The Data Center Debate

Arguing against

The case against data centers

Eight arguments opponents make, each with the evidence behind it, the best response from the other side, and the reply. Excerpted from the handbook, which is written as a guide for arguing a side, so the voice is 'you'.

This page presents one side as its own advocates would. The other side's response is included under every point, and the opposite page does the same in reverse. Numbers in superscript link to the source list, which says who paid for each document.

1 Regular families are paying part of the bill, and the deals are kept secret so nobody can check.

Last updated October 13, 2026 by the daily research pass. Tracks: Who pays for the grid buildout, Secret contracts and disclosure.

The claim in one line

When a data center needs new power lines and power plants built for it, somebody has to pay. A lot of that cost is landing on ordinary customers instead of on the data center. And the contracts are sealed, so the public cannot verify it.

What it means

Go back to the group dinner. The power company and the data center negotiate a private deal on how much the data center will pay. Then the power company asks the referee to keep that deal confidential, meaning the public is not allowed to see it. So when your bill goes up, you cannot check whether the big customer paid its share, because the receipt is sealed.

The evidence

Two researchers at Harvard Law School, Eliza Martin and Ari Peskoe, read through nearly fifty of these utility cases. They concluded that rules originally built to share the cost of keeping the lights on for everybody are now being used to make the public pay for equipment that exists to serve a handful of extremely rich companies. Their sharpest finding is about the secrecy: utilities routinely ask for confidential treatment, which limits public scrutiny and narrows what the referee can realistically do. In their words, the complexity of the process conceals utilities overcharging captive customers in order to funnel money to their own side businesses.16

How big is it, really

Unknowable, and that is the entire point, so be honest about both edges of it. A sealed contract is not proof of a bad deal. It is proof that nobody outside the room can check the deal, in a system where you are legally required to buy the product. Concretely: when your utility asks the referee for a rate increase, the data center's contract is in the file with the numbers blacked out. Your consumer advocate argues about a document they cannot fully read. So argue the checking, not the cheating. The moment you claim to know what is inside the sealed deal, you are guessing, and they will catch you guessing.

How this side says it

"I do not have to prove the exact dollar amount. The public is being handed a bill it is not allowed to inspect. That by itself is the problem."

Why this lands

It does not require you to win a math fight. It is an argument about fairness and transparency, and everyone understands both.

What the other side says back

"Harvard read legal paperwork and made a fairness argument. It never measured a single dollar actually moving onto anyone's bill. The one study that tried to measure it, by the consulting firm E3, found no evidence that families are subsidizing data centers."8 A smart opponent will then concede the secrecy and say the contracts should be public, to take the sting out of your point.

The reply

Their measuring study was paid for by the Data Center Coalition, the industry's own lobbying group, which reviewed it before publication. It is printed on the report's acknowledgments page. And the report's own fine print admits only a few studies exist, most show correlation rather than cause, and the risk "is not absent."8 Say: "Not finding something in a nearly empty filing cabinet is not proof it is not there." Then, if they conceded the secrecy, do not let it float away as a nice gesture. Pin it: "Then you will support a rule making every one of these contracts public, starting now?" If sealed deals were fair deals, nobody would need to seal them.

What has changed since the handbook

  • October 13, 2026 Helps the case against The October 9 Senate Democratic report concludes that Amazon, Google, Meta and Microsoft continue to request nondisclosure agreements from utilities, landowners and some local officials despite recent public pledges by Microsoft and Amazon to stop. That shifts the public evidence on this argument toward the view that voluntary rollbacks have not ended the practice.
  • October 13, 2026 Helps the case for The Indiana regulator's October 7 approval of the AES Google deal in Monrovia is a worked example of a but-for allocation: Google is contractually responsible for the roughly 1.3 billion dollars in solar, storage and transmission needed to serve its site. If the structure holds, it supports industry arguments that large customers can be charged for the costs they cause without automatic ratepayer exposure.
  • October 11, 2026 Helps the case against The U.S. Department of Energy made its first FERC filing in five years to push PJM to reallocate the costs of its 6.8 GW capacity backstop so new data centers pay rather than existing households. Even an administration generally friendly to data center buildout now agrees the current cost allocation would shift costs onto ratepayers.
  • October 10, 2026 Helps the case against The Arizona attorney general's October 8 request that regulators investigate whether APS's Power Supply Adjuster allocated about $587 million of large-user costs to other customers from 2023 to 2025 moves what was an advocacy analysis into a regulatory question. APS disputes the figure and no commission finding has been made.
  • October 9, 2026 Helps the case against Duke Energy and all major hyperscalers filed a proposed large load tariff settlement in North Carolina on October 7 to recover costs from customers using 50 megawatts at least 80 percent of the time. The deal still needs commission approval, and consumer and environmental groups say it leaves cost gaps.
  • October 8, 2026 Helps the case for Pennsylvania's HB2828 moved unanimously out of the House Consumer Protection, Technology and Utilities Committee on October 6. It would make data center developers apply for electric service and pay infrastructure fees before seeking zoning, strengthening the case that ratepayer protections can be codified rather than left to utility tariffs.
  • October 6, 2026 Helps the case against The Minnesota attorney general's October 1 PUC filing on the Xcel-Google Pine Island deal is the first state-level official estimate putting a specific nine-to-ten-figure cost-shift risk on a single data center ESA, warning of about $1 billion in exposure for other customers if the deal is approved as filed.
  • October 5, 2026 Helps the case against Minnesota's attorney general filed formal comments with state regulators arguing a specific Xcel-Google special contract would leave other customers on the hook for potentially more than a billion dollars because the deal does not fully assign all costs to the data center as a very large customer under state law. The filing adds a state-level enforcement voice to the ratepayer cost-shift debate.
Latest evidence tagged to this argument

Tagged automatically by the daily update from the updates feed. New evidence does not rewrite the argument above; read both and judge whether the point still stands.

Sources cited in this point: 8paid for, 16

2 The grid's own watchdog blames data centers for billions in higher costs.

Last updated October 12, 2026 by the daily research pass. Tracks: Capacity prices and grid costs.

The claim in one line

The standby fee everyone pays to keep power plants ready went up more than ten times over. The independent referee who polices that market says data centers are the biggest reason.

What it means

Remember the babysitter retainer. Across thirteen states plus Washington, DC, there is a grid operator called PJM that runs this standby auction. Every household in that region pays into it. The price exploded because expected demand jumped while new power plants did not arrive fast enough.

The evidence

PJM has an independent watchdog, a firm called Monitoring Analytics, whose entire job is checking whether that market is run fairly. It reports to federal regulators, not to PJM's management. It found data centers responsible for $6.3 billion, or 38 percent, of $16.4 billion in charges from one recent auction, and that across the last four auctions, data center related charges came to nearly half of $63.6 billion in total.17 The standby price went from $28.92 to $329.17 and then hit its ceiling of $333.44. PJM expects peak demand to grow by 32 gigawatts by 2030, and all but 2 of those gigawatts come from data centers.18

How big is it, really

The PJM region serves about 65 million people. Spread $6.3 billion across them and it is nearly $100 per person, for one year of standby charges alone, before anyone uses a single unit of electricity. In practice it does not arrive evenly, because businesses pay a large share first and pass it along in prices, but the money all starts on that region's electric bills. On a single home's bill, the whole price spike showed up as somewhere between a few dollars and roughly $25 a month depending on the state, with New Jersey households near the top of that range.27 Small enough that many families never noticed why, large enough that it is real money, every month, with no end date announced.

How this side says it

"That is not an environmental group saying this. That is the official market watchdog, the one whose job is policing the grid operator itself. They put a number on it. $6.3 billion in a single auction."

Why this lands

The source is nearly impossible to smear. The watchdog has publicly criticized PJM itself, so it is nobody's ally.

What the other side says back

The group dinner distinction, used against you. "The watchdog counted who caused the cost, which nobody disputes. What a family actually pays depends on how the check gets split, and that is exactly what the new rules govern: at least 38 special rules for giant customers since 2018, and Virginia's new customer class that makes data centers pay their own way."8 9 They will also note the watchdog blames other things too, like old power plants retiring and the market's own design.

The reply

Timing. The auction charges are not a forecast, they already happened and already landed on bills across thirteen states: nearly half of $63.6 billion over the last four auctions.17 The rules they are bragging about came after the money moved, and the Virginia protections do not even start until January 1, 2027.9 Say: "You are describing a fix for a check that has already been paid. Who refunds the last four years?" On the other causes: agree those exist, then repeat that the watchdog still named data centers the primary reason, and it is the referee, not you, saying so.

What has changed since the handbook

  • October 9, 2026 Helps the case against The Federal Reserve's September 15 to 16 meeting minutes, released October 7, identified the AI and data center buildout as a growing driver of goods inflation, overtaking tariffs in the committee's narrative. It gives those arguing data centers raise costs beyond electric bills a direct macro citation.
  • September 17, 2026 Helps the case against The Pennsylvania PUC's newly released Synapse study projects PJM will fall below its reliability planning standard from 2027-2030 and could hit more than 13 loss-of-load hours per year by 2030 under high data center growth. That gives opponents a state-agency source, not a modeling firm alone, to cite when arguing that current interconnection practice cannot absorb pending demand.
Latest evidence tagged to this argument

Tagged automatically by the daily update from the updates feed. New evidence does not rewrite the argument above; read both and judge whether the point still stands.

Sources cited in this point: 8paid for, 9own account, 17, 18, 27own account

3 The jobs do not show up, and states are paying enormous amounts for the few that do.

Last updated October 12, 2026 by the daily research pass. Tracks: Jobs, State tax breaks.

The claim in one line

These buildings are designed to run with almost nobody inside. Meanwhile states are giving away more in tax breaks than they collect back.

What it means

There is an old assumption that a big investment means a lot of jobs. It is usually true. It is not true here. The money goes into computers and cooling equipment, not payroll. So a county can host billions of dollars of construction and end up with a permanent headcount you could fit on a school bus.

The evidence: part one, the jobs

Sean O'Leary at the Ohio River Valley Institute argues data centers are capital-heavy and labor-light by design, so they create few jobs and put little money into the local community, and that the impressive investment figures simply do not 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 still lost population.19 Big numbers on paper, empty storefronts in real life.

The evidence: part two, the tax giveaways

Good Jobs First tracks these deals using the states' own legally required financial disclosures, so these are governments' own books, not an advocacy group's estimates. Georgia, Texas, Virginia, and Ohio each have exemptions capable of costing more than $1 billion a year. Ohio's cost jumped from $555 million in 2024 to $1.6 billion in 2025, which works out to roughly $330 per Ohio household in a single year (that per household figure is our arithmetic on the state total, not a number Good Jobs First published). Indiana gave up $655 million, with more than 83 percent of it going to a single company, Amazon. Texas gave up $1 billion in one year. Oregon handed $616 million in property tax breaks to facilities owned by Amazon, Apple, Alphabet, and Meta. And states that actually measured the return found they lose between 52 and 70 cents on every dollar they give away. Fourteen states with these exemptions publish no annual total at all.20

What that 52 to 70 cents actually measures, because this is where the honest fight is

The natural objection is: "Sure, the state skipped the sales tax, but now there is a building paying property tax that paid nothing before, plus income tax from workers. Doesn't the math flip after a few years?" Virginia ran exactly that experiment. Its watchdog added up a full decade, 2014 through 2023, of everything the state got back because the buildings existed: income taxes from the workers, corporate taxes, sales tax on everything not exempted, all of it. The answer was 48 cents returned for every dollar given up.22 Ten years in, the flip has not come, and here is the reason it does not come with time: the exemption is not a one-time discount on building the place. It covers the computers, and data centers replace their computers roughly every three to five years. Every refresh is a new tax-free purchase, forever, for as long as the building runs. The giveaway renews itself on the same schedule the benefit does.

Now the part of the objection that is genuinely right: the property tax is real, and it goes to a different pocket. The sales tax exemption is the state's loss. The property tax windfall belongs to the county, which is exactly why Loudoun County can cut its tax rate ten years running while Virginia as a state gets back 48 cents on the dollar. Both things are true at once. When you argue this point, say which pocket you are talking about, because your opponent will happily switch pockets mid-sentence.

And the last piece, the "without the exemption there would be no building at all" argument: that is only true if the tax break is what won the building. The best pro-side study found the incentives are just 2 percent of the money behind the big company-owned facilities.21 Those companies pick sites for power, land, and fiber. For them, the exemption is a bonus on a decision already made, and a bonus for a decision already made is pure loss.

The evidence: part three, and this is your best shot

Use their own study. The strongest pro-data-center employment research, by two economists at Brookings, found that tax incentives make up only 2 percent of the investment in the giant company-owned facilities, but 62 percent of the investment in the rental facilities.21 Translated: the buildings that produce the better jobs barely need the subsidy, and the subsidy is flowing overwhelmingly to the ones that produce fewer.

How this side says it

"Virginia measured a full decade of this and got back 48 cents on the dollar. That is not my estimate, that is the state's own audit. And the industry's own best jobs study says the subsidy flows hardest to the buildings that create the fewest jobs."

What the other side says back

They will not claim thousands of jobs, the good ones know better. Expect four moves. First, the honest numbers: the Brookings study found computing employment up 56 percent, roughly 100 to 200 permanent jobs per county, and home values up 2 to 5 percent.6 Second, the pivot to construction: electricians' union hours reached 28 million a year with at least half from data centers, and it is unions saying so, not companies.7 Third, your own 48-cent figure turned around: the same Virginia audit found that 48 cents compares with 41 cents for the average Virginia incentive, and the auditors themselves rate the exemption moderate and similar to that average, and a 2026 state incentive evaluation ranked the data center break Virginia's second-best incentive, producing $6.10 of labor income for every exempted dollar.22 28 Fourth, the multiplier: "the industry supports 4.7 million American jobs, because every direct job supports 6.5 more in the supply chain."31 Fifth, their smartest move: they may simply agree to cap or kill the tax breaks, splitting the subsidy question away from the industry entirely.

The reply

On "best of a bad bunch": being the least money-losing giveaway is still losing money, and 48 cents back is a failing grade on any test. On the $6.10 of labor income: ask what share of that is construction-phase wages that end when the building is finished, because that number counts the wave, not what stays. On the 4.7 million supported jobs: that study was paid for by the industry's lobbying group and produced by economic modeling, not measurement. The study that measured, county by county, found 100 to 200 jobs and flat wages.6 Say: "One of these numbers came out of a model the industry paid for. The other one came out of counting. I will take the counting." If they offer to give up the tax breaks, take the win and make it stick: "Say that on the record. Cap the exemptions." That concession is most of what your side is actually asking for.

The trap question inside this exchange, be ready for it

"If the deal loses money, why do more than thirty states keep making it? They must see an upside you are not counting." Do not dodge it, because there is a real answer, and it has three parts. One, the 48 cents counts only the state treasury, and the people signing these deals are buying things the treasury number does not count: the ribbon cutting, the announcement with a tech giant's logo on it, construction jobs they can point to, and property tax that flows to their counties even when the state's own books lose. The benefit is visible and concentrated. The cost is invisible, forgone revenue never shows up as a line item, and fourteen states do not even publish the total.20 Two, most states never actually decided to spend this much. The exemptions were written years ago, when data centers were small, with no caps and no annual vote. Then AI arrived and the cost exploded on autopilot: Ohio nearly tripled in a single year, from $555 million to $1.6 billion, without any legislature voting to spend $1.6 billion.20 Three, it is a bidding war. Each state matches its neighbors because losing the building feels worse than losing the money, even when the math says otherwise. That is not evidence the deal is good. It is the reason the fix has to be caps and disclosure rather than trusting each state to stop on its own. Say: "They see an upside, sure: an announcement. The audit that came later is the part they were not counting."

On the home values line, size it honestly

the Brookings 2 to 5 percent is the extra rise, above what near-identical counties saw, spread across the first decade of operation.6 On a typical house that is a couple of thousand dollars a year of added price, real but modest. Use it as a footnote to the sharper finding, which is that wages did not rise at all: the building made housing a little more expensive without making paychecks any bigger. Do not lead with it, and do not let it get inflated into "data centers priced out the neighborhood," because the study will not support that.

What has changed since the handbook

  • October 12, 2026 Helps the case for Fresh Indiana numbers report about 18,000 trades union members on data center projects in 2026 and 12,000 new union members added in two years, with Google expecting roughly 5,000 construction workers daily across three sites and Meta's Lebanon project 4,000 at peak. These are construction jobs that last for the project, not permanent operations jobs.
  • October 11, 2026 Helps the case against A Senate Democrats report based on responses from Amazon, Google, Meta, Microsoft, CoreWeave, Digital Realty and Equinix says the companies plan to keep pursuing state sales tax exemptions and NDAs. The report sharpens the case that state exemptions are a cost to public budgets, though it is a partisan Senate product rather than a bipartisan or GAO finding.
  • October 10, 2026 Helps the case against Scioto Analysis estimates Ohio's current data center policy will yield only about $2.4 billion in net benefits from 2027 to 2034, a benefit-cost ratio of 1.04, with electricity costs and air pollution driving most of the downside. The report finds a construction ban or a 90 percent clean-power requirement would score roughly 2.7 to 2.9.
  • October 8, 2026 Helps the case against A new Illinois Economic Policy Institute report puts firm modeled numbers behind a longstanding skeptics' claim. It projects $57 billion in planned Illinois data center investment through 2035 would create about 121,000 total jobs across the economy but only 2,800 direct permanent positions, a ratio of roughly 43 modeled jobs to one permanent one.
  • September 30, 2026 Helps the case for Industry and labor rolled out coordinated jobs messaging: Blackstone, OpenAI, QTS and SoftBank launched a coalition with five building-trades unions targeting seven states, and Google's data center chief used a Fortune interview to point at hundreds of thousands of unfilled trades roles. The argument becomes harder to characterize as anti-worker where the IBEW and pipefitters are publicly on the developers' side.
  • September 28, 2026 Both sides The confirmed Alabama Delta Forge 2 announcement offers a fresh order of magnitude data point: about 1,000 peak construction jobs and about 100 permanent onsite jobs for a $3.2 billion facility, with the developer's CEO publicly acknowledging that hyperscale data centers do not create large permanent workforces relative to their power use. That both supports the construction jobs case and reinforces the critique on permanent job counts.
  • September 27, 2026 Helps the case against Maryland's governor called on the General Assembly to repeal the state's 2020 data center sales and use tax exemption while codifying a uniform state review for large data centers, adding a fourth state after Ohio, Utah, and Nebraska to publicly move against its own data center tax break within one year.
  • September 20, 2026 Helps the case against Nevada's Executive Order 2026-005 removes Local School Support Tax abatements from the state data center incentive program and caps other sales and use tax breaks going to the general fund. It follows Arizona, Illinois and Washington actions earlier this year and adds another large state to the list rolling back existing incentives.
Latest evidence tagged to this argument

Tagged automatically by the daily update from the updates feed. New evidence does not rewrite the argument above; read both and judge whether the point still stands.

Sources cited in this point: 6, 7, 19funder unknown, 20, 21, 22, 28, 31paid for

4 Virginia studied this honestly, and the permanent half of the verdict is bad.

Last updated October 12, 2026 by the daily research pass. Tracks: How much electricity data centers use, Jobs.

The claim in one line

The state with more data centers than anywhere on earth had its own nonpartisan investigators look. They found real benefits that mostly end when construction ends, and permanent costs that do not end.

What it means

Construction work is genuinely good work and there is a lot of it. But it stops. The power demand does not stop. So the real question is whether you are trading a temporary boom for a permanent obligation.

The evidence

Virginia's JLARC credited data centers with 74,000 jobs, $5.5 billion in wages, and $9.1 billion in economic activity per year, and stated that those benefits come mostly during initial construction. It then found their power demand will likely increase costs for all customers, including people with nothing to do with data centers, and that if demand grows unchecked, Virginia's electricity needs would double within ten years, with data centers as the main driver. To supply that, the state would have to add new solar power at twice the rate it managed in 2024, and more new wind power than every offshore site it has secured could produce.22

How big is it, really

Doubling electricity needs in ten years means building, in one decade, a second copy of the generating capacity that it took Virginia the better part of a century to build the first time. And the yardsticks in the study are not abstractions: twice the solar the state managed in its best year ever, every single year, plus more offshore wind than every site Virginia has secured put together. Not impossible. But nobody has ever done it, and if it falls short, the shortfall gets managed with higher prices, which land on everyone.

How this side says it

"Virginia's own legislature looked at this. A construction boom that ends, in exchange for a power bill that does not. That is the trade."

A caution for this side

Quote the good half out loud first. Saying "74,000 jobs, and here is the rest of the finding" makes the rest land far harder than hiding it would.

What the other side says back

"The same study you are quoting found 74,000 jobs, $5.5 billion in wages, and $9.1 billion a year. Do you accept the whole finding or only your half? And construction here is not one event. It is a decade-long rolling pipeline, which is exactly why union hours doubled over ten years instead of spiking once and collapsing."7 22

The reply

You already quoted their half first, so the trap does not close on you. Then answer the pipeline: a rolling pipeline is still temporary, it rolls until the buildout stops, and then it stops. What does not stop is the other half of the finding: electricity demand that would double the state's needs within ten years, with data centers as the main driver, and costs that land on every customer.22 Say: "You have described a wave. I am talking about what the wave leaves behind."

What has changed since the handbook

  • October 12, 2026 Helps the case for Fresh Indiana numbers report about 18,000 trades union members on data center projects in 2026 and 12,000 new union members added in two years, with Google expecting roughly 5,000 construction workers daily across three sites and Meta's Lebanon project 4,000 at peak. These are construction jobs that last for the project, not permanent operations jobs.
  • October 8, 2026 Helps the case against A new Illinois Economic Policy Institute report puts firm modeled numbers behind a longstanding skeptics' claim. It projects $57 billion in planned Illinois data center investment through 2035 would create about 121,000 total jobs across the economy but only 2,800 direct permanent positions, a ratio of roughly 43 modeled jobs to one permanent one.
  • September 30, 2026 Helps the case for Industry and labor rolled out coordinated jobs messaging: Blackstone, OpenAI, QTS and SoftBank launched a coalition with five building-trades unions targeting seven states, and Google's data center chief used a Fortune interview to point at hundreds of thousands of unfilled trades roles. The argument becomes harder to characterize as anti-worker where the IBEW and pipefitters are publicly on the developers' side.
  • September 28, 2026 Both sides The confirmed Alabama Delta Forge 2 announcement offers a fresh order of magnitude data point: about 1,000 peak construction jobs and about 100 permanent onsite jobs for a $3.2 billion facility, with the developer's CEO publicly acknowledging that hyperscale data centers do not create large permanent workforces relative to their power use. That both supports the construction jobs case and reinforces the critique on permanent job counts.
  • September 19, 2026 Helps the case against A Pennsylvania PUC-commissioned Synapse study finds that data center load could nearly double state electricity demand between 2025 and 2040 and pushes the PJM grid below its own reliability planning standard in every year from 2027 through 2030 in the base case.
  • September 18, 2026 Helps the case against Moody's on Sept 14 estimated that meeting U.S. data center demand through 2030 will require about 45 GW of new generation costing $110 billion, more than 30 GW of it from natural gas. The number is a rating-agency analyst estimate, not a utility integrated resource plan.
Latest evidence tagged to this argument

Tagged automatically by the daily update from the updates feed. New evidence does not rewrite the argument above; read both and judge whether the point still stands.

Sources cited in this point: 7, 22

5 The water use is real and largely hidden, and the pollution lands on the same neighborhoods it always does.

Last updated October 11, 2026 by the daily research pass. Tracks: Air quality and on-site power, Water use.

The claim in one line

Many data centers cool themselves by evaporating clean drinking water, like sweating. That water does not come back. And the power plants built to run them get put next to poor neighborhoods.

What it means: on water

Computers get extremely hot. A common way to cool them is to evaporate water, which carries the heat away. That water goes into the air, not back into the local river or aquifer. There is also a second, invisible amount: making the electricity uses water too, at a power plant somewhere else. Companies often report only the water used at the building, which makes the number look far smaller than the real total.

The evidence

Researchers led by Shaolei Ren at the University of California, Riverside, published in a leading computer science journal that training one AI model in Microsoft's U.S. facilities could evaporate on the order of 700,000 liters of clean freshwater on site, and consume about 5.4 million liters once the water used to generate its electricity is counted, and projected global AI water use at 4.2 to 6.6 billion cubic meters by 2027, more than several European countries withdraw in a year.23 Ren has separately argued that company disclosures mislead precisely by counting only the water at the building.

How big is it, really

Say which number you mean, because they differ by roughly eight times and the smaller one travels further. 700,000 liters is about 185,000 gallons, roughly what a typical American family runs through in a year and a half, for every faucet, shower, load of laundry, and sprinkler, evaporated to train one model one time. Be honest about the other end of the scale too: any single AI question costs a sip, and your opponent will say so. The argument was never about one glass. It is that the industry is pouring billions of glasses, in specific places, some of which were short on water before it arrived.

What it means: on air

Some operators are not waiting for the grid. They install their own gas-burning turbines on site, because it is faster. Turbines emit pollution that causes smog and breathing problems. Where they get placed is the issue.

The evidence

In South Memphis, the company xAI put gas 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. The turbines emit nitrogen oxides and formaldehyde. The Southern Environmental Law Center appealed the air permit on behalf of the local NAACP chapter and a group called Young, Gifted and Green. Hundreds of residents attended the public hearing, and every single person who spoke opposed it. The permit was issued anyway. The NAACP later sued over illegal pollution and asked a court for emergency action, and reporting in 2026 documented 59 turbines running without permits at a second site nearby.24

How this side says it

"Hundreds of people showed up. Not one spoke in favor. The permit was issued anyway, in a neighborhood that already fails the federal air standard. That is not a coincidence. That is a pattern."

What the other side says back

On water, the direction of travel: "Microsoft cut data center water use roughly 90 percent from its earliest buildings, put back more freshwater than it took in 2025."14 If they add that the newest designs use no water at all for cooling, or that Google has committed to the same by 2030, ask where that is published: it is widely repeated but not in the source usually given for it. On Memphis: "That is one company at one site behaving badly. It is exactly what siting standards are for." A shrewd opponent will even offer to accept siting restrictions in water-stressed regions, to look reasonable.

The reply

On water, name whose numbers those are: the company's own, self-reported, unaudited. The peer-reviewed research exists precisely because company disclosures count only the water used at the building and leave out the water used making the electricity, which makes the total look far smaller than it is.23 And the newest designs do not retrofit the hundreds of buildings already running. On Memphis, accept their "siting standards" answer and then spring it: standards existed, hundreds of residents spoke, not one in favor, and the permit was issued anyway.24 Ask: "The standard was on paper. It did not bind. Tell me exactly what would have stopped this, and I will ask why the industry is not demanding it."

What has changed since the handbook

  • October 7, 2026 Helps the case against Virginia DEQ documented seven sediment runoff incidents at a Google Chesterfield data center construction site between December 2025 and August 2026, affecting roughly 7,735 linear feet of stream channel and about 1.4 acres of wetlands, with some incidents not reported within required windows. The notice adds a construction phase example to the usual focus on operating water use.
  • October 4, 2026 Helps the case against New Jersey's $1.07 million administrative order against the DataOne Vineland facility for running 62 unpermitted gas generators for roughly 46,000 hours gives opponents a concrete regulator finding, not just a modeled estimate, that on site data center generators can create significant NOx emissions without timely oversight.
  • September 25, 2026 Helps the case against The Texas attorney general opened an investigation into whether hundreds of data centers filed the water use reports state law requires, giving weight to critics who say water disclosure has been largely voluntary in practice.
  • September 24, 2026 Helps the case against New Jersey issued its largest ever data center enforcement action on September 22, fining the Vineland Data One project 1.07 million dollars for operating 62 large natural gas generators without Air Pollution Control Act permits.
  • September 17, 2026 Helps the case against Virginia DEQ's $2.5 million penalty against Microsoft over generator-runtime emissions in Leesburg is the largest data center air-quality penalty in state history, and Microsoft's settlement funds a new mobile monitoring station and 45 sensors in high-density data center areas.
Latest evidence tagged to this argument

Tagged automatically by the daily update from the updates feed. New evidence does not rewrite the argument above; read both and judge whether the point still stands.

Sources cited in this point: 14own account, 23funder unknown, 24own account

6 Communities across the country are saying no, and the industry's own tracker proves it.

Last updated October 13, 2026 by the daily research pass. Tracks: Local opposition and consent, Siting, zoning and standards.

The claim in one line

Opposition is not a few cranks. It is enormous, it is growing fast, it crosses both parties, and the people counting it work for the AI industry.

What it means

This is an argument about consent. The benefits get argued at the state and national level. The costs, meaning the noise, the traffic, the water, the power bill, get experienced by the people who live next to it, and they generally were not asked.

The evidence

A tracker called Data Center Watch counts blocked and delayed projects. It is run by a firm whose clients are AI companies, so it has every reason to make opposition look small. These are their numbers. $64 billion in projects blocked or delayed through 2025. Then in the first three months of 2026 alone, about 75 projects worth roughly $130 billion, matching the entire previous year in a single quarter. They describe it as a structural change rather than a blip: communities have learned from each other, opposition groups more than doubled across 49 states, and fourteen state legislatures took up pause proposals in three months.25 In March 2026, Senator Bernie Sanders and Representative Alexandria Ocasio-Cortez introduced a bill for a national pause.26

How big is it, really

A large data center campus runs one to several billion dollars, so $130 billion in one quarter is on the order of fifty to a hundred large projects turned away or stalled in three months, by zoning boards and county meetings, which are the smallest and slowest levers in American government. Be precise about what the dollars measure, though: that is the price tag of investment that did not land, which the pro side counts as the communities' loss, not their savings. So use the number for what it honestly shows: the size and speed of the revolt, not money anyone banked.

And know the denominator, because they will

In the very same record quarter for blocking, the industry also set its all-time record for building: $44.7 billion of actual construction spending in Q1 2026 alone, with full-year investment forecast near $700 billion.29 So do not claim communities are winning, they are not, most projects still go through. The claim that survives is different and still strong: market analysts now estimate 30 to 50 percent of the 2026 pipeline is on track to be delayed or canceled,29 which means local consent has become one of the largest business risks this industry faces. You are not describing a movement that is stopping the buildout. You are describing a movement big enough that the buildout's own analysts now price it in.

How this side says it

"This tracker is funded by a firm that works for AI companies. They have every reason to undercount. And even their numbers say $130 billion of projects got blocked or stalled in three months."

What the other side says back

Full concession, then a reframe: "You are right that opposition is real. Now look at what people actually object to: power bills, water, noise, traffic, property values. Every one of those is a siting and standards problem with a known fix. Opposition to a badly placed project is not opposition to the industry."

The reply

Scale breaks that story. One badly placed project is a siting problem. $130 billion stalled in a single quarter, opposition groups more than doubling across 49 states, and fourteen legislatures moving pause bills in three months is the same set of complaints showing up everywhere the buildings show up.25 If they answer with the denominator, "more got built than blocked, spending set a record that same quarter," concede it straight, because it is true, and then land on the number that survives: analysts put 30 to 50 percent of this year's pipeline on track for delay or cancellation.29 A risk that size is not a few cranks at a zoning meeting, it is a line item. Then take their own fix and hand it back as a demand: "If the fixes are known, why is the industry not writing them into its contracts before anyone forces it? Make them binding. The day they are binding, most of this opposition, including mine, goes quiet."

What has changed since the handbook

  • October 13, 2026 Helps the case against Pew's October 8 release puts a national number on political risk for pro data center officials: 61 percent of US voters dislike local officials who support new data center sites, with majorities in both parties. That strengthens the opposition side's argument that public consent is actively eroding, not just locally.
  • October 11, 2026 Helps the case against In the same week, developers advanced a $66 billion plan in Wyoming's Fremont County, which has no zoning code, and a $15 billion plan in Hudspeth County, Texas, where the county judge said she had very little advance information. The two cases strengthen the argument that the biggest projects are migrating to jurisdictions with the weakest local review.
  • October 9, 2026 Helps the case against A new Heatmap Pro tracker puts canceled data center investment at about $260 billion year to date, with the Q3 cancellations larger than all of 2025. The number comes from a proprietary database, not a public filing, but is the clearest indication yet that local resistance is reshaping the buildout.
  • October 5, 2026 Helps the case against A tribal government in Oklahoma voted a three-year moratorium on data center construction, adding a tribal sovereign to the list of local jurisdictions that have paused or banned data centers in 2026.
  • October 2, 2026 Both sides Two local votes on October 1 showed the split in how counties are handling moratoriums. Forsyth County, North Carolina rejected a pause after its attorney warned a Chatham County developer had successfully challenged a similar moratorium, while Mercer County, Kentucky unanimously killed draft zoning that would have allowed hyperscale sites near a historic landmark.
  • September 30, 2026 Helps the case against Local restrictions kept accumulating: Milwaukee's plan commission unanimously advanced a one-year moratorium, Buffalo's council adopted a resolution for a two-year pause, and about 40 Aurora residents spoke against a 500 MW Vantage proposal. The pattern shows opposition is broadening into large blue-city and suburban councils, not just rural counties.
  • September 30, 2026 Both sides Saginaw's new ordinance, adopted 8-0 and effective October 9, illustrates a middle-path approach: data centers are allowed but only in industrial districts, at least 500 feet from homes, with a 55-decibel noise cap, mandatory water and energy disclosures and two neighborhood meetings before approval.
  • September 29, 2026 Helps the case against A new Washington Post satellite-image analysis reports that only about 40 percent of large US data centers were built on previously developed land, with about a third on former farmland, 12 percent on former forest and 11 percent on shrubland. This strengthens the case for tighter siting rules on greenfield conversion.
Latest evidence tagged to this argument

Tagged automatically by the daily update from the updates feed. New evidence does not rewrite the argument above; read both and judge whether the point still stands.

Sources cited in this point: 25funder unknown, 26own account, 29

7 They are enormous, they hum all night, and they change what a place is.

Last updated October 13, 2026 by the daily research pass. Tracks: Noise, appearance and rural character, Siting, zoning and standards.

The claim in one line

A data center is a windowless box the size of several football fields, ringed by fences and lights, with hundreds of rooftop fans running around the clock. Put one next to homes and the neighbors do not get a factory's jobs; they get a factory's noise and a factory's silhouette, permanently.

What it means

The complaints that fill county meetings are rarely about terawatt-hours. They are about a hum that comes through closed windows, a skyline of gray walls where a farm used to be, and the sense that a town made a decision nobody who lives there was asked about. Those are real harms even though nobody can put them on a utility bill.

The evidence: part one, the noise

In the Great Oak neighborhood south of Manassas, Virginia, an Amazon data center went up next to homes. Residents measured up to 65 decibels at night and described it as "like being on a tarmac with an airplane engine running constantly," except that the plane never leaves. One neighbor replaced every window in his house and it did not help.87 After two years of pressure, Amazon replaced all 424 rooftop exhaust tubes with taller ones, cutting the noise by about 10 decibels, roughly half as loud, to around 50 decibels.88 Then the county wrote a new noise ordinance, 52 decibels in the daytime and 47 at night, about the level of moderate rainfall. Testing showed the Amazon complex still would not comply, that the low-frequency roar residents feel inside their houses is not regulated by the old rule or the new one, and that emergency diesel generators are not covered by either.89 Next door in Loudoun County, the county's own website says noise complaints "are among the most persistent issues brought forward to Loudoun County officials," admits it has no rule at all for tonal or low-frequency noise, and does not expect one until 2027.90 This is not new. Chandler, Arizona residents began complaining about a million-square-foot data center in 2014 and the city did not adopt an ordinance until 2023.91

The evidence: part two, the scale and the setting

The Prince William Digital Gateway was 37 data center buildings, "roughly the equivalent of 144 Walmart Supercenters," with 14 electrical substations, planned on the edge of Manassas National Battlefield Park. County planning staff recommended denial twice, in part because without drawings of the actual buildings nobody could trust the studies of what they would look like from the battlefield. The Virginia Court of Appeals voided the rezoning in March 2026 and the developer dropped its appeal that July.92 In North Carolina, the Forsyth County planning board recommended denying a hyperscale project in Rural Hall in June 2026 after residents argued it would end the area's rural character, and Stokes County residents sued over a zoning change that opened a dozen rural sites to data centers.93

How big is it, really

A decibel scale is logarithmic, so the numbers understate it. The 65 decibels measured at Great Oak is a vacuum cleaner in the next room, all night, every night. Even after Amazon's fix, 50 decibels is a refrigerator that never cycles off. And the buildings do not go away: a data center is a thirty to fifty year land use, so the question is not whether the hum is tolerable this year but whether the next owner of every house nearby gets asked.

How this side says it

"You are describing megawatts. The people who live there are describing a hum they cannot turn off and a wall where their view used to be. Both are real. Only one of them is in your spreadsheet."

Why this lands

It is the argument that actually wins zoning votes. Power bills and water get argued at the state capitol; noise, lights and character get argued at the microphone in the county boardroom, by people who can point out the window.

What the other side says back

"Every one of those complaints has an engineering answer, and the answers are being written into law right now. Amazon cut the noise in half once it was made to. Chandler's ordinance has required a sound study, mitigation and generator testing windows since 2023. Loudoun ended by-right data centers in March 2025 and is writing design and compatibility standards. Prince William is adopting a rainfall-level noise limit. And the one measured study of home values found they went up 2 to 5 percent near data centers, not down."88 91 94 6

The reply

Concede the fixes, then point at the dates. Great Oak's neighbors waited two years for a fix that still fails the county's own new standard, Loudoun says it will not have a low-frequency rule until 2027, and Chandler took nine years.89 90 91 Every fix arrived after the building did, and the building does not move. Say: "You are right that the standards exist. Now tell me why the industry keeps building before they do." On home values, do not fight the number; narrow it. A county-wide average rising 2 to 5 percent says nothing about the twenty houses on the fence line, which is where the noise lives.6

What has changed since the handbook

  • October 11, 2026 Helps the case against In the same week, developers advanced a $66 billion plan in Wyoming's Fremont County, which has no zoning code, and a $15 billion plan in Hudspeth County, Texas, where the county judge said she had very little advance information. The two cases strengthen the argument that the biggest projects are migrating to jurisdictions with the weakest local review.
  • October 2, 2026 Both sides Two local votes on October 1 showed the split in how counties are handling moratoriums. Forsyth County, North Carolina rejected a pause after its attorney warned a Chatham County developer had successfully challenged a similar moratorium, while Mercer County, Kentucky unanimously killed draft zoning that would have allowed hyperscale sites near a historic landmark.
  • September 30, 2026 Both sides Saginaw's new ordinance, adopted 8-0 and effective October 9, illustrates a middle-path approach: data centers are allowed but only in industrial districts, at least 500 feet from homes, with a 55-decibel noise cap, mandatory water and energy disclosures and two neighborhood meetings before approval.
  • September 29, 2026 Helps the case against A new Washington Post satellite-image analysis reports that only about 40 percent of large US data centers were built on previously developed land, with about a third on former farmland, 12 percent on former forest and 11 percent on shrubland. This strengthens the case for tighter siting rules on greenfield conversion.
  • September 27, 2026 Helps the case against The Maryland executive order requires any data center at 25 megawatts or larger that seeks any state action to be scored against five public criteria and posted on a monthly Data Center Dashboard, setting a floor for state review that mirrors calls in Virginia and California to end approvals with no state visibility.
  • September 24, 2026 Helps the case against Prince William County, one of the largest existing data center markets in the country, voted 8 to 0 on September 23 to end by-right data center approvals and shrink its overlay district from about 9,700 to roughly 3,641 acres. Future projects outside the reduced overlay now require a special use permit.
  • September 23, 2026 Helps the case against Texas broadened its data center pause to include all TCEQ environmental permits, meaning water, air and land use approvals now sit on hold statewide until ERCOT and Texas Water Development Board audits are complete. That extends the moratorium beyond grid interconnection into the full permitting stack.
  • September 19, 2026 Helps the case against Virginia's executive order proposes ending by-right approval and requiring local government sign-off for any data center using more than 25 megawatts, and directs an accelerated noise rule and water-use standards. Loudoun County advanced a 12-month pause on new data center and substation applications on September 16.
Latest evidence tagged to this argument

Tagged automatically by the daily update from the updates feed. New evidence does not rewrite the argument above; read both and judge whether the point still stands.

Sources cited in this point: 6, 87, 88, 89, 90own account, 91own account, 92own account, 93own account, 94

8 The largest one ever built is a live experiment in who holds the bag, and so far the answer is the people who live there.

Last updated September 24, 2026 (handbook revision; no newer evidence yet).

The claim in one line

Louisiana is building Meta's biggest data center anywhere. To power it, regulators approved billions of dollars of new gas plants. Those plants are built to run 30 to 40 years. Meta's contracts run 15 to 20. Somebody pays for the leftover years, and it is not Meta.

What it means

Picture a town borrowing for a forty year bridge because one factory promised to use it for twenty. If the factory leaves on schedule, the bridge still has twenty years of payments on it, and the townspeople make them. That is the shape of this deal, except the bridge is a fleet of gas fired power plants and the town is everyone on the utility's bill.

The evidence

In Richland Parish, Louisiana, Meta is building a campus it calls Hyperion, announced at $10 billion in December 2024 and raised to a commitment of more than $50 billion by July 2026, with 5 gigawatts of computing power planned.60 In August 2025 the state's utility commission approved about $3.2 billion of new gas generation to serve it, three turbines totaling roughly 2,200 megawatts, on a 4 to 1 vote.76 In April 2026 the same commission fast tracked a roughly $21 billion plan for seven more plants, again 4 to 1, with the full decision expected in December 2026.65 Intervenors working the docket calculate that when Meta's contract ends, about $3.4 billion of plant cost could still be unpaid, and that ratepayers carry it.63

How big is it, really

Big enough that the commission's own judge went looking for the arithmetic and was overruled. In 2026 the commission's chief administrative law judge ordered Meta to disclose how it calculated the job numbers and the electricity demand it had been publicizing, and Meta moved to quash the order as unreasonable and oppressive.65 In August 2026 the commissioners voted 3 to 1 to overrule their own judge and let the figures stay private.76 Separately, a nonprofit newsroom documented that a state senator who sponsored bills enabling the project bought seven properties within five miles of the site after it was announced, and sold about 300 acres to the utility roughly four weeks after the approval vote.66 None of that proves the numbers are wrong. It means nobody outside the company has been allowed to check them.

How this side says it

"The power plants outlive the contract by twenty years. Tell me who pays for those twenty years, and show me where it says so."

Why this lands

It is specific, current, and entirely on the public record: real dockets, real votes, named dissenters. You are not predicting a disaster. You are pointing at a mismatch between two numbers, a plant's lifetime and a contract's length, that nobody disputes.

What the other side says back

"This is the poorest corner of a poor state and the money is real. Nearly one person in four in that parish lives below the poverty line.67 The project has already put more than $1.6 billion of contracts into Louisiana businesses, and the state counts 7,500 construction workers at peak and about a thousand permanent jobs.60 The utility says its agreements with Meta will hand customers about $2.65 billion in savings.62 Meta has funded bill assistance and teacher bonuses on top of that.61 You are asking people with nothing to turn down the only offer they have had in decades."

The reply

Every one of those numbers comes from somebody selling the deal: the company, the utility, or the state agency that recruited the project.60 61 62 The savings figure is the utility's own and has not been independently verified. And when an independent judge asked to see the arithmetic behind the jobs figure, the commission voted to keep it sealed.76 Say: "I am not against the money. I am against the part where the only people allowed to see the math are the people who profit from it." Then make the ask that costs them nothing if they are right: publish the demand and jobs calculations, and put the leftover plant cost on the company rather than the parish.

Sources cited in this point: 60own account, 61own account, 62own account, 63own account, 65, 66, 67, 76

Answering the other side's opening points

The counters above cover what they say when you attack. This section is for when they are on offense with their own affirmative points. Each of their six points in Part 4 ends with a "what they will say back" block. That block is you. Those are your lines. Read Part 4 the same way they will read Part 3, and you have both halves of every exchange. Two of their openers deserve special drilling because they come up in nearly every debate:

Their Loudoun County opener (taxes down ten years running): your answer is under their Point 2. The short version: do not fight the numbers, fight the concentration. 39 percent of a county budget on one industry is a household living on one paycheck.

Their "efficiency fixed this before" opener: your answer is under their Point 6. The short version: that research covers 2010 to 2018, the one-time savings of consolidating server closets are spent, and the same authors now publish the projections showing steep growth.1 13 You can only empty the closets once.

Questions this side asks the other

  1. "Your rate study was paid for, and reviewed before publication, by the industry's lobbying group. Is there an independent study that reached the same conclusion?"8
  2. "Virginia's protections start in 2027. How much got connected before that date, and under what terms?"9
  3. "You say data centers will ease off during peak hours. Will you support making that a binding requirement instead of voluntary?"11
  4. "The grid's own independent watchdog says data centers drove $6.3 billion in one auction. Is the watchdog wrong?"17
  5. "If these are as profitable as you say, why do they need a tax exemption that costs Ohio $1.6 billion a year?"20
  6. "What share of a county's budget depending on one industry would you consider too risky? Is 39 percent fine?"4

Where this side is weak, in its own words

Every honest case has soft spots. These are the places this side's own handbook warns not to overclaim.

  • Do not say data centers are the main reason electricity prices are rising nationally. They are not, and you will get fact-checked. Natural gas price swings, old power plants retiring, grid upgrades, wildfire protection, and equipment costs are all major drivers.8 10 Say data centers are a large and fast-growing driver. That is true and defensible.
  • Do not dispute the Loudoun County tax numbers. They come from public budget documents anyone can pull.4 Attack the concentration risk and the lack of consent instead.
  • Do not use "$2 million per job" as a current figure. It comes from an older analysis of the very biggest deals. Use it as a sense of scale and lead with the state-by-state revenue losses instead.20
  • Your best jobs evidence is your opponent's own study. Quote it accurately, including the 56 percent rise in tech employment it found, or you will be caught cherry-picking and lose the exchange.21
  • Memphis is one company at one site. Do not claim it is what every operator does. Use it as proof of what happens when nothing constrains siting.24
  • The Virginia study also found large benefits.22 Say them out loud first. It costs you nothing and makes you credible.
  • On noise and looks, do not claim the buildings are unfixable. Amazon cut Great Oak's noise in half once made to, and Chandler's rules have worked since 2023.88 91 Argue timing and enforcement: the fixes arrive after the building does.
  • Do not argue for a national construction ban.26 It is the maximum possible ask, it is not going to pass, and defending it makes you easy to dismiss. Argue for the specific tools instead: mandatory flexibility, public contracts, caps on tax breaks, and real siting standards. Those are much harder to argue against.
Now read the case for Get the full handbook