U.S. AI Data Center Timeline 2010–2026: When the AI Boom Met America’s Electric Grid

America’s data centers used about 192 terawatt-hours of electricity in 2024 — 4.7% of total U.S. electricity consumption, according to Lawrence Berkeley National Laboratory (LBNL), a U.S. Department of Energy national lab. LBNL’s 2025 Update to that report now puts 2030 data-center electricity use at a Reference Case of 649 TWh, or 11.8% of forecast U.S. electricity that year, with a modeled range of 9.5% to 15.3% (521–843 TWh) depending on how fast GPU shipments, AI-server lifetimes and inference workloads actually play out. Generative AI is the single biggest reason that curve bent upward after 2022. But every one of those figures is for the broader U.S. data-center sector — cloud storage, streaming, enterprise IT and AI workloads together — not for AI alone. What’s real: demand growth is fast and regionally concentrated. What’s uncertain: exactly how much of it gets built, how fast, and who pays for the wires and power plants underneath it.

Lawrence Berkeley National Laboratory’s original December 2024 U.S. Data Center Energy Usage Report estimated 2023 consumption at ~176 TWh (4.4% of U.S. electricity) and projected a 2028 range of 325–580 TWh (6.7%–12%). LBNL’s 2025 Update, published in 2026, revised the near-term numbers — 2024 data-center electricity use came in at 192 TWh (4.7%), and its Reference Case now puts 2028 at 464 TWh. Extending that same Reference Case to 2030 gives 649 TWh, or 11.8% of forecast U.S. electricity that year — a share calculated against the North American Electric Reliability Corporation’s (NERC) 2025 Long-Term Reliability Assessment. Combining LBNL’s sensitivity scenarios (lighter or heavier GPU shipments, shorter AI-chip lifetimes, higher inference idle power) produces a Compounded Uncertainty range of 521–843 TWh, or 9.5%–15.3%. The spread exists because nobody can yet say how many announced gigawatt-scale campuses actually get built, financed and energized on schedule.

Methodology changed between the 2014 and 2023/2028 figures — treat this as directional growth, not one continuous, perfectly comparable trend line.

LBNL’s 2025 Update Reference Case: 649 TWh, 11.8% of forecast 2030 U.S. electricity, inside a 521–843 TWh scenario range. Source: LBNL.

AI development can move fast because its inputs — chips, code and money — are things a company can buy or build in-house on a compressed timeline. Electricity infrastructure can’t move that fast, because it depends on things no single company controls: utility interconnection studies, state permitting processes, transformer and turbine manufacturing queues, and, often, entirely new transmission lines that cross multiple jurisdictions. Software can scale in months. Power systems can take years. A utility cannot serve a new 1-gigawatt customer with a software update.

A rack of AI accelerators doesn’t care whether its electricity came from a nuclear plant, a wind farm or a gas turbine — the grid does the work of turning fuel or sunlight into electrons, and the data center just consumes them. What makes AI workloads unusually power-hungry is scale and density: training a large model means running thousands of GPUs continuously for weeks, and serving millions of users means running inference around the clock. Add networking, high-bandwidth memory, and the cooling systems needed to keep dense server racks from overheating, and a single AI campus can draw as much power as a mid-sized industrial facility — or more.

As AI adoption grows, inference’s share of total electricity use tends to become more important simply because it never stops, while training runs are finite events. No credible public research currently gives a precise, universal split between the two at the U.S. national level — treat any specific percentage you see with caution.

Megawatts (MW) and gigawatts (GW) measure power — the rate at which electricity is used at any given moment, like the speedometer in a car. Megawatt-hours (MWh) and terawatt-hours (TWh) measure energy — the total amount used over time, like the odometer. One gigawatt sustained for one hour equals one gigawatt-hour. The same 1 GW sustained continuously for a full year equals about 8.76 TWh.

Electricity doesn’t stop at the meter — it’s stepped down, distributed, computed with, and cooled.

Every new generation project needs site development, financing, permitting, an interconnection study, equipment procurement, fuel supply arrangements where applicable, environmental review and construction — then transmission capacity to actually move the power to where it’s needed. But timelines vary enormously by technology: utility-scale solar and battery storage can sometimes be sited and built in one to three years, while new nuclear plants, large gas turbines (currently facing multi-year order backlogs) and major transmission lines routinely take five to ten years or more. AI capacity, grid capacity and generation capacity do not scale at the same speed — and pretending they do is why “just build more power plants” undersells how the constraint actually works.

Connecting a new large customer to the grid isn’t just a matter of running a wire. It requires an interconnection study to confirm the local substation, transformers and transmission lines can handle the added load without destabilizing the grid for existing customers, plus often new transmission or generation to serve it. For a 1-gigawatt data center, that can mean new substations, transformer orders (some with multi-year lead times), and transmission upgrades spanning dozens or hundreds of miles.

By 2026, reported large-load interconnection requests nationally — overwhelmingly from data centers — had climbed past 700 gigawatts, more than ten times LBNL’s estimate of actual current U.S. data-center electricity consumption. Texas alone saw its ERCOT queue grow from roughly 48 GW in 2023 to about 474 GW by 2026. A meaningful share of that requested load is what researchers and grid planners now call “ghost demand”: duplicate applications from the same developer filed with multiple utilities to keep options open, speculative site reservations without secured financing or chip orders, and projects that never break ground. Filing an interconnection request has historically cost a developer almost nothing — some paperwork, a modest deposit, maybe a feasibility fee — with no requirement to show financing, permits or a signed power contract first, so there was until recently little cost to over-requesting.

This matters because exaggerated queues distort planning: utilities may study or even build infrastructure for demand that never materializes, other customers can face longer interconnection wait times behind inflated queues, and long-term capital gets allocated based on unreliable signals. It’s why Texas and other markets have started requiring large-load applicants to show financial and technical commitment — a signed equipment order, secured financing, site control — before their request is treated as real in grid planning.

ERCOT’s data-center interconnection queue grew from about 48 GW in 2023 to roughly 474 GW by 2026 — nearly a tenfold jump, and far more than Texas could ever realistically connect, let alone power. In response, Texas Governor Greg Abbott signed Senate Bill 6 on June 20, 2025, directing the Public Utility Commission of Texas (PUCT) and ERCOT to overhaul how large loads connect to the grid. In March 2026, the PUCT published a draft rule (16 TAC §25.194) requiring any new load of 75 MW or more to prove site control (a deed, purchase option, or a lease running at least five years past its contracted peak-demand date) and sign an intermediate agreement with the transmission provider before ERCOT even begins an interconnection study — alongside a proposed $50,000-per-MW interconnection fee meant to filter out speculative filings. A final large-load rule is expected by December 31, 2026. Separately, the PUCT has already affirmed ERCOT’s authority to curtail co-located data-center load during grid emergencies, in its first net-metering case decided under SB6. None of this amounts to a ban on data centers — it’s a readiness and cost-allocation filter aimed squarely at Texas’s own ghost-demand problem.

PJM Interconnection is a regional transmission organization — not a utility — that coordinates wholesale electricity markets and grid reliability across 13 states and Washington, D.C., including Virginia, Ohio and Pennsylvania. PJM’s territory contains the country’s densest concentration of data centers, concentrated heavily in Northern Virginia. Rising large-load forecasts, alongside power-plant retirements and market-rule changes, have pushed PJM’s capacity-market auction clearing prices sharply higher: from $28.92 per megawatt-day for the 2024/2025 delivery year to $269.92/MW-day for 2025/2026, and $329.17/MW-day for 2026/2027 (a figure PJM’s own price cap held down from going higher still). PJM’s Independent Market Monitor has attributed 63% of the 2025/2026 increase — about $9.3 billion — to data-center load growth specifically, with the remainder tied to generator retirements and other market dynamics. That single-auction attribution is the most-cited figure; multi-year cumulative projections (one advocacy estimate puts $100–163 billion through 2033 absent reform) are forward-looking projections, not settled facts, and depend on auctions that haven’t happened yet.

PJM’s capacity clearing price roughly tripled year-over-year, then rose again under a price cap. Source: PJM auction results.

Northern Virginia became “Data Center Alley” for reasons that predate AI: dense fiber-optic connectivity, major internet exchange points, an established cloud-computing ecosystem, available land and historically favorable tax policy. Those same advantages now collide with real constraints — transmission capacity, community opposition to new substations and transmission lines, and Dominion Energy’s own long-range load forecasts. These constraints are specific to Virginia’s grid geography; they don’t automatically apply to every U.S. data-center market.

Maybe — but the answer depends heavily on where you live and how costs are allocated, not on a single national rule. New transmission lines, generation, substations and capacity-market costs can all flow into utility rates if regulators let data-center-driven costs be spread across all customers. Some states and utilities are moving toward large-load tariffs and minimum-demand contracts specifically designed to make data centers, not households, shoulder the infrastructure risk they create. Distinguish four different things reporters often blur together: the wholesale price (what generators are paid), the capacity price (paid for guaranteed future availability), the utility retail rate (what you actually pay per kWh) and separate transmission/distribution charges on your bill. A rise in one doesn’t automatically mean a rise in the others. Concrete, market-specific examples exist: Washington, D.C.-area utility Pepco raised residential bills by roughly $21 a month starting in June 2025, and one advocacy group’s modeling projects PJM-territory households could see bills rise by roughly $70 a month by 2028 versus pre-surge levels if capacity costs keep climbing and aren’t reallocated. Those are PJM-specific, capacity-market-linked figures — not a claim that every U.S. household’s bill is rising because of AI.

Nuclear plants offer qualities that make them attractive to data-center operators seeking firm power: large, steady output; a high capacity factor (they run near-full power most of the time); low operational carbon emissions; and, for existing plants, a connection to the grid that already exists.

Natural gas remains the largest single source of U.S. electricity generation, and gas turbines can be dispatched quickly and paired relatively fast with new data-center load compared to nuclear or major transmission builds — though current turbine order backlogs still stretch several years. Gas plants emit CO2 and are exposed to future price volatility and long asset lifespans that could clash with state and corporate climate targets, plus pipeline capacity constraints in some regions. It is neither an obviously good nor obviously bad choice — it’s a trade-off between near-term reliability and long-term emissions commitments.

Solar can supply large volumes of low-cost electricity during sunny hours; wind adds output that peaks at different times and in different locations, improving diversity across a portfolio. Batteries can shift some of that electricity across hours, but most deployed storage today covers short durations (hours, not days) — a different problem from a data center’s need for essentially uninterrupted power. The honest answer is neither “solar alone can’t power data centers” nor “solar plus batteries can always deliver 24/7 supply” — it depends on the specific site, the portfolio mix and how much firm backup (gas, nuclear, or grid import) sits alongside the renewables.

Cooling systems in many data centers use water, and electricity generation itself (especially thermal plants) can also be water-intensive depending on the cooling technology used. Water intensity varies enormously by data-center cooling design (air-cooled vs. evaporative vs. liquid cooling) and by region — there is no single “liters per AI prompt” figure that applies everywhere. Local community concerns about water use are real and documented in specific markets, but they don’t generalize into a single national number.

Better accelerators, quantization, smaller and more specialized models, Mixture-of-Experts architectures, model distillation, software optimization and higher data-center utilization are all reducing the electricity cost of a given AI task over time. But cheaper, more efficient AI tends to get used more — a pattern economists call the rebound effect. Efficiency gains lowering the cost per task don’t automatically translate into lower total electricity demand if usage grows faster than efficiency improves.

Not every AI company is a data-center owner — some (OpenAI, xAI) primarily lease or co-develop capacity rather than build and operate facilities outright.

OpenAI, Oracle and SoftBank’s Stargate program is among the largest announced AI-infrastructure efforts, with sites including one near Abilene, Texas. Announced financing and site plans are real, but announced capacity and actually energized, operating capacity are not the same thing — large infrastructure programs routinely take longer and cost more than their initial announcements suggest, and not every announced phase is guaranteed to be built as originally described.

One bottleneck — a single missing transformer, a permit stuck in review — can delay an entire facility’s energization by months or years, regardless of how quickly the chips and buildings themselves are ready.

This tracker will be updated as new verified developments are confirmed; it is not exhaustive.