Big Tech’s race for dominance in artificial intelligence has collided with an old constraint. Electricity. Data centers that train and run the latest models now threaten to double their draw on the world’s power supplies by 2030. Shortages loom. Prices rise. Grids strain.
The numbers tell a stark story. Global data center electricity consumption climbed 17% in 2025, according to the International Energy Agency. AI-focused facilities grew even faster, up 50%. By 2030 the IEA projects total data center demand will reach 950 terawatt-hours. That’s roughly 3% of global electricity. AI-specific load could triple in the same period.
The Scale of the Surge
Servers the size of a refrigerator will soon pull as much juice as dozens of homes. One rack in an advanced AI data center could demand peak power equal to 65 households by 2027, the IEA warns. Individual campuses now eye gigawatts of supply. A single proposed site in Texas seeks 5 gigawatts. Another in Louisiana matches it. These aren’t outliers. They signal a structural shift.
Hyperscalers have responded with capital on a historic scale. The five largest technology companies poured more than $400 billion into data centers and related infrastructure in 2025. Plans call for a 75% jump in 2026, pushing annual spending above $700 billion. That figure exceeds total annual U.S. energy sector investment. And much of it flows straight into facilities that will consume power around the clock.
But supply hasn’t kept pace. In the United States, data centers could claim between 9.5% and 15.3% of national electricity by 2030, a Department of Energy report updated in 2026 finds. The reference case lands at 11.8%, or 649 TWh. Other analysts paint an even sharper picture. A Business Insider review of permits shows potential annual consumption jumping 50% from prior levels if all projects materialize. At the midpoint that exceeds electricity use of any U.S. state except Texas.
Regional grids feel the pinch first. PJM Interconnection, which serves 67 million people across 13 states and Washington, D.C., saw wholesale power costs surge 76% in early 2026. Data centers drove much of that increase, according to the grid’s independent market monitor. Its latest capacity auction cleared at the price cap, short of reliability needs. Similar pressures grip ERCOT in Texas, where data center interconnection requests once topped 474 gigawatts, nearly five times the grid’s record peak demand.
Transmission lines represent another choke point. New large loads can connect in two to five years. Upgrading the grid to support them often takes a decade or more. The mismatch has forced emergency measures. In January 2026 the Department of Energy authorized certain data centers to switch to backup generators during winter peaks to protect residential service. Ireland has restricted new connections in Dublin. The Netherlands limits where such facilities can locate. These steps mark the first cracks in what could become broader rationing.
Power density compounds the headache. AI servers now dominate new shipments. Gartner forecasts they will consume more electricity than all conventional data center hardware combined by 2027. Global data center power demand is set to climb 26% in 2026 alone to 565 TWh, the research firm projects. TrendForce puts the rise in power demand capacity at 31% that year, with AI servers accounting for over a third. Efficiency gains per task have been remarkable. Yet surging usage, more users and more intensive applications such as AI agents have overwhelmed those improvements.
So hyperscalers hunt for firm, always-on supply. Natural gas turbines have emerged as the default near-term choice. Many projects now include on-site generation, sometimes called behind-the-meter power. Microsoft has signed more than 5 gigawatts of such capacity this year. Google, once reluctant, is installing aeroderivative turbines at a major campus. OpenAI prepares a 1.4-gigawatt facility in Texas backed by hundreds of gas engines. These arrangements bypass congested grids but raise emissions questions and local opposition.
Nuclear power draws fresh interest. Constellation Energy plans to restart Three Mile Island’s Unit 1, now renamed, under a 20-year deal to supply Microsoft data centers directly. The pipeline for small modular reactor commitments tied to data centers has grown from 25 gigawatts at the end of 2024 to 45 gigawatts. Yet new nuclear plants still face long timelines. Restarts help. Full new builds do not arrive fast enough for the current wave of demand. One analysis from Moody’s Ratings calculates the U.S. will need $110 billion in new generation through 2030 to meet data center growth. Over 30 gigawatts of that is likely gas-fired.
Financial engineering has kept some costs off balance sheets. A Financial Times investigation revealed Big Tech uses guarantees and other structures to hold roughly $300 billion of AI-related exposure away from their books. That tactic preserves credit ratings while funding aggressive buildouts. It also obscures the full picture for investors and regulators.
Markets have begun to price the strain. Capacity prices in constrained regions have multiplied. Ratepayers foot part of the bill. PJM’s recent auction is projected to add billions to ordinary electricity costs. Public pushback grows as residents face higher bills, water consumption concerns from cooling systems and noise from backup generators. Some projects have been canceled. Others delayed. Roughly half the data center capacity slated for 2026 in the U.S. may not come online on schedule, multiple reports indicate.
Yet the demand signal shows no sign of easing. Major model providers reported a threefold rise in active users and fivefold revenue growth over the past year. AI agents, multimodal models and enterprise adoption all point higher. Forecasts continue to be revised upward. Goldman Sachs, BloombergNEF and others have lifted their projections for data center power demand in recent months. The IEA’s latest update, released this week, largely holds its prior trajectory but stresses the need for better data from tech companies on actual consumption.
Searching for Balance
Some see opportunity amid the pressure. AI could optimize energy systems, accelerate discovery of new materials for batteries or fission, and improve grid management. The IEA notes efficiency gains per AI task have improved at rates rarely seen in energy history. If those continue and if new applications don’t explode usage further, the trajectory could bend. Real-time monitoring, frequent forecast updates and cooperation with the technology sector will prove essential.
For now the tension dominates. Data centers are being built faster than the infrastructure to power them. That gap will define which companies secure compute capacity, which regions attract investment and which technologies win the next round of energy infrastructure. Gas buys time. Nuclear offers cleaner firmness but arrives late. Renewables scale quickly yet require storage and firming to match always-on AI loads. Behind-the-meter solutions proliferate. Each choice carries trade-offs in cost, carbon and reliability.
The hyperscalers hold the purse strings. Their capital decisions will shape power markets for the next decade. Utilities, regulators and generators scramble to respond. Ordinary consumers watch their bills. And the AI models at the heart of it all keep getting larger, more capable and more power-hungry. The question isn’t whether electricity will limit artificial intelligence. It’s which parts of the economy will sacrifice to keep the models running.