How Many Data Centers Are in the USA, and Why the Numbers Don't Agree

By Muninder Adavelli. Scraping, SEO, and proxies for over a decade.

The most common answer to "how many data centers are there in the US" is off by a factor of somewhere between two and ninety, depending on which provider you ask. That's not a rounding problem. It's a definitional one, and it collides with a public that, in the same twelve-month window, went from mostly not thinking about data centers to mostly not wanting one built nearby. This is a look at the 2026 US data-center map: how large it actually is, where the buildout is landing, why opinion turned negative in seven months, and what the pipeline says about the next two years.

1. How many data centers are in the US, and why the count depends on who's counting

1.1 The Cleanview count: 3,543 total, 1,302 operating, 2,196 planned

The United States hosts 3,543 data centers across 419 counties, according to Cleanview's 2026 dataset: a figure that includes 1,302 currently operating facilities and 2,196 projects in the planning or land-banked phase. That total represents 456,532 MW of combined capacity: 61,751 MW already online, and 385,893 MW in the pipeline. The planned capacity is 6.2 times the current operating stock.

The real story, though, isn't the total. It's the variation in how different providers define "data center." Aterio tracks 8,049 US facilities: a scope that includes colocation hubs, edge nodes, and hyperscaler campuses under one roof. ElectricChoice reports over 2,100 tracked facilities, restricted to those in active development or operational status. Cleanview sits between these two, and its count is the reference used through the rest of this piece; no single number is stable. The US data-center map isn't one map; it's three, drawn by three different providers with three different counting rules.

1.2 The state ranking depends on the source

Pew Research Center's Feb 2026 analysis of Data Center Map ranks Virginia first with 685 total data centers, followed by Texas (466), California (331), Illinois (262), and Georgia (235). But cross-referenced with Cleanview's 2026 data, the rankings shift. Virginia climbs to 749, Texas to 511: small gaps, both about 1.10×. California drops to 86, a 3.85× gap that's by far the widest among the top-15 states.

The California divergence isn't error. It's methodology. Cleanview counts only facilities with confirmed capacity or construction; Pew's underlying Data Center Map source includes land-banked sites and smaller colocation clusters. California has an unusually large enterprise-cloud footprint relative to its commercial data-center count, and the definitional cut hits harder there than it does in states where the market is dominated by standalone hyperscaler and colocation campuses. The takeaway for a reader is not "one source is right" but "state-level counts should be treated as a range, not a point."

1.3 The pipeline dwarfs the fleet

The US data-center pipeline is 6.2× the size of what's currently operating. Cleanview's 2026 data shows 385,893 MW of planned capacity across 2,196 projects, compared to 61,751 MW already operating. If even half of that pipeline materialises by 2030, the US would double its data-center capacity in under four years.

Some 846 data centers are currently under construction in the US, and over 620 additional projects are in active planning across 47 states. Texas leads with 91 projects in the active pipeline, followed by Virginia (70), Ohio (48), Iowa (28), and California (27). The gap between operating and planned capacity isn't a lag; it's a structural feature of the current buildout, and it's the number the grid, utility commissions, and county planning boards are already stress-testing against.

1.4 New buildout is landing in rural counties

Most new US data centers are being built in rural counties: a sharp reversal from the historical pattern of urban clustering. Pew Research Center's April 2026 analysis found that 87% of currently operating data centers sit in urban areas, but 67% of planned data centers are in rural counties. Some 39% of planned projects are in counties that don't currently have a data center at all. Three-quarters (75%) of all planned data centers will be built in the South and Midwest combined, with the South alone accounting for 48% of the total.

That geographic shift matters for the next section. The counties absorbing the buildout are, in general, not the same places driving national opinion on it. The people who feel the noise, the truck traffic during construction, the pressure on the local grid, and the visible cooling towers on the edge of town are increasingly not the people whose views show up in national polls with the most weight.

2. The public turned negative in seven months

2.1 Awareness jumped from 75% to 88% in the same year

When Pew Research Center surveyed US adults in January 2026, three-quarters had heard of data centers: a baseline that already signaled infrastructure visibility. By August 2026, that share had climbed to nearly nine in ten. The share who said they hadn't heard of data centers dropped from 25% to 12%. That isn't incremental awareness; it's a threshold crossing. Data centers are no longer niche tech infrastructure to the American public. They're part of the national conversation, and with that comes scrutiny.

Awareness alone would predict either support or ambivalence. What happened instead was sharp negative movement across every dimension Pew measured.

2.2 Every category moved more negative in only seven months

The seven-month span between Pew's January and August 2026 surveys wasn't just a window of time; it was a turning point. The share of Americans saying data centers are "mostly bad" rose in every category measured. On environment: 39% → 54% (+15pt). On home energy costs: 38% → 50% (+12pt). On people's quality of life nearby: 30% → 49% (+19pt), the largest single-category jump. On local jobs: 15% → 24% (+9pt). On local tax revenue: 12% → 21% (+9pt).

Local jobs and local tax revenue are the two categories where "mostly good" still slightly leads "mostly bad": 22% good vs 24% bad on jobs, 22% good vs 21% bad on tax revenue. But those gaps closed sharply over the same seven months. Everywhere else (environment, home energy costs, quality of life), "mostly bad" is now the plurality view. The pattern is consistent across dimensions the industry has traditionally treated as its strongest arguments.

2.3 Six in ten Americans wouldn't want a data center in their area

Abstract concern has become site-level refusal. Pew's Aug 2026 survey asked whether Americans would be comfortable with a new data center operating in their area: 60% said "not too" or "not at all" comfortable. Only 7% said "extremely or very comfortable"; another 19% said "somewhat comfortable." NIMBY is now the default posture.

That number matters most in context with the rural-siting pattern from earlier. Data centers are being built in counties where residents have limited state-level advocacy infrastructure and no dedicated federal lobbying presence, and the opposition sentiment is running well ahead of the industry's ability to convert local jobs numbers into local political support.

2.4 Democrats are now more worried than Republicans about AI's job impact

Historically Republicans were more skeptical of AI's job impact. That polarity flipped. Pew's Sept 2026 short-read documents a 17-percentage-point rise since 2023 in the share of Democrats who think AI will lead to fewer jobs, and a 10-point drop for Republicans (now at 49%). On the broader "more concerned than excited about AI in daily life" measure, 56% of Democrats say concerned, up from 46% in 2023: a 25-point cumulative shift since 2021. Liberal Democrats' concern rose from 45% to 63%; conservative Republicans' fell 14 points over the same period.

The alignment matters here because it maps onto geography. Democrat-leaning urban areas concentrate the federal regulatory influence that shapes national rules on emissions, transmission, and permitting. Data centers, meanwhile, are being sited in more Republican-leaning rural counties. For most of the past decade, that separation kept siting decisions local. What Pew is measuring is a partisan alignment where the political will to intervene from the outside is now rising in the places that could actually intervene.

3. The grid is the constraint

3.1 Data centers used 4.4% of US electricity in 2023, projected at 426 TWh by 2030

Data centers consumed 4.4% of total US electricity in 2023: roughly 176 TWh. The International Energy Agency estimates US data-center electricity use reached 183 TWh in 2024. A DOE-commissioned Berkeley Lab report projects consumption will surge to 426 TWh by 2030: a 133% increase from 2023 levels. The IEA's international framing is that US data-center demand is already roughly equivalent to the entire annual electricity demand of Pakistan.

The 4% share is misleading in isolation. The US grid was designed for a mix of industrial loads none of which absorbed more than a few percent of national output, and data centers are the fastest-growing industrial electricity consumer since air conditioning entered wide residential deployment. The 426 TWh projection isn't a scenario; it's built on current buildout pipelines with named tenants and signed contracts.

3.2 The strain is a state-level story, not a national one

The national 4.4% figure hides how concentrated the load actually is. Virginia's data centers consume more than 1 in 4 kilowatt-hours generated in the state: an estimated 32 TWh out of 128 TWh in 2023, according to an EPRINC analysis. The Electric Power Research Institute puts North Dakota at 15%. Dominion Energy reported that data centers accounted for 21% of its sales as of late 2022. State-share is what utility regulators, ratepayers, and voters actually see, which is why the political fight over electricity prices is concentrated in the states with the biggest existing clusters, not the ones with the fastest growth rate.

3.3 Behind-the-meter is the industry's answer to a grid that can't keep up

Behind-the-meter (BTM) generation lets a data center connect directly to a power producer, skipping the transmission grid. Grid Strategies puts online BTM capacity at roughly 2 GW today, with the projected pipeline reaching approximately 90 GW. That's a 45× gap between what's running and what's announced. Of the 90 GW total, roughly 60% is still at announcement or early-planning stage, another 36% has been permitted, and just 1.2% is under active construction. In a scenario where many projects experience delays, cumulative BTM capacity could be as low as 5 GW by the end of 2027.

The pipeline size, more than the online capacity, is the point. The industry is telling regulators, utilities, and communities that it doesn't believe the standard grid-connection process can move at the pace the buildout requires. Whether the 90 GW target holds or slips to 5 GW is now largely a regulatory and permitting question, not a technical one.

3.4 Household bills are the mechanism the public feels

The abstract "grid strain" story has a concrete household-level readout. Carnegie Mellon University projects that data centers and cryptocurrency mining could increase the average US electricity bill by 8% by 2030, and could push bills up by more than 25% in the highest-demand markets of central and northern Virginia. The typical US household paid $142 per month for electricity in 2024, up from $114 per month in 2014: a 25% increase over ten years, according to the US Energy Information Administration. In 2025 alone, over $60 billion in rate increases were approved nationwide.

The health-cost side of the ledger is emerging too. One analysis in Virginia estimated that even limited use of backup diesel generators is already associated with roughly $300 million in annual public health costs and 14,000 asthma-related health impacts across multiple states. Pew's tracking data on the "mostly bad for home energy costs" question ties directly to this: the share of Americans saying data centers are bad for home energy costs jumped 12 points between January and August 2026, from 38% to 50%. That's not an opinion drifting untethered from data. That's opinion catching up to the utility bill.

4. Texas and Virginia are two very different states of buildout

4.1 Virginia already runs the country's densest cluster: 17,325 MW operating

Virginia has 352 operating data centers with 17,325 MW of combined capacity: the highest operating capacity of any US state. Seventy-one unique developers are active in the state. Northern Virginia was already the largest data-center market in the world before the current AI wave; the current planned buildout of 34,758 MW across 392 projects is roughly double the operating stock. That doubling isn't a runway; for a market this mature, it's consolidation. Virginia's story from here is optimisation and defence of grid capacity, not first-time buildout.

4.2 Texas is where the pipeline is: 97,905 MW planned, 13.7× its operating capacity

Texas has only 139 operating data centers (7,169 MW) but 371 planned projects totalling 97,905 MW: a 13.7× expansion over the operating stock. Ninety-five unique developers are active in the state, more than any other. Individual named projects show the pattern: the Fermi Project Matador Expansion (Gas) at 5,778 MW and Fermi Project Matador East Expansion (Gas) at 5,000 MW, both in Carson, TX; the Fermi Project Matador Expansion (Nuclear AP1000) at 4,400 MW in the same location; Nexus Hubbard Expansion at 5,630 MW; Data City Expansion at 4,700 MW in Webb County. Texas is where the AI-era buildout is being paired with dedicated gas-fired and nuclear generation at gigawatt scale: a very different footprint profile than Virginia's grid-connected hyperscaler cluster.

4.3 The count discrepancy is largest in California, not Texas

Second methodology callout in this article, because it matters. Cleanview says California has 86 data centers total; Pew Research Center (using Data Center Map) says 331. That's a 3.85× gap in a state with a sophisticated commercial-real-estate reporting infrastructure. Virginia's and Texas's cross-source gaps are 1.09× and 1.10×. The California gap is a definitional one: different providers count colocation, hyperscaler, and edge deployments differently, and California's dispersed enterprise-cloud presence means the definitional cut hits especially hard. When comparing states, the range matters more than the point estimate.

4.4 Washington shows what a two-firm mature state looks like

Washington's data-center market is what saturation looks like. Forty-one operating data centers with 1,069 MW of capacity, only eleven planned adding 1,434 MW. Fifteen unique developers are listed, but two (Microsoft and Amazon) account for the overwhelming share of both existing and planned load. Microsoft operates multiple 57 MW facilities across Grant, Douglas, and Chelan counties. Amazon's four planned Wallula Gap phases, at 300 MW each and 1,200 MW combined, dominate the pipeline. That pipeline is only 1.3× the operating stock: well under half the multiplier of Virginia and well under a tenth of Texas's. Two-firm markets are stable, and Washington's stability is the reason its pipeline is small. There's no runway left for a Texas-scale expansion in the state.

5. Environmental pressure and the regulatory fight

5.1 Water, air, and noise: the physical footprint of a cluster

Two-thirds of new hyperscale campuses built since 2022 are in high water-stress counties, and roughly one in five servers now sits in a water-stressed watershed. A single large modern AI data center can consume up to 5 million gallons of water per day: comparable to what a small town uses, or roughly the indoor water use of 360,000 households. Hyperscale data centers alone are projected to consume 16-33 billion gallons of water annually by 2028. According to IEA analysis, the larger AI-focused data centers currently under construction are expected to use around 20 times as much electricity as a typical AI-focused hyperscaler, and their water footprint scales with that load.

Efficiency has moved. Modern closed-loop cooling systems operate at less than 0.2 litres per kWh, while traditional power generation still consumes around 20 L/kWh: a 100× gap that means the on-site water number understates the true watershed cost by an order of magnitude. Noise is the third footprint dimension. Larger industrial cooling units at data-center campuses can approach 100 decibels (comparable to a motorcycle or jackhammer at close range), while smaller diesel generators reach around 85 decibels, a level associated with long-term hearing damage on prolonged exposure. The environmental cost of a cluster isn't a general "they use a lot of resources" argument; it's a set of specific measurements that vary sharply by siting choice.

5.2 Developers are pushing to weaken federal pollution rules

Data-center developers aren't just building facilities; they're actively lobbying to reshape the rules that would otherwise constrain them. The Guardian reported on 27 September 2026 that developers are pushing the current administration to weaken federal pollution rules governing diesel generator emissions from backup and peaking systems. The specific stakes are visible in North Carolina, where Amazon and Duke Energy are installing 649 diesel generators to back up a new hyperscale center. Amazon and Duke's combined plans call for nearly 250 tons per year of nitrogen oxides, carbon monoxide, and volatile organic compounds, roughly 98 tons of particulate matter annually, and 6.4 tons of hazardous air pollutants including benzene and formaldehyde.

Enforcement is uneven but not absent. A New Jersey data center was fined $1.1 million after a visual investigation showed dozens of unpermitted generators: a signal that regulators are willing to prosecute when they can document facility-level violations, but also that documentation is spotty. The regulatory fight now visible in the Guardian's reporting is not a debate about whether emissions should be regulated in principle. It's a debate about whether existing rules should be relaxed to fit the pace of the AI-era buildout.

5.3 Behind-the-meter is regulatory posture as much as it is engineering

The 90 GW BTM pipeline from earlier in this piece isn't only an engineering response to grid interconnection delays; it's a regulatory posture. Projects structured as behind-the-meter deals can be arranged to sit outside standard utility and environmental review processes, which is why community groups and organisations like WRI treat BTM growth as a red flag independent of any given project's technical merits. Only 1.2% of that 90 GW is currently under construction, another 36% has been permitted, and 60% remains at the announcement or early-planning stage. Which of those cohorts becomes real infrastructure will depend on whether the review pathways that BTM structures are designed to bypass hold or get formally relaxed.

5.4 County-level review is where residents actually have leverage

Federal pollution rules are being contested. State-level siting review is inconsistent from one state to the next. But county planning boards are where residents have concrete, exercisable review power, and Pew's tracking data suggests the political appetite to use that power is rising fast. The 19-point jump in "mostly bad for people's quality of life nearby" from January to August 2026 is the largest single-category shift in Pew's dataset. Even "mostly bad for local tax revenue," historically the industry's most durable argument, moved 9 points more negative in the same window.

The relevant constituency is not small. Pew reports that 38% of Americans now live within five miles of at least one operational data center, and another 4% live within five miles of a planned one. That's more than four in ten Americans with direct proximity, and it's growing. County-level review is slow, granular, and inconsistent; it's where residents facing a proposed facility actually have a voice.

6. Where this is heading

6.1 The pipeline is 6.2× the current operating stock

The single most compressed forward-looking fact in this piece: US data centers currently operate at 61,751 MW of capacity, and 385,893 MW is planned. That's a 6.2× multiplier. Even a half-build-out scenario over the next four years doubles the operating stock. The pipeline isn't an abstraction; it's a queue of named projects with named developers and named locations. The top three planned sites alone (Delta Gigasite Expansion at 9,700 MW in Utah, New Era Lea at 7,000 MW in New Mexico, and Wonder Valley Stratos Phase 2 at 6,000 MW in Utah) represent over 22,000 MW of individually announced capacity. Texas leads the state-level active pipeline at 91 projects, followed by Virginia (70), Ohio (48), Iowa (28), and California (27).

6.2 The opinion trend, if it holds, crosses a threshold within a year

Public sentiment has moved faster than any comparable infrastructure trend of the past decade. Negative view on "environment" went from 39% (January 2026) to 54% (August 2026) in seven months. Quality of life moved from 30% to 49% in the same window. Awareness climbed from 75% to 88%. On three of five Pew categories, negative view is already the plurality. If the trajectory of the last seven months holds (and there's no obvious reason the drivers behind it will reverse), negative view on environmental impact would reach 65-70% by mid-2027. That's an extrapolation, not a forecast. But the direction is clear: data centers have moved from invisible infrastructure into the same public-opinion phase pipelines and cell towers occupied a decade ago.

6.3 The three variables that decide the next 24 months

Three variables will decide how the buildout actually lands. First: the pollution-rule fight documented in the Guardian's reporting. If the federal administration relaxes diesel-generator emissions standards, a major regulatory brake on rural, high-generator-count sites comes off. If it doesn't, the North Carolina Amazon-Duke pattern (649 diesel generators for a single hyperscale facility, with hundreds of tons per year of criteria pollutants) becomes harder to replicate.

Second: whether the 90 GW BTM pipeline reaches its projection or gets throttled. State utility commissions and transmission planners have the ability to slow BTM projects through review-timing decisions. If they do, more of the buildout ends up on the standard grid, where cost pass-through to households is direct and visible.

Third: whether the recent partisan alignment on AI-and-jobs holds. Democrats' concern that AI will lead to fewer jobs has risen 17 percentage points since 2024; Republicans' has fallen. That alignment is what currently keeps federal regulatory pressure on the buildout from mapping onto the states doing most of the buildout. A shift in that alignment changes the political economy of rural-red-state expansion fast, in either direction, depending on which party moves.