For most of the 2000s and 2010s, American electricity demand was essentially flat. Population grew, the economy expanded, and appliances multiplied — yet total kilowatt-hours barely budged. Efficiency gains from LED lighting, better motors, and tighter building codes absorbed most of the growth. Grid planners could forecast a decade out and count on demand arriving roughly where they expected it to.
That era is over. Across much of the country, load forecasts have been revised upward sharply, and the cause is largely concentrated: data centers, particularly facilities built to support artificial intelligence workloads. Understanding why these facilities matter so much — and how they differ from ordinary commercial load — is now foundational to understanding the power sector itself.
The end of flat demand
To appreciate the shift, it helps to put numbers on what “flat” meant. Between 2008 and 2023, total U.S. electricity sales grew at well under 1 percent per year in real terms, and in some years fell outright. This was historically unusual. For most of the 20th century, electricity demand grew at several percent annually, doubling roughly every decade. The grid was built for growth, then spent two decades managing stability without it.
The turnaround began around 2023 and 2024, when regional grid operators and utilities started revising load forecasts upward. The revisions were not small. In the PJM region — the wholesale market covering the Mid-Atlantic and parts of the Midwest — summer peak forecasts rose by several gigawatts in a single planning cycle. Similar revisions appeared in Texas (ERCOT), the Pacific Northwest, and the Southeast. The common thread in every revision: large new commercial loads, dominated by data centers.
These revisions matter because the grid plans in decades. A power plant approved today may not generate electricity for five to ten years. Transmission lines take even longer. When forecasted demand jumps by gigawatts on short notice, the entire planning pipeline — generation procurement, transmission planning, resource adequacy modeling — has to recalibrate at once.
Why data centers are different from other load
A new office building, warehouse, or shopping mall adds load to the grid in familiar ways. The utility knows roughly how much power it will draw, when it will draw it, and how that demand will behave over the building’s lifetime. Data centers break these assumptions in several ways.
First is sheer scale. A large modern data center can draw 100 to 300 megawatts of continuous power — comparable to a small city. Hyperscale campuses under development reach 500 megawatts or more, and some planned sites target a full gigawatt. For illustration, a single 200-megawatt data center consumes roughly as much electricity as 150,000 to 200,000 homes, but it arrives at the grid in one place, on one contract, with one customer’s timeline.
Second is concentration. Historically, load growth was diffuse: millions of small additions scattered across a territory. Data center load arrives in lumps. A utility may receive interconnection requests for several gigawatts of new data center load in a single year — more than the territory’s organic growth over a decade. This lumpiness strains every part of the planning process, from substation capacity to generation adequacy.
Third is the speed of the request. A residential subdivision grows over years. A data center developer may ask for 200 megawatts with an energization date two to three years out. The mismatch between the customer’s timeline and the utility’s infrastructure timeline is one of the central tensions in modern grid planning.
Fourth is the load profile. Data centers draw power nearly constantly — often at 70 to 90 percent of their contracted capacity, around the clock. This is very different from residential or commercial load, which peaks in the afternoon and falls overnight. Constant, flat load is actually efficient for the grid in some ways (high load factor, predictable dispatch), but it means every megawatt of data center load shows up in the peak that sets capacity requirements.
The anatomy of the demand
Data center power goes to three main places. The largest share — typically 40 to 60 percent — feeds the servers themselves: the processors, memory, and networking that do the computing. The second-largest share goes to cooling: removing the heat those servers generate. In older facilities, cooling could consume nearly as much power as the IT equipment itself; modern designs have cut that ratio substantially, as discussed in our piece on PUE. The remainder goes to power conversion losses, lighting, and support systems.
The rise of AI training has changed the composition of this load. A traditional enterprise data center might fill a building with general-purpose servers drawing a few kilowatts per rack. An AI training cluster uses specialized processors that draw far more — racks of 50 to 100 kilowatts are becoming common, and some designs push higher. This raises both the total load and the density of the load, which creates new challenges for cooling and for the local distribution system.
It also changes the economics. AI hardware is expensive, and the companies deploying it want it running at maximum utilization as quickly as possible. This creates intense time pressure on interconnection — the process of connecting a new large load to the grid — which was designed for a slower era.
How big is the growth, really?
Quantifying the data center share of future demand is genuinely difficult, and honest estimates carry wide ranges. What analysts agree on is the direction and the order of magnitude: data centers are expected to be the largest single source of new electricity demand in the United States through the end of this decade, growing from a few percent of national electricity consumption today toward the high single digits or low double digits by 2030.
For illustration, if national electricity consumption is roughly 4,000 terawatt-hours per year, a shift of even three to four percentage points of that total toward data centers represents more than 100 terawatt-hours of new annual demand — equivalent to adding tens of gigawatts of new generation capacity running year-round. And that growth is not evenly distributed: it concentrates in a handful of regions with favorable land, power, fiber, and tax conditions, where the local impact is far larger than the national average suggests.
The uncertainty cuts both ways. Some announced data centers will never be built; interconnection queues always contain speculative projects. Forecasting methodology is adapting — utilities and grid operators are developing new techniques to distinguish real projects from placeholders, including requiring larger deposits, tighter milestone schedules, and minimum take-or-pay commitments.
The regional picture
National averages obscure where the growth actually lands. Data center development clusters in a small number of regions — Northern Virginia, central Texas, the Pacific Northwest, parts of the Southeast and Midwest — where land, fiber, water, tax treatment, and power availability align. In these concentrated markets, data centers can account for a double-digit share of local electricity consumption, and the grid impact is transformative rather than incremental.
This concentration creates a two-speed grid. In cluster regions, utilities face the full force of the load-growth challenge: multi-gigawatt interconnection queues, transmission buildouts, and rate cases dominated by cost-allocation disputes. In regions with little data center development, the old world of flat demand and gradual efficiency-driven decline persists. National policy debates about load growth often talk past each other because participants are describing different regional realities.
The clustering also creates feedback effects. Once a region proves it can interconnect large loads quickly, more developers choose it, which lengthens its queues and strains its grid — success breeds the congestion that undermines it. Meanwhile, regions with ample grid headroom but no data center ecosystem struggle to attract the first movers. State economic development agencies have noticed: power availability and interconnection speed are now central pitches in the competition for data center investment, alongside traditional tax incentives.
Electrification’s supporting role
Data centers dominate the load-growth story, but they are not the only new demand. Electric vehicles, building electrification (heat pumps replacing gas furnaces), and industrial reshoring — new manufacturing plants, particularly in semiconductors and batteries — all add load. In most forecasts, these sources are individually smaller than data centers through 2030, but collectively significant.
The interaction matters for planning. EV charging and heat pumps have very different load profiles from data centers — peakier, more weather-sensitive, more distributed. A grid planned only for flat data center load would misserve the peaks driven by electrification. Conversely, the steady baseload character of data centers actually helps absorb variable renewable generation: a constant load is easier to match with wind and solar portfolios than a spiky one. Integrated planning has to serve both profiles simultaneously, which is one reason the forecasting problem has gotten so much harder.
Why this matters beyond the grid
Load growth reshapes the power sector’s economics. For two decades, utilities in flat-demand regions could focus on replacing aging infrastructure and integrating renewables. Growth changes the investment calculus: it creates a reason to build new generation, which creates jobs, tax revenue, and political interest — but also raises questions about who pays.
The central economic question is cost allocation. Serving a gigawatt of new data center load may require billions of dollars in new generation and transmission. If those costs are socialized across all ratepayers, residential bills rise even though households caused none of the growth. If the data center bears the full cost, the economics of the facility may not pencil out. Utilities, regulators, and large customers are now negotiating this boundary in rate cases across the country — through new tariffs, minimum bills, and extended contract terms that keep the large load from shifting costs to smaller customers.
There is also a reliability dimension. Grid operators maintain a margin of generation above expected peak demand — the reserve margin — to cover plant outages and extreme weather. When peaks rise faster than new generation arrives, reserve margins shrink, and the risk of shortfalls during extreme events rises. Several regions are watching their reserve margins tighten as data center load arrives faster than planned capacity.
What to watch
The load-growth story is still being written, and several variables will determine its shape. The first is how much AI demand materializes versus how much is speculative buildout — the difference between sustained growth and a boom-bust cycle. The second is how quickly utilities and regulators adapt their planning and interconnection processes to large, fast-moving loads. The third is how the costs of new infrastructure get divided between large customers and everyone else.
What is not in doubt is that the era of flat demand has ended. The grid is growing again, and data centers are the reason. Every subsequent question — about interconnection, efficiency, power procurement, and planning — starts from that fact.