatility. AI may scale according to ambition and capital. The grid scales according to thermodynamics and balance.

American Electrification

The American grid did not emerge fully formed. It evolved slowly, often unevenly, in response to the physical constraints we have just outlined above. In the late nineteenth century, electricity was local. Early generating stations served a handful of city blocks. Power was produced near where it was consumed because there was no efficient way to move it over long distances. Direct current systems, championed by Thomas Edison, operated at largely low voltages, because the high voltage systems of the time tended to kill people. They worked, but only within tight geographical limits. Push current too far at low voltage and an increasing share of the energy is converted into heat rather than delivered as usable power. Distance was therefore expensive, and scale was constrained.

The breakthrough came with the development of alternating current. Unlike direct current, which flows in a single direction, alternating current reverses direction many times per second. The mechanics matter less than what this made possible. The oscillation of alternating current allowed engineers to build transformers, devices made of wire coils wrapped around magnetic cores, which were capable of raising or lowering voltage efficiently and without using moving parts. The ability to “step” voltage up and down solved the central transmission problem. Power could be generated at one voltage, raised dramatically for long-distance travel away from human hands, and then lowered again near homes, factories, and businesses to levels that machinery, and people, could safely handle. High voltage made distance economical, lower voltage made electricity usable.

That flexibility changed everything. Generation no longer had to sit next to consumption. Hydroelectric dams, coal plants, and later nuclear stations could be located where geography and fuel made sense, while cities, many built long before large-scale electrification, and rarely eager to sit beside smokestacks, could draw power from miles away. Industry could cluster around ports and rail lines; urban life could expand without tethering itself to the nearest turbine. What began as small, local systems evolved into interconnected networks spanning states and regions, and by the mid-twentieth century had merged into a continental grid. That grid emerged from an industrial economy in which growth was physical and therefore time-bound; homes needed to be framed, factory floors needed to be poured, and railway lines needed to be laid before anything could be electrified. Expansion was constrained by the pace of construction, and the grid expanded at that same speed, a tempo that became embedded into its architecture.

For decades, that tempo held. Load increased, but in patterns utilities could anticipate; population growth, suburbanization, industrial expansion and contraction, electrified appliances, air conditioning. Generation was added in large blocks consisting of coal units, nuclear plants, and later combined-cycle gas turbines. And transmission expanded gradually to connect them.

Point A

Before we speak of where artificial intelligence intends to take the grid, before we sketch Point B with its promises of scale and acceleration, we must sit for a moment with what already exists. The system described above is the machine we have inherited. What began as small municipal systems now delivers roughly 4,000 terawatt-hours of electricity each year across the United States. That scale was not achieved through sudden leaps, but through decades of incremental expansion and disciplined planning.

Load has risen and fallen with population growth, suburbanization, industrial cycles, and electrified appliances, but largely within bands utilities could anticipate. Generation has been added in large blocks and transmission expanded gradually to connect them. Utilities forecast years in advance, model contingencies, and adjust reserve margins with care. The architecture of the grid reflects that expectation; steady movement rather than abrupt transformation.

Generation itself has diversified, but not infinitely flexible. Natural gas plants now provide the largest share of electricity, valued for their relative speed in ramping output up or down. Nuclear facilities, by contrast, are designed to operate continuously, their economies dependent on steady baseload production. Coal stations remain embedded in certain regions. Wind and solar have grown rapidly, particularly across the Midwest and Southwest, adding variability that must be balanced elsewhere on the system. Each resource contributes, but each does so with constraints; physical, political, geographic.

Transmission, the long steel lattice that binds these resources into a single network, expands more slowly. High voltage lines are capital intensive, politically sensitive, and procedurally dense. Siting decisions stretch across jurisdictions. Environmental reviews take years. Land must be acquired, rights negotiated, corridors cleared. A single project can require a decade from conception to energization. Large power transformers, essential for stepping voltage up and down across the system, are custom-built pieces of equipment often sourced globally, with lead times measured not in months but in years. Meanwhile, interconnection queues swell, as proposed generation projects await study and approval before they may connect to the grid’s existing backbone, a reminder that generation is only as useful as the wires that carry it.

And yet, despite these frictions, the system functions. Reserve margins are maintained. Operators balance supply and demand in real time. Transmission corridors operate near thermal limits during peak periods but remain within them. Extreme weather events test reliability, sometimes severely, but the machine endures. It was built for stability under gradual expansion; for factories added one by one, for suburbs extending mile by mile, for air conditioning, and refrigeration, and incremental industrial growth distributed across millions of customers.

Point B

Now consider what is being asked of that system.

Artificial intelligence is not a marginal load added quietly at the periphery of the network. It is industrial demand at a scale and concentration the grid was never designed to quickly absorb. Frontier training clusters are already measured in the hundreds of megawatts. Projections now speak openly of single campuses approaching one gigawatt in the near term, and several gigawatts within the decade. Eight gigawatts, the scale RAND describes as plausible by 2030, is the electrical profile of a major metropolitan region redirected toward one fenced in industrial complex.

Training is only the most visible layer. Inference extends the load outward and makes it permanent. Search queries, document drafting, logistics optimization, fraud detection, image generation, drug discovery, battlefield analysis; each inference call is small in isolation, but trivial calls multiplied across hundreds of millions of users become persistent base demand. Data centers that once served storage, cloud computing, and search are being retrofitted with dense racks of accelerators designed to operate at high utilization around the clock. Facilities are expanded. New campuses are announced in rapid succession. The buildout is not incremental, it is structural.

What distinguishes this demand from traditional industrial growth is not merely its size but its concentration and tempo. A steel mill may draw hundreds of megawatts, but such facilities are geographically dispersed and historically added over decades. AI campuses, by contrast, cluster rationally around existing transmission nodes, fiber corridors, and labor pools where capacity already exists. Those nodes, however, were built for prior demand patterns, not for repeated multi-gigawatt additions on compressed timelines. These facilities require high power reliability and have minimal tolerance for interruption. These loads do not idle at night or retreat with the seasons. The hardware inside them costs billions and only generates value when it is running, which makes continuous operation an economic necessity.

The cadence of AI deployment differs fundamentally from the cadence of grid expansion. The system of Point A was shaped by caution and sequence; permit, finance, build, connect. The digital economy operates differently; iterate, deploy, scale. Model performance improves quarterly. Compute requirements rise accordingly. Chips are ordered years in advance. Capital is committed before revenue is proven. Hyperscalers announce new facilities on timelines measured in months. The governing assumption beneath that tempo is straightforward; if the models are ready and the severs are installed, the electricity will be there. That assumption, that physical system will scale to meet digital ambition without delay is Point B.

The Electron Gap

What matters now is not the current scale of the grid, but the delta between where demand is headed and what the system can realistically supply. Even before AI entered the picture, electrification was poised to push the baseline higher. EV adoption, heat pumps, industrial electrification, and data driven manufacturing are projected to add between 1,500 and 2,400 terawatt-hours of demand annually by mid-century, depending on adoption rates.

AI accelerates that trajectory and compresses its timeline. If gigawatt-scale campuses operate continuously across multiple regions, and enterprise inference becomes embedded across major sectors, incremental annual demand could rise by 200 to 500 terawatt-hours by 2030 alone. This means our 4,000 terawatt-hour system must expand not only to accommodate long-run electrification, but to absorb a near-term surge layered on top of it. The distance between what is being demanded and what can be built in time is the electron gap.

Let’s run the numbers. The required expansion above what the grid can realistically supply could approach 370 terawatt-hours by 2030 and potentially 1,900 terawatts-hours by mid-century. 370 terawatt-hours approaches the annual electricity consumption of the United Kingdom. 1,900 is roughly equivalent to that of India. Even at the lower end, this is not marginal growth, it is the addition of a major industrial economy to the system within a handful of years.

Meeting that requirement means adding generation capacity, but capacity additions unfold through industrial supply chains that are neither frictionless nor infinitely elastic. Utility-scale gas turbines are manufactured in a limited number of facilities worldwide, with multi-year backlogs that lengthen as orders accumulate. Nuclear construction involves specialized components, heavy forging, regulatory review, and workforce constraints that make dramatic schedule compression unlikely even under favorable policy conditions. Renewable generation can be installed more quickly once permitted, but utility-scale wind and solar projects still require land acquisition, environmental review, financing, and grid impact studies before construction begins.

Transmission presents an additional layer of complexity. High-voltage lines must move large quantities of power over distances while keeping current within thermal limits. As stated above, when current rises, heating increases with its square, which constrains how much additional load existing conductors can safely carry. Expanding transmission therefore means more than stringing new wire; it often requires taller towers, wider rights-of-way, upgraded substations, and in some cases reconductoring with higher capacity materials. These materials are not generic commodities. High voltage conductors, large transformers, and associated switchgear are specialized industrial products with limited manufacturing capacity. Lead times for large power transformers frequently extend beyond two years, and domestic manufacturing capacity is limited after decades of modest grid growth and cannot be expanded overnight. Even where equipment is available, siting new corridors requires negotiation across counties and states, engagement with landowners, environmental assessments under federal and state law, and, increasingly, litigation. Projects that appear straightforward on engineering diagrams often spend a decade navigating procedural steps before energization.

Even where generation and transmission upgrades are feasible, connection is not automatic. Any new generator or large load seeking connection must undergo a series of studies to determine its impact on voltage stability, fault levels, system balancing, and congestion across the broader network. These studies proceed sequentially; one project’s required upgrade can alter the assumptions for the next project in line, forcing re-analysis. If reinforcement is required, a new substation for example, or upgraded protection systems or transmission capacity, those upgrades must be engineered, cost-allocated, financed, and constructed before the project proceeds. In many regions, queues now contain years’ worth of proposed capacity awaiting review. The bottleneck, despite what some hyperscalers may say, is not run-of-the-mill bureaucratic delay, it’s the cumulative engineering work required to ensure reliability across a synchronized system operating in real time.

These processes were built to accommodate gradual expansion. They rely on engineering review, material supply chains, regulatory approval, and synchronized planning across jurisdictions. Expanding national electricity output by 370 to 1,900 terawatt-hours requires more than capital commitments or announced projects; it requires equipment to be manufactured, corridors to be approved, substations to be built, and systems to be tested before they can operate within a tightly balanced network. The electron gap is the difference between what aggressive AI-driven demand, layered on top of robust national electrification, would require and what the grid can realistically deliver; a shortfall of about 370 terawatt-hours by 2030, which widens to nearly 1,900 terawatt-hours by mid-century.

Capital Under Constraint

AI firms today are some of the most capital-intensive enterprises to ever be assembled. Billions of dollars have been committed to specialized accelerators, hyperscale campuses, multi-year power purchase agreements, transmission upgrades, and long-term land positions. The largest firms are underwriting construction programs that resemble full industrial buildouts, traditionally the prevue of governments and conglomerates, not software startups. At the same time, revenue models remain in flux. Adoption is expanding, but monetization per unit of compute remains uncertain. Valuations are divorced from present cash flow, responding instead to adoption curves, projections that often assume scaling can proceed without interruption.

Capital markets are tolerant of heavy upfront investment when scale appears inevitable and timelines feel compressible. Perhaps that is why we continually hear AI evangelists predicting full AI adoption within twelve months. Capital markets are less tolerant when bottlenecks emerge that cannot be resolved by writing larger checks. Debt taken on to finance campuses assumes utilization. Equity valuations assume forward revenue expansion tied to that utilization. If deployment slows because electrons are scarce or expensive, cash burn extends while revenue realization shifts outward. In that environment the weakest balance sheets feel pressure first. And failure at the margin affects perception at the center.

When smaller or overextended firms falter, unable to secure sufficient power, unable to sustain burn while waiting for interconnection, unable to justify capex against delayed monetization, the signal travels. Valuations adjust. Discount rates rise. Capital becomes more selective. Investment committees that once funded expansion on the assumption of immediate scale begin to demand clearer paths to profitability under constrained conditions. In a sector that has absorbed extraordinary levels of funding, even moderate repricing can reverberate widely.

Electricity scarcity introduces another dynamic. In competitive markets, power is allocated through price and access, and when concentrated AI load begins pressing against transmission and generation limits in specific regions, costs tend to rise locally before new capacity can be built. In that environment, AI firms are not only competing with one another for model performance and customers; they are competing with a local population for electrons coming from the same grid. Utilities, under reliability obligations and political pressure, may prioritize existing customers. Faced with uncertain access or escalating prices, firms may seek greater control over their energy supply. Long-term power purchase agreements are already common, and co-location with generation assets is increasing. Beyond that lies deeper verticalization: underwriting dedicated generation by financing gas turbines, small modular reactors, or large renewable installations paired with storage to supply specific campuses. Under scarcity, control over electrons becomes a strategic asset rather than a background utility service.

That strategy offers control, but it comes with costs. Modern electrification achieved scale by pooling fixed costs across millions of ratepayers. Large plants fed wide transmission networks, and reserve margins were shared across regions. A gigawatt scale campus that finances its own generation internalizes those fixed costs. The balance sheet absorbs the construction risk, fuel price exposure, regulatory uncertainty, and maintenance obligations that utilities have historically managed within regulated frameworks. The capital requirements alone are substantial. A single combined-cycle gas plant can cost more than a billion dollars. Multiply that across multiple campuses and later in transmission reinforcement and storage, and the financial profile begins to resemble heavy industry. Returns must justify not only silicon and talent, but turbines and steel.

There are additional tradeoffs. Dedicated generation reduces exposure to regional scarcity but concentrates operational risk. Weather events, fuel supply disruptions, or local transmission constraints affect larger shares of compute capacity. Political scrutiny intensifies when highly visible corporate campuses secure private power while residential consumers face rising rates. Vertical integration might be rational in a world of scarce electrons, but it is also expensive. It shifts fixed costs from regulated utilities onto corporate balance sheets and concentrates risks that large, interconnected grids were designed to diffuse. For more than a century, electrification worked because scale lowered cost and shared uncertainty. Large plants fed wide networks, reserve margins were pooled, capex were amortized across millions of customers. That system’s efficiency rested on aggregation.

A turn toward self-supplied power marks a departure from that logic. It implies that pooling can no longer be relied upon, that access to shared infrastructure may not keep pace with industrial ambition. In effect, some of the most capital-rich firms in the world would be rebuilding elements of the electrical system for their own use, an inversion of the model that made modern economic growth reliable and relatively inexpensive to power. The scale of capital being committed to AI is extraordinary. To layer onto it the costs of generation, transmission, and storage is to bet not only on model capability, but on the willingness to abandon the shared efficiencies that have underwritten industrial expansion for nearly a century.

None of this implies collapse. It implies correction. When infrastructure imposes discipline, industries reorganize around constraint. The early internet did not disappear after the dot-com bubble burst, it adapted, converging on revenue models based on surveillance and advertising. AI will also need to adapt. If the electron gap widens, efficiency will cease to be a technical afterthought and become a strategic priority. Utilization rates will be scrutinized. Geographic siting decisions will weigh energy access as heavily as tax incentives or labor pools. Some firms will find ways to secure stable power and integrate it into their long-term cost structures. Others, particularly those whose expansion depended on uninterrupted scaling, may consolidate or fail as capital becomes more selective.

In that environment, the trajectory for AI will be shaped as much by energy economics as by algorithmic breakthroughs. Model capability will continue to improve. But the pace and distribution of deployment will reflect the cost, timing, and availability of electricity. The electrons do not determine whether innovation occurs; they determine how far and how fast it travels.

Numbers to Watch

If the electron gap is emerging, it will show up in permitting timelines, transformer backlogs, queue lengths, and price signals. The transition from Point A to Point B can be tracked.

Start with generation. New capacity additions matter less in press releases than in interconnection approvals. Watch how many gigawatts actually clear regional transmission organization queues each year, not how many are proposed. Proposed capacity can reach into the thousands of gigawatts, cleared capacity is what counts. Additionally, watch annual net additions of firm generation; natural gas plants, nuclear restarts, long-duration storage, not just nameplate renewable announcements. Intermittent capacity adds megawatts on paper; firm capacity determines how much load can operate continuously.

Transmission is even more telling. High-voltage line milage permitted and completed each year is a leading indicator of the grid’s ability to absorb concentrated industrial load. If transmission expansion continues at a 20th century pace while large scale data center projects accelerate, the electron gap widens. Transformer lead times are another signal. Large power transformers are specialized, custom-built pieces of equipment with global supply chains. If delivery timelines extend further, the pace of grid expansion slows regardless of capital availability.

The third category is simpler; price. Sustained increases in wholesale electricity prices in regions with heavy data center development indicate congestion or capacity scarcity. Capacity market clearing prices, where they exist, are especially informative. Rising capacity prices signal tightening reserve margins.

The fourth category is geography. Track where hyperscale AI campuses are announced versus where generation and transmission capacity are expanding. If projects cluster heavily in a handful of regions, concentration risk rises. If buildouts begin to disperse geographically, that suggests constraint is shaping deployment decisions.

Finally, watch power purchase agreements signed by AI firms themselves. Long-dated, large-scale contracts with dedicated generation providers indicate vertical integration pressure; firms securing electrons directly rather than relying on grid availability alone.

None of these metrics independently confirms the electron gap, but taken together they reveal whether infrastructure is scaling in step with ambition. The arithmetic is simple. The indicators above show whether it is being reconciled in practice.

The Grid Will Have Its Say

When hyperscalers speak of the need for “absurd” amounts of energy, they are acknowledging something fundamental. AI does not run on venture capital, valuation multiples, or benchmark improvements. It runs on electricity, and electricity is governed by physical laws, industrial supply chains, and regulatory process.

The grid does not respond to earnings calls. It does not accelerate because capital is impatient. It expands when turbines are manufactured, when substations are upgraded, when high-voltage lines clear years of permitting and litigation, when transformers arrive after long production queues. Its pace is determined by steel, copper, concrete, and coordination across jurisdictions. That pace can increase, but it cannot compound at software speed.

The electron gap describes a measurable expansion requirement; roughly 370 terawatt-hours by 2030, widening to nearly 1,900 terawatt-hours by mid-century, layered onto a system that was not designed for abrupt growth. Closing that gap depends on how quickly generation, transmission, and associated infrastructure can be permitted, financed, manufactured, and brought online without undermining reliability. Where expansion keeps pace, deployment can proceed with fewer financial distortions. Where it does not, rising costs, delayed energization, and tighter reserve margins will feed directly into balance sheets and investment decisions, reshaping the competitive landscape in ways that extend beyond customer acquisition to the underlying availability of power.

None of this need halt innovation necessarily. Models will improve. Chips will become more efficient. Engineers will find ways to extract more output per watt. But the trajectory of deployment will be shaped by energy economics as surely as by algorithmic progress. The twentieth century built an electrical system that enabled industrial expansion by pooling cost and spreading risk across millions of users. The twenty-first century now asks that system to accommodate industrial-scale computation at unprecedented concentration. Whether that transition unfolds smoothly or under strain depends less on ambition than on infrastructure. Ambition can compound. Steel cannot.

AI may redraw the boundaries of knowledge and productivity. But it will do so inside a physical system that answers to thermodynamics, supply chains, and time. The grid does not negotiate. It balances. And in the years ahead, it will have its say.