Where does the electricity for AI come from?

There’s also a shortage of workers. The U.S. data center construction sector faces a labor gap of about 650,000 people. The Bureau of Labor Statistics projects an average annual shortfall of roughly 81,000 electricians between 2024 and 2034. Nationwide, there are approximately 720,000 licensed electricians, with an average age of 44; over 200,000 are expected to retire within the next decade, while only 15,000–20,000 new entrants join the profession each year. Chips can advance generation after generation annually—but transformers cannot, high-voltage transmission lines cannot, and electricians certainly cannot. The dot-com bubble hit a financial wall: once funding dried up, everything collapsed. This time, before hitting that financial wall, we’re first hitting a physical one. And there’s something even more troublesome than total electricity demand. Data centers’ power consumption profile is abnormal. Traditional factories operate in shifts—running by day and shutting down at night; residential electricity use follows predictable morning and evening peaks, as regular as clockwork. Data centers, by contrast, run near full capacity around the clock—hardly catching their breath. Even more challenging are large-scale training jobs. When tens of thousands of GPUs start and stop simultaneously, the instantaneous power fluctuation reaches gigawatt scale. Such sudden, massive load shedding is far harder for the grid to manage than sustained high load. In fact, there’s a solution the power industry has used for decades: demand response. Large users voluntarily reduce their load when the grid is stressed—and receive compensation afterward. Aluminum smelters and steel mills have done exactly this. There are only dozens of truly critical hours per year; if major users would yield during those few hours, the grid could absorb significantly more capacity—many projects wouldn’t need to wait seven years for infrastructure upgrades. The problem is this mechanism doesn’t work well for AI. If a training job is interrupted, that segment may need to restart from scratch; inference services simply cannot go offline—users are waiting on the other side of the screen. The computing industry touts itself as highly flexible in its marketing—but when it comes to electricity use, it’s the most inflexible class of large consumers. The thing hailed as the most adaptable is precisely the one least willing to compromise. So where, then, is electricity coming from in the near term? Tech companies have signed numerous nuclear power agreements—sounding promising on paper. Yet existing nuclear plants have fixed output capacity; signing more contracts mainly locks up already-available electricity rather than creating new supply. Small modular reactors (SMRs) are widely seen as promising, but commercial deployment timelines generally point beyond 2030. Wind and solar power are cheap—but intermittent. To provide round-the-clock stable power to data centers, massive energy storage is required—and storage systems themselves face the same seven-year lead times. Counting them off one by one, the only scalable option likely to fill the gap over the next three to five years is natural gas. Thus emerges an awkward situation: several companies that have been loudest in pledging carbon neutrality are now rapidly deploying natural-gas-fired generators to power AI. Meta, for instance, built a $1.6 billion natural-gas power facility at its Ohio site, complete with a dedicated microgrid—a configuration known in industry jargon as “island mode.” The commitment is written into ESG reports; the power plant is built on the ground in Ohio. There’s another cost to self-built power plants—one rarely accounted for. On the surface, tech companies generating their own electricity appear to withdraw from competition for public grid power. In reality, they’re merely competing for something else under a different name: natural gas pipeline capacity is limited, gas turbine delivery timelines are lengthening, and emissions remain concentrated in local air. More critically, there’s the grid’s fixed cost. Transmission lines, substations, dispatch systems—these investments must be allocated across users. When large users build their own power plants and exit this shared pool, the remaining fixed costs fall to those who stay behind.

Why are ordinary households beginning to foot the bill for AI giants’ electricity bills?

When a large data center prepares to locate in a region, its first step is to apply to the local utility for power capacity. To the grid, it isn’t just another new customer—it’s more like a medium-sized city suddenly relocating nearby: highly concentrated load, year-round operation, and extreme sensitivity to outages. To accommodate it, utilities may need to build new substations, transmission lines, natural-gas peaking units—or even upgrade networks dozens of kilometers away. These projects routinely take years—and require upfront investment. The crux lies in “who pays first.” If the data center bears all costs, the project may become too expensive and relocate to another state or country. If the utility builds first and recoups expenses gradually through the broader rate structure, part of the cost may ultimately be passed on to residential and small-business customers. And if demand forecasts prove overly optimistic—or if projects are delayed or canceled—the already-built infrastructure leaves behind “stranded costs.” So residents aren’t paying for the reading on the AI company’s meter. Instead, they may be covering system-level costs: reserving capacity for an ultra-high-load user, reinforcing the grid, and preventing blackouts. The statement “AI consumes a lot of electricity” is both true—and insufficiently precise. In October 2026, the International Energy Agency (IEA) released its report *Powering Artificial Intelligence*, covering roughly 90% of global data center capacity. It identifies affordability, sustainability, and security as the three core challenges of AI’s energy use. Earlier, the IEA disclosed that global data center electricity consumption will rise about 17% in 2025—with AI-dedicated data centers growing roughly 50%. Major tech firms’ capital expenditures in 2025 exceeded $400 billion—and are projected to surge further in 2026. These figures confirm mounting pressure—but blaming every electricity price hike on a single chat session is misleading. Even among requests labeled “AI,” energy consumption varies enormously. A simple text Q&A differs vastly—in energy terms—from generating long videos, performing extended-chain reasoning, or running always-on intelligent agents. The IEA estimates that replacing traditional search with a brief AI text query adds only a limited share to global data center electricity use; what steepens the curve are heavier models, longer inference durations, and broader deployment scales. The issue isn’t that you asked AI one extra question today—it’s that the entire industry is transforming “computation” into a new kind of infrastructure demand. Why isn’t “100% renewable” enough? Many tech companies purchase renewable energy equivalent to their annual electricity consumption—and even announce achievement of “100% clean energy matching.” This is progress—but “annual totals matching” doesn’t mean “matching hour-by-hour, location-by-location.” A data center might run at full load overnight, while its contracted solar farm generates only during daytime; the data center may be sited in a grid-congested area, while its green-power source sits in another state. On annual balance sheets, the two sides cancel out—but the real-world grid still relies on fossil-fueled generation or storage to cover gaps during certain hours. Moreover, there’s water. A World Bank article published early in 2026 notes that data centers account for roughly 1.5% of global electricity use—and cites studies estimating AI-related water demand could reach 4.2–6.6 billion cubic meters by 2027. That range carries significant uncertainty, and corporate disclosure standards remain inconsistent—but it reminds us: data center siting doesn’t just reshape electricity demand curves; it may also pit data centers against local residents and agriculture for cooling water and power-generation water.

Energy Transfer Approves $1.6 Billion Gas Pipeline as AI Use Soars

12 hours ago

Energy Transfer LP approved a $1.6 billion natural gas pipeline to meet soaring demand from AI data centers and LNG exports. The 215-mile South Mississippi Pipeline is expected to start operations in 2030 with a capacity of 1.2 billion cubic feet per day.

bloomberg

2026-06-20

Energy Transfer to Expand Ethane, LPG Export Capacity at Nederland

2026-11-03

Energy Transfer Approves $1.6 Billion Gas Pipeline as AI Use Soars

Energy Transfer Approves $1.6 Billion Gas Pipeline as AI Use Soars

FAQ