FAQ center
Questions people ask, answered from both sides
Each answer starts with what is not in dispute, then gives the strongest honest version of each side, then tells you what to check before you repeat any of it.
How much electricity data centers use
What exactly is a data center?
A data center is a huge building full of computers that run the internet, banking, streaming and AI, working around the clock. Two facts drive this whole debate: one large data center can use as much electricity as a small city, and once built it typically employs only a few dozen permanent workers. That mix of huge power use and tiny headcount is unusual, so the usual playbook for a big new employer does not apply cleanly here.
Supporters say comparing data centers to old-style factories misses the point. These buildings are the physical infrastructure behind an AI-driven economy, more like a power plant or a highway than a factory, so judging them mainly by headcount is the wrong yardstick for critical infrastructure.
Opponents say the low headcount is exactly the problem: communities are asked to host the power lines, water use, noise and traffic of a small city's worth of demand without a small city's worth of jobs or tax base in return, so the usual civic bargain does not hold.
Check it: Before accepting either framing, ask what a specific proposed project's actual permanent headcount and power draw will be, since both numbers vary a lot by facility size and design.
Sources: 1
How much electricity do data centers use in the US right now?
About 192 terawatt-hours in 2024, or roughly 4.7 percent of all U.S. electricity, according to Lawrence Berkeley National Laboratory, the federal energy lab that produces the official government count. Both sides in this debate accept this figure.
Supporters point out that even at 4.7 percent, data centers remain a modest slice of the total grid, smaller than what many other large uses of electricity draw.
Opponents say the current 4.7 percent is not the issue, it is the direction: usage has grown fast and is projected to grow much faster by 2030, straining grid capacity that takes years to build.
Check it: Check the date on any figure you see. Data center electricity estimates change quickly as new facilities come online, so a number even a year old may already be out of date.
Sources: 1
How much electricity will data centers use by 2030?
Berkeley Lab's best estimate is 649 terawatt-hours, about 11.8 percent of all U.S. electricity, with a full range of 578 to 782 terawatt-hours, 9.5 to 15.3 percent, depending on assumptions. Both sides accept this as the neutral federal estimate, though they draw very different conclusions from it.
Supporters treat the growth as proof AI and computing are becoming central to the economy and national security, and argue the grid has absorbed big jumps in demand before.
Opponents point to Virginia, where the state's own watchdog found electricity needs could double within ten years with data centers as the main driver, requiring new solar and wind at a pace the state has never achieved.
Check it: Notice this is a range, not a single prediction. The report's own authors give a range because a single confident number would not be honest, so treat anyone quoting only the middle figure as overselling certainty.
What does TWh mean when people talk about data center power use?
TWh stands for terawatt-hour, a unit for measuring electricity, the same way a gallon measures liquid. One terawatt-hour is roughly what 95,000 American homes use in a year, so when you see a number like 192 TWh, it is meant to convey scale, not anything more mysterious.
Supporters use the unit to show data center demand in context next to the country's total electricity use, arguing the numbers are large but manageable within a multi-trillion-kilowatt-hour national grid.
Opponents use the same unit to show how fast the total is climbing, arguing a jump from 192 to a projected 578 to 782 TWh in about six years is a pace the grid has never had to match.
Check it: When you see a TWh figure, ask what year it covers and whether it is measured or projected. Measured past-year figures and forward-looking projections get quoted side by side as if equivalent, and they are not.
Sources: 1
Who pays for the grid buildout
How does my electricity bill actually get decided?
A utility commission, a government referee, reviews what the power company spends on plants and wires, then divides that cost among customer groups such as homes, small businesses and big factories, since each group costs a different amount to serve. When a giant new customer like a data center needs billions of dollars of new plants and wires, the central fight is over whether it pays the full cost or whether part of it spreads across everyone else's bill.
Supporters say regulators are actively creating new rules and separate customer classes specifically so huge users pay their own way, and that this is how the process is supposed to work: rates follow real costs, reviewed by an independent referee.
Opponents say the process is opaque by design, that contracts between utilities and data centers are often sealed, and that ordinary customers and their advocates cannot fully check whether a large customer is really paying its share before their own bill goes up.
Check it: Ask whether the specific rate case or contract you are hearing about has been finalized by the regulator yet, and whether the terms are public or sealed. A proposal and an enacted rule are very different things.
What is 'cost shifting' and why is it central to this debate?
Cost shifting is when the cost of serving one customer, like a data center, gets spread across other customers' bills instead of being paid fully by the customer who caused it. Think of a group dinner: if one person orders lobster and everyone splits the check evenly, that is cost shifting. The entire fight is whether data centers are paying for their own lobster.
Supporters point to a growing set of new rules, at least 38 since 2018 according to one industry-funded study, aimed at making sure giant customers pay for their own infrastructure rather than splitting the check evenly with everyone else.
Opponents point to Harvard Law School researchers who reviewed nearly fifty real utility cases and concluded that rules meant to share the cost of keeping the lights on for everyone are being used to bill the public for equipment that serves a handful of very large companies.
Check it: Separate two different questions before judging any claim: who caused a cost, and who actually gets billed for it. Both sides agree those are different questions, and much of the confusion in this debate comes from blurring them together.
Are ordinary families subsidizing data centers' power use?
This is genuinely unresolved. A study by the consulting firm E3 found no measurable evidence to date that other customers are subsidizing data centers, and noted some large customers actually paid more than they cost. Harvard Law School researchers, reviewing nearly fifty real utility cases, argue the secrecy of these deals means the public cannot properly check whether costs are being shifted onto them.
Supporters lead with the E3 finding of no measurable subsidy, and note Texas and Virginia, the states with the fastest-growing data center demand, had among the smallest price increases, while California and New York, where demand was falling, had the largest increases.
Opponents say the E3 study was paid for by the Data Center Coalition, the industry's own lobbying group, which reviewed it before publication, and that its own fine print admits only a few studies exist and the risk 'is not absent.' They say Harvard's finding that contracts are routinely sealed means nobody can fully verify E3's conclusion.
Check it: Find out who funded the study you are being shown and whether it measured actual bills or reviewed legal filings and arguments. Those are different kinds of evidence answering different questions.
Capacity prices and grid costs
What is a 'capacity auction' and why does it affect my bill?
The grid must be ready for the single worst day of the year, usually the hottest August afternoon, so a grid operator called PJM, which serves parts of thirteen states plus Washington, DC, pays power plants a standby fee to promise to be available whether or not they actually run. That standby promise gets bought and sold at an auction, and the price is folded into everyone's electric bill in that region.
Supporters say capacity auctions are simply how grid reliability has always been paid for, and that new rules requiring big customers to pay for a large share of the capacity they reserve are designed to keep this standby cost fair.
Opponents point out that data centers being announced faster than power plants get built is exactly what pushed this auction's price up more than tenfold, and that the system was not designed with a customer this large and fast-growing in mind.
Check it: Remember that capacity auction prices are a wholesale number, what a power plant gets paid to stay ready, not your actual bill. Ask how much of your specific bill is capacity cost before assuming a spike changes your bill by the same percentage.
Sources: 18
The auction price spiked over tenfold. What did that do to home bills?
In one PJM auction, the standby price went from $28.92 to $329.17, more than ten times higher, and PJM's own independent market watchdog found data centers responsible for $6.3 billion, or 38 percent, of $16.4 billion in charges from that auction. That wholesale spike does not translate one-to-one into a home bill, but it is real money: New Jersey's own auction translated it into residential increases of about 17 to 20 percent for the year, roughly $25 a month on a typical bill, while PJM's own gentler estimate for some states was 1.5 to 5 percent.
Supporters do not dispute the $6.3 billion figure, since it comes from an independent watchdog, not activists. They argue the fix is underway: at least 38 new rules for giant customers since 2018 and Virginia's new customer class requiring data centers to pay for the capacity and wires they reserve.
Opponents say the new rules mostly have not started yet, Virginia's do not begin until January 1, 2027, while the $6.3 billion in one auction and nearly half of $63.6 billion across the last four auctions already landed on bills across thirteen states. They ask who refunds the years before the fix takes effect.
Check it: Look for your own state or utility's specific translation of the price spike into a bill increase, since the honest range spans from a few dollars a month to about $25, depending heavily on which state you live in.
Are data centers the main reason electricity prices are rising nationally?
No. Natural gas price swings, aging power plants retiring, grid upgrades, wildfire protection spending and equipment costs are all major drivers of rising electricity prices nationally. Data centers are a large and fast-growing driver in specific places, especially concentrated grid regions like PJM, but they are not the single explanation for national price trends.
Supporters point to Texas and Virginia, where data center demand grew fastest but price increases were among the smallest, while California and New York saw the biggest increases with demand actually falling, as evidence data centers are not the main national story.
Opponents agree data centers are not the sole cause everywhere, but say that concession should not be used to wave away the PJM watchdog's finding that data centers were the single largest driver in that specific, thirteen-state region, where the impact is concentrated and real.
Check it: When you see a claim about why bills are rising, check whether it is about the whole country or one specific grid region. National averages and PJM-specific findings tell different stories that get mixed together constantly.
Secret contracts and disclosure
Can the public see what data centers actually pay for electricity?
Often no. Utilities regularly ask regulators to keep the contracts between them and data centers confidential, which means the public and even consumer advocates cannot fully verify whether a large customer is paying its fair share before rates go up. Harvard Law School researchers who reviewed nearly fifty real utility cases found this secrecy limits what regulators can realistically scrutinize.
Supporters tend to concede this point, since defending secrecy is a weak position, and many say they would support making these contracts public. They separately argue that even without full visibility, evidence like the E3 study found no measured harm to other customers.
Opponents say secrecy is the problem regardless of whether harm is later proven, since a sealed contract means the public is asked to trust, not verify, and 'we found no evidence in filings we cannot fully read' is not the same as 'there is no cost shifting.'
Check it: If you are told a rate increase is justified partly by a data center's contract, ask whether that contract, or at least its key terms, is public. If it is sealed, note that as a fact on its own, separate from any suspicion of wrongdoing.
Jobs
How many permanent jobs does a data center actually create?
Roughly 100 to 200 permanent jobs per county, according to Brookings Institution economists who compared about 1,500 built facilities against 52 similar projects that were announced and then canceled. The same study found computing employment rose 56 percent and telecom employment rose 43 percent in host counties over a decade, but wages did not rise.
Supporters use this study themselves rather than overstating the number, and pair it with the industry's own PwC-funded study claiming the sector supports 5.5 million jobs nationally through supply chains, while noting that broader number is modeled, not counted.
Opponents say 100 to 200 jobs is a small return for the billions in tax breaks some states give up, and that the same Brookings study found flat wages and home values rising an extra 2 to 5 percent, a cost for renters and first-time buyers.
Check it: Distinguish a measured county-level jobs number from a modeled economy-wide multiplier. Ask whether a jobs claim was counted directly, like Brookings did, or estimated with economic modeling, like the industry-funded PwC study.
What about construction jobs? Are those real?
Yes, and this is the strongest jobs evidence for the pro-data-center side. Local 26 of the electricians' union, covering Washington DC, Maryland and Virginia, reports its members worked 28 million hours last year, double a decade earlier, with data centers accounting for at least half of that, plus 600 new apprentices in a single year.
Supporters emphasize this is organized labor making the argument, not a company press release, and that an apprenticeship is a credential that outlasts any single job site, even as the buildout is a rolling, decade-long pipeline rather than one short spike.
Opponents accept these construction numbers are real but say construction work is temporary by nature. It rolls until the buildout slows and then it stops, while the electricity demand and grid costs the buildings leave behind do not stop with it.
Check it: Check whether a jobs figure describes construction-phase work, which ends, or permanent operational staff, which does not. The two get blended together in ways that can make a facility's lasting local impact look bigger than it is.
Where does the industry's '5.5 million jobs' figure come from?
It comes from a PwC study commissioned by the Data Center Coalition, the industry's own lobbying group, using economic modeling that estimates each direct job supports 4.5 more jobs across construction, utilities, logistics and services. It measures gross economic activity, not net new jobs, and it is a model, not a headcount.
Supporters cite it as evidence the industry's footprint extends far beyond the buildings themselves, into suppliers, utilities and services nationwide, arguing a narrow focus on on-site headcount misses most of the real economic effect.
Opponents point out this figure was paid for and commissioned by the industry itself, that it uses input-output economic modeling rather than counting actual jobs, and that the much smaller, measured Brookings finding of 100 to 200 jobs per county suggests most of that modeled multiplier lands somewhere other than the host community.
Check it: Ask whether a jobs number was measured, by comparing real places over time like Brookings did, or modeled, by applying an economic multiplier like PwC did. The Brookings study is the one number both sides can honestly cite.
Does a data center's investment size guarantee real local benefit?
Not necessarily. Sean O'Leary at the Ohio River Valley Institute argues data centers are capital-heavy and labor-light by design, so large investment figures do not reliably translate into local benefit. His evidence is regional history: the thirty biggest gas-producing counties in Appalachian Ohio, Pennsylvania and West Virginia posted far better than average economic growth during the fracking boom and still lost jobs and population.
Supporters note the Ohio River Valley Institute study is about the natural gas industry, not data centers directly, making it a comparison rather than a direct measurement, and point to the more directly relevant Brookings study finding real, measured job and income gains in data center host counties.
Opponents say the parallel is exactly the point: an industry can post enormous investment numbers on paper while the community around it sees little translate into jobs or population growth, and data centers, with even less labor per dollar invested than gas extraction, are structurally likely to repeat that pattern.
Check it: Ask whether a specific economic benefit claim is a direct study of data centers, like Brookings, or an analogy from a different industry, like the gas-boom comparison. Both are legitimate evidence, but they carry different weight.
State tax breaks
Do state sales tax exemptions for data centers pay for themselves?
Generally no, by the states' own numbers. Virginia's nonpartisan watchdog found the state got back 48 cents for every dollar given up in tax exemptions over a full decade, 2014 through 2023. Good Jobs First, using states' own required financial disclosures, found states that measured their return lose between 52 and 70 cents on every dollar given away, and Georgia, Texas, Virginia and Ohio each have exemptions capable of costing more than a billion dollars a year.
Supporters note the 48 cents is nearly three times better than the 17-cent average return of Virginia's other sales tax exemptions, and that a 2026 state evaluation ranked the data center break Virginia's second-best incentive, producing $6.10 of labor income for every exempted dollar.
Opponents say the strongest pro-side employment study found tax incentives make up only 2 percent of investment in large company-owned facilities, meaning the companies getting the biggest breaks barely needed them to decide to build, while the exemption renews itself every time a data center replaces its computers, which happens every three to five years.
Check it: Check whether a figure like $6.10 of labor income measures money flowing through the economy, mostly during construction, or money that actually lands back in the state treasury, which is what the 48-cents figure measures. Those are different questions that get swapped mid-argument.
How much are individual states giving up in data center tax breaks?
Ohio's exemption cost jumped from $555 million in 2024 to $1.6 billion in 2025, roughly $330 per Ohio household in a single year. Indiana gave up $655 million, with more than 83 percent going to one company, Amazon. Texas gave up $1 billion in one year, and Oregon handed $616 million in property tax breaks to facilities owned by Amazon, Apple, Alphabet and Meta. Fourteen states with these exemptions publish no annual total at all.
Supporters argue these figures reflect the scale of investment the exemptions are attracting, and note the money represents value the state chose not to collect rather than money spent, with real construction and tax revenue arriving in return.
Opponents note most of these exemptions were written years ago, before AI arrived, with no caps and no annual legislative vote, so costs exploded on autopilot. They also flag that fourteen states publish no annual total at all, making the true national cost unknowable.
Check it: Look for whether a state's data center tax exemption has an annual cap or requires a legislative vote to renew. Many do not, which is part of why costs can jump dramatically from one year to the next without anyone formally deciding to spend more.
Sources: 20
Local property tax revenue
Do data centers really lower property taxes, as in Loudoun County, Virginia?
Yes, this is documented, not a projection. Data centers produced roughly $1.2 billion in property tax revenue in Loudoun County in fiscal year 2026, about 39 percent of the county budget and nearly half of all property tax collected. The county cut its property tax rate every year for a decade, from $1.145 per $100 of assessed value in 2016 to $0.805 in 2026, and average homeowner bills went down even as home values rose.
Supporters say this proves the tax benefit is real and already happening, not hypothetical, and that families in Loudoun County pay less than they otherwise would because one industry covers a large share of the cost of schools, police and roads.
Opponents do not dispute the numbers, which are public budget records, but say Loudoun is the most favorable example on earth and proves little about other counties. They call 39 percent of a budget depending on one industry a concentration risk, like a household living on one paycheck, and note Loudoun residents are among the loudest opponents of new projects in the country.
Check it: Loudoun County's figures are public budget documents anyone can look up, so do not dispute the numbers themselves. Instead ask how typical Loudoun's circumstances are compared to the county or state you actually care about.
Water use
How much water does it take to train an AI model?
Researchers at the University of California, Riverside, published in a peer-reviewed computing journal that training one AI model in Microsoft's U.S. facilities could directly evaporate on the order of 700,000 liters, about 185,000 gallons, of clean freshwater. They projected global AI water use could reach 4.2 to 6.6 billion cubic meters by 2027, more than several European countries withdraw in a year. They also say company disclosures typically count only water used at the building, leaving out water used to generate the electricity elsewhere.
Supporters note any single AI question or interaction costs only a small sip of water, and that the biggest technology companies are actively redesigning cooling systems, including some newer designs that use no water at all.
Opponents say the argument was never about one glass of water, it is that billions of glasses are used in specific places, some already short on water before facilities arrived, and that counting only building-level water understates the real total.
Check it: Ask whether a water figure counts only water used on-site at the data center, or also the water used to generate its electricity elsewhere. The gap between those two counting methods is the core of this dispute.
Sources: 23
Microsoft and Google say they are cutting water use a lot. Is that true?
According to their own self-reported figures, yes. Microsoft says it put back more freshwater than it took in during 2025, five years ahead of its own target, cut data center water use roughly 90 percent compared to its earliest facilities, and designs newest facilities to use no water at all for cooling in normal operation. Google has committed to a similar water goal by 2030.
Supporters point to this as proof the industry is engineering out the water problem, replacing evaporative cooling with closed-loop systems that use none, arguing the newest designs show where the industry is headed.
Opponents note these are company-reported, unaudited figures from companies with a direct interest in looking good, and that the peer-reviewed water research exists precisely because such disclosures typically miss the water used to generate electricity. New designs also do not retrofit hundreds of older buildings already running.
Check it: Check whether a water reduction claim is independently audited or self-reported, and whether it applies only to newly built facilities or to a company's entire existing footprint.
Air quality and on-site power
What happened with xAI's gas turbines in Memphis?
In South Memphis, xAI installed gas-burning turbines next to Boxtown, a mostly Black neighborhood already surrounded by industrial polluters, in a county that already fails the federal health standard for smog and that the American Lung Association graded F for ozone. Hundreds of residents attended the public hearing, and every single person who spoke opposed it, but the air permit was issued anyway. The NAACP later sued over illegal pollution, and 2026 reporting documented 59 turbines running without permits at a second nearby site.
Supporters frame Memphis as one company at one site behaving badly, the kind of case siting standards are meant to prevent, not evidence of what the industry does everywhere. They note operators use on-site turbines because it is faster than waiting for grid connections.
Opponents say Memphis shows what happens when siting standards are not enforced: a permit process existed, hundreds of residents used it and were unanimous in opposition, and it did not change the outcome. They call the pattern, pollution equipment near a neighborhood already failing air standards, foreseeable, not a fluke.
Check it: This case involves an active lawsuit, so separate facts that are independently checkable, like permit terms, turbine counts and the Lung Association's grade, from characterizations made by one side in the case, like the Southern Environmental Law Center.
Sources: 24
Noise, appearance and rural character
Are data centers noisy, and can anything be done about it?
They can be. Hundreds of rooftop cooling fans run around the clock, and backup generators are tested regularly. In the worst documented case, the Great Oak neighborhood next to an Amazon data center south of Manassas, Virginia, residents measured up to 65 decibels at night, about a vacuum cleaner in the next room. After two years of pressure Amazon replaced all 424 rooftop exhaust tubes and cut the noise roughly in half, to around 50 decibels. Most data centers are not that loud, but the low, steady hum is the single most common quality-of-life complaint at county hearings.
Noise is a design problem with known fixes: taller exhaust stacks, sound walls, setbacks and generator testing windows. Chandler, Arizona has required a pre-construction sound study, mitigation to the baseline and testing limits since January 2023, and Loudoun County now requires a special exception with conditions for every new data center. Written into the approval, those conditions make a data center a quieter neighbor than most industrial uses, with almost no truck traffic.
Every fix arrived after the building did. Great Oak waited two years for a fix that still fails the county's proposed limit of 52 decibels by day and 47 at night, the low-frequency hum that carries into houses is not regulated by the old rule or the new one, and emergency generators are exempt. Loudoun admits it has no rule for tonal or low-frequency noise and does not expect one until 2027. Chandler took nine years from the first complaints to an ordinance.
Check it: Ask what the noise limit is, whether it is measured in A-weighted decibels only (which understate low-frequency hum), whether generators are covered, and whether the standard applies to buildings already approved. A county-wide average or a company's own reading tells you little about the houses on the fence line.
Do data centers ruin the look of a place, and does that ever stop a project?
A data center is a windowless building the size of several football fields, often with fencing, lighting and a substation, and it stays for decades. Appearance and rural character are among the first objections raised at hearings, though on their own they rarely decide a case. They matter most when combined with scale: the Prince William Digital Gateway, 37 buildings described as the equivalent of 144 Walmart Supercenters beside Manassas National Battlefield Park, was voided by Virginia's Court of Appeals in March 2026 and abandoned that July, after planning staff had objected that the visual studies could not be trusted without drawings of the actual buildings.
Looks are a siting and design question, not an argument against the industry. Loudoun County is writing use-specific standards that name aesthetics and compatibility, special exception approvals can require screening, berms, setbacks and facade treatments, and the one careful study of home values found they rose 2 to 5 percent more in counties that got a data center than in comparable counties. The market did not treat the buildings as a blight.
A county-wide home value average says nothing about the homes that share a fence with the building, and a rural community that chose open land is not compensated by a facade treatment. Forsyth County's planning board recommended denying a hyperscale project in Rural Hall in June 2026 over rural character, and Stokes County residents sued when a zoning change opened a dozen rural sites. The standards that would govern appearance in Loudoun are not expected until 2027, after most of the buildout.
Check it: Separate the visual complaint from what actually decided the case: the Digital Gateway rezoning was thrown out over how the county ran its hearings, not over the view. Ask whether the county has adopted design standards or only announced a plan to write them, and whether the home value figure quoted is county-wide or fence-line.
Local opposition and consent
How much local opposition is there to data center projects?
A great deal, and it is growing quickly. Data Center Watch, a tracker run by a firm whose clients are AI companies, recorded $64 billion in projects blocked or delayed through 2025, then about 75 projects worth roughly $130 billion in the first three months of 2026 alone, matching the entire previous year in a single quarter. Opposition groups more than doubled across 49 states, and fourteen state legislatures took up pause proposals in three months.
Supporters point out the same period set an all-time construction record, $44.7 billion actually spent in the first quarter of 2026, with full-year investment forecast near $700 billion, arguing most projects still go through and opposition is not stopping the buildout.
Opponents say the scale is the story: this much opposition, spreading this fast across so many states, is not a handful of critics at a zoning meeting, and market analysts now estimate 30 to 50 percent of the 2026 pipeline is on track to be delayed or canceled.
Check it: Notice that record construction spending and record opposition happened in the very same quarter. A claim that cites only one of those two facts is telling you half the story.
Who runs the Data Center Watch tracker, and can its numbers be trusted?
Data Center Watch is run by a firm called 10a Labs, whose stated clients are AI companies and large technology firms. It does not disclose who specifically funds this research. Because the firm's own commercial relationships give it every incentive to make opposition look smaller, its finding of $130 billion in blocked or delayed projects in a single quarter is considered hard for the industry to wave away.
Supporters can still use these numbers, since the firm's incentives run toward undercounting opposition, but note the same data also shows record actual construction spending happening alongside the opposition, meaning most projects still get built.
Opponents say the funding is the argument: a tracker built by and for the AI industry finding this much opposition is more credible, not less, than if a group with the opposite incentive had produced the same numbers.
Check it: Ask who funds any tracker or study before trusting its numbers, and specifically whether its funder's incentive runs toward inflating or deflating the finding shown to you. Here the incentive runs against the number reported, which is part of why it carries weight.
Sources: 25
Flexible load
What does 'flexible load' mean?
Flexible load means a data center agrees to ease off its power use during the small number of hours each year when the grid is most strained, like the hottest August afternoon, instead of running at full power around the clock. Because the grid is built for that one worst day and sits mostly underused the rest of the year, a customer willing to step back during peak hours can use existing spare capacity without anyone building expensive new power plants and wires.
Supporters call this the strongest argument for the pro-data-center side, since it lets huge new demand get served using equipment already in the ground, and point to live tests already running at nine sites with 49 participants including major tech companies and grid operators.
Opponents agree the concept is sound but say it means little as currently practiced, since agreeing to ease off is voluntary today and most contracts do not require it. A promise to maybe step back is not the same as guaranteed grid capacity.
Check it: Ask whether a specific data center's contract makes flexible load a binding requirement or a voluntary option. That difference changes what the whole concept is actually worth.
How much extra demand could the grid handle with flexible load?
Duke University researchers found that roughly 100 gigawatts of new demand, an enormous amount, could be added across grid regions serving about 95 percent of the country using equipment already in the ground, as long as those customers accept short, occasional cutbacks during peak hours. This is university research, not funded by industry, and it is the most downloaded study the sponsoring institute has published in twenty years.
Supporters say this shows the grid has far more room than commonly assumed, and that the finding has already moved from paper to practice with live tests at nine sites across the U.S. and Europe with 49 participants including major tech and utility companies.
Opponents say the entire 100-gigawatt finding depends on data centers actually easing off at peak hours, which is voluntary today, and most contracts signed right now do not require it. Without a binding commitment, the number describes potential capacity, not guaranteed capacity.
Check it: Check whether a cited 100-gigawatt figure is described as available today or as conditional on a policy change, mandatory flexibility, not yet widely adopted. It is a real finding about potential, not a current guarantee.
Efficiency and forecasts
Have predictions about data center energy use been wrong before?
Yes, once. A peer-reviewed paper in the journal Science showed widely repeated predictions from the 2000s that data center energy use would double, triple or quadruple were wrong, because efficiency gains, mainly consolidating computers out of thousands of small server rooms into a few large, efficient buildings, held global data center electricity use nearly flat from about 2010 to 2018 even as computing output grew enormously.
Supporters use this history as a reason for humility about today's scary forecasts, since the last set of confident predictions was wrong, and the same underlying force, engineers finding new efficiency gains, is still at work.
Opponents say the comparison does not hold up: the 2010 to 2018 savings came from a one-time consolidation move that can only happen once, and it has already happened. The same authors who wrote that paper now publish the projections showing steep growth from 192 TWh toward 578 to 782 TWh by 2030.
Check it: Confirm the exact years any efficiency claim covers. 'Data centers got more efficient before' is true for 2010 to 2018 specifically, and using it to describe the current AI-driven period without that caveat is misleading.
Do all announced data center projects actually get built?
No. An analyst at the Information Technology and Innovation Foundation found announced projects consistently overstate what actually gets built: deals fell more than 40 percent in late 2025, and only about a third of announced projects are actually under construction. Separately, market analysts estimate 30 to 50 percent of the 2026 pipeline is on track to be delayed or canceled.
Supporters use this to argue alarming forecasts based on announced capacity overstate the real-world grid impact, since a large share of announcements never turn into finished, power-drawing buildings.
Opponents note that even if not every project gets built, the same quarter these delay estimates came from also set an all-time construction spending record of $44.7 billion, with full-year investment forecast near $700 billion, so the buildout overall is not slowing, only some individual projects are.
Check it: When you see a headline data center capacity figure, check whether it describes projects that are announced, under construction, or actually operating. Those are different stages, and announcement totals are the least reliable predictor of final impact.
AI race and national security
Is the data center buildout really a national security issue?
This is a contested framing, not an agreed fact. Industry trackers estimate the United States holds roughly 75 percent of the world's AI computing capacity against about 15 percent for China, and U.S. AI computing capacity is forecast to grow from 62 gigawatts in early 2026 to 152 gigawatts by 2030. Advocates for this framing argue a chip only becomes usable capability once installed in a powered building, so the data center is what turns America's chip advantage into something real.
Supporters argue whoever has more of these buildings trains stronger AI models faster and can deploy AI across more of its economy and military, and that of the four pillars of AI strength, energy, compute, talent and adoption, compute is the one pillar America still clearly leads.
Opponents do not necessarily reject the underlying facts but note the framing assumes AI capability keeps scaling with computing power the way it has so far, which is not guaranteed, and that the same framing concedes China's electric grid is growing faster than America's.
Check it: Notice that the group most often cited making the strongest version of this argument, the American Edge Project, is a technology industry advocacy group funded by tech companies. That does not make the underlying compute-share estimates wrong, but the framing built on them is advocacy, not neutral analysis.
Sources: 30
Who holds the risk
If a data center project is cancelled, who pays for the plant built for it?
It depends on the contract. If the data center must pay regardless of whether it uses the power, called a minimum take contract, the company holds the risk. If not, the cost can land on everyone else's bill. Virginia's regulator created a rule in November 2025 requiring large customers, those using more than 25 megawatts, to sign fourteen year contracts and pay for at least 85 percent of the wires capacity and 60 percent of the generating capacity they reserved, whether they use it or not, starting January 1, 2027.
Supporters point to Virginia's new rule as proof the fix is being installed: a customer that reserves the table pays for the table, even if it never shows up or the project falls through, which directly answers the who-holds-the-risk question.
Opponents note Virginia's protection does not start until January 1, 2027, years into the buildout, so it does not cover plants already built or being built now. Consumer and land conservation groups also argued the regulator ducked the harder question of who pays for transmission lines specifically.
Check it: Ask what date a risk-allocation rule takes effect and whether it applies to contracts signed before that date. A rule covering only future contracts leaves existing risk exactly where it was.
Sources: 9
What is Virginia's new rule for very large electricity customers?
In November 2025, Virginia's State Corporation Commission created a separate customer class for anyone using more than 25 megawatts of power, which mainly means data centers. Those customers must sign fourteen year contracts and pay for a large share of the capacity and wires they reserve whether they use it or not, starting January 1, 2027. It is the concrete example of regulators writing a new rule to keep a giant customer's costs with that customer.
Supporters call this proof that cost-shifting concerns are being actively addressed through real regulation, not just discussed, and note Virginia has the most data centers in the country, making this the highest-stakes test case.
Opponents note the rule does not start until 2027 and only applies going forward, so it does not retroactively fix any cost shifting from projects already connected. They also cite consumer and land conservation groups who say the commission avoided ruling on transmission line costs specifically.
Check it: Look up whether the rule you are hearing about is in effect yet or scheduled for a future date, and whether it was adopted by the regulator, as the Virginia rule was, versus merely proposed.
Sources: 9
Siting, zoning and standards
If my county blocks a data center, does that stop it being built elsewhere?
No. Demand for computing comes from customers worldwide, not from any single local zoning board, so turning down a project does not reduce how much gets built nationally. It generally shifts where a project gets built, along with the associated tax revenue, jobs and rules governing it, to a different community.
Supporters argue the only thing a community actually controls is whether a project gets built inside its borders under its own rules and taxes, not whether the underlying demand for computing exists, so refusing a project trades away local revenue and control rather than preventing the building from existing somewhere.
Opponents say this argument, taken to its end, becomes a race to the bottom where every community feels pressure to accept weaker standards rather than lose the project to a neighbor, and argue it should be treated as a case for consistent minimum standards everywhere, not a reason for any one community to waive its own.
Check it: Ask what specific siting standards, on water use, air permits, noise or setback distance, the project already has to meet in your area, and whether blocking it would actually change the standards it meets elsewhere or just move them to a different zip code.
Laws, bills and regulation
What did the federal data center permitting executive order do?
In July 2025, alongside a national AI action plan with more than ninety actions to speed construction, the administration issued an executive order instructing federal agencies to exempt qualifying data center projects from standard federal environmental review, put them on an accelerated permitting track, and open federal land for their construction.
Supporters frame this as removing unnecessary federal delay for infrastructure the government has determined is central to the economy and national security, arguing qualifying projects should not be held up by review processes not designed for this kind of build.
Opponents say the provision worth targeting is the environmental review exemption specifically, since that review is normally the step where local communities and objections get formally heard, and removing it federally makes local opposition, already large, a community's main remaining check on a project.
Check it: These are official policy documents, so cite them as statements of current federal policy, not independent research. Their economic assumptions and projected benefits are contested by other sources in this debate.
What is the Sanders and Ocasio-Cortez data center moratorium bill?
In March 2026, Senator Bernie Sanders and Representative Alexandria Ocasio-Cortez introduced the AI Data Center Moratorium Act, a bill for a national pause on data center construction. It is a political announcement describing the most restrictive position formally proposed in this debate, and it is not expected to pass.
Supporters of the buildout treat this bill as the maximum possible anti-data-center position, and argue most opponents of specific projects do not actually support a nationwide construction ban, making the bill easy to paint as outside the mainstream.
Even many opponents of specific projects do not endorse a full national moratorium, arguing instead for narrower, more defensible tools like mandatory flexible load requirements, public contracts, caps on tax breaks and real siting standards, since a total ban is unlikely to pass.
Check it: Check whether an opposition argument you are hearing is actually calling for a full moratorium like this bill, or for narrower reforms such as more transparency or stricter siting rules. Conflating the two makes it easy to unfairly dismiss more modest opposition.
Sources: 26
No questions match. Try another word.