On 22 July I wrote up Gallup's first survey of American attitudes toward data centers and flagged what looked like its weakest joint. Gallup had asked about AI data centers, not data centers. I guessed the word was doing work, and said we had no way to know how much.
We do now. It is worth about eight points.
Cremieux published a piece this morning titled Can the Truth Convince People to Like Data Centers? His answer is mostly no, and the essay is a genuinely useful tour of the NIMBY messaging literature, which is a body of work almost nobody in the AI infrastructure debate has read. But the poll he links to is the June 2026 Echelon Insights omnibus for Puck, and that topline contains two experiments he walks straight past. One of them settles the framing question. The other answers the question in his title, and points somewhere other than where he ends up.
Question 30 asked 1,012 likely voters whether they would support or oppose each of eight things being built in their community. Randomized row order, four point scale, one grid, one sample, one sitting. Two of the eight rows were data centers, identical except for what the building is for.
Support and opposition, in that order. The only difference between the first row and the third is the phrase to power artificial intelligence in place of to power digital services (e.g., online searches and video streaming). That phrase costs eight points of support and buys nine points of opposition. The nuclear plant sits between them.
The full ladder is worth seeing, because it kills the lazy reading that Americans have simply turned against buildings.
| Proposed in your community | Support | Oppose |
|---|---|---|
| An Amazon distribution center | 68% | 24% |
| A traditional manufacturing factory | 67% | 24% |
| A solar farm | 66% | 26% |
| A wind farm | 61% | 30% |
| A chip manufacturing plant | 49% | 35% |
| A data center to power digital services | 35% | 53% |
| A nuclear power plant | 34% | 57% |
| A data center to power artificial intelligence | 27% | 62% |
A warehouse polls at 68. A steel mill polls at 67. These are the same respondents on the same afternoon. Whatever is happening here is specific, and a good deal of it is attached to two words rather than to a building.
True of the AI framed row, at 27 against nuclear's 34. False of the other one, at 35. The essay's own primary source draws a distinction that the essay collapses into a single sentence.
Why it mattersThe sentence is the hook for the whole piece. If the industry's problem is partly a word, then the recommended remedy of permitting workarounds and tax abatements is aimed at the wrong organ. Nobody is packing a county hearing over online video streaming.
Before the poll re asked the build question, it tested six arguments in favor of data centers and asked how convincing each one was. This is the actual persuasion experiment, and its ranking is the most useful thing in the document.
| Argument tested | Convincing |
|---|---|
| Generate billions in local tax revenue that can benefit schools, roads, and other essential services | 42% |
| Make powerful computing accessible to anyone, anywhere | 37% |
| Create thousands of high paying jobs for local community members | 36% |
| Power modern business and healthcare without costly hardware investment | 34% |
| Use less water than steel manufacturing does | 32% |
| The scary water figure rested on a mathematical error the author has acknowledged and corrected | 27% |
The winner is public money for schools and roads. The loser, by fifteen points, is the correction. The best available fact, delivered with an admission of error attached, was the least persuasive thing tested.
Then the poll re asked the build question. Support for the digital services center moved from 35 to 40. Support for the AI center moved from 27 to 31.
Four and five points against a margin of error of 3.7, with no control arm, measured by re asking the same people the same question immediately after showing them six arguments on one side. Every pressure in that design points up.
Credit where dueCremieux flags this himself in a footnote and calls the result an upper bound. He is right, and most people quoting a within subject shift would not have said so.
The essay treats the belief battery as a list of things the public has wrong. Six items, and they do not all behave the same way when you go looking for the evidence.
The public is right, and it is not close. Virginia's JLARC found in December 2024 that a typical 250,000 square foot facility employs about 50 full time workers, roughly half of them contract staff. Statewide, just over 8,000 direct data center jobs against more than 63 million square feet of data center space. Georgia's state auditors project 1,873 permanent data center jobs in the whole state for 2025, against $474.2 million in forgone revenue. That is about $253,000 of forgone state revenue per permanent job per year.
Where the essay goes wrongListing "they create high paying jobs" among the truths people refuse to accept. The jobs are high paying. Virginia's exemption requires 150% of local prevailing wage. There are about fifty of them per building, and Brookings finds county wages unchanged, because fifty good salaries do not move a county.
Nobody knows. There is no peer reviewed hedonic study of data center proximity and home values. Every paper is from 2026 and every one is a working paper. Two of them use transaction level difference in differences and find negative effects: a national analysis finds hyperscale openings cut nearby prices 6.8%, fading past 14 km, and a Loudoun County analysis finds homes within half a mile of an announced facility sell about 2.8% below comparable homes slightly farther out. County level papers, including an NBER working paper and a Brookings analysis, find prices up.
Where the essay goes wrongCalling this "extraordinarily obviously untrue." The two designs are not in conflict. A county can gain while the two hundred houses across the road lose, and the public's belief is a claim about the second thing. On the single item where the essay uses its strongest adjective, the literature is eighteen months old, unrefereed, and split by unit of analysis.
A draw, and it depends entirely on which tax. Loudoun County collected $733 million in data center tax revenue, 31% of all its local tax revenue. That is a spectacular result and it is real. Georgia's Department of Audits found the other side of the ledger: the state's sales and use tax exemption produced a net state fiscal impact ranging from minus $17.0 million in 2018 to a projected minus $780.2 million in 2030. Virginia's exemption cost $928.6 million in FY23 alone.
The contradiction that isn't oneThe essay says people believe data centers raise taxes and lower property values, and that they do not see the contradiction. There is no contradiction to see. A topline cannot tell you whether the same respondent holds both beliefs, and even if one did, a falling assessment on your house and a rising rate on an eroded commercial base is an ordinary story, not a paradox. Note also that taxes drew 29% unsure, the highest in the battery. This is the item people are least confident about, and it is the one the essay is most confident they are wrong about.
This is the weakest ground the essay stands on, and the case against it comes from a market monitor rather than an advocacy group. PJM's 2025/2026 capacity auction cleared at $269.92 per MW day, up from $28.92, taking total capacity cost from $2.2 billion to $14.7 billion. Monitoring Analytics, PJM's own Independent Market Monitor, attributes $9.3 billion of that increase, a 174% rise, to data center load, and calls it misleading to describe the result as ordinary market tightening. PJM's own December 2025 release states that of roughly 5,250 MW of forecast peak load growth, nearly 5,100 MW is attributable to data center demand. Synapse, analyzing for the DC Office of the People's Counsel, put the retail effect at about $10 a month on an average Pepco residential bill.
The honest qualificationCutting the other way: JLARC found no historic cost shifting in Virginia rates as of December 2024, and the viral claim that bills are up 267% near data centers refers to wholesale nodal prices, not retail bills, and was rated Mostly False. Whether tomorrow's load raises your bill turns on tariff design, which is precisely why Ohio's PUCO and the Virginia SCC spent 2025 rewriting large load tariffs. Regulators do not demand $1.5 million per megawatt in collateral against a risk they consider imaginary.
Right nationally, wrong locally, and the essay quietly drops the larger of the two terms. Lawrence Berkeley National Laboratory put direct water consumption by all US data centers at 66 billion litres in 2023, about 48 million gallons a day. Against 322 billion gallons a day of national withdrawals that is a rounding error, and on that framing the essay wins outright.
The term that goes missingThe 176 TWh those facilities consumed in 2023 carried roughly 800 billion litres of water consumed at power plants, about twelve times the direct figure. Brian Potter has shown that number is inflated by counting hydroelectric reservoir evaporation, and correcting for it still leaves the indirect term the larger one. Meanwhile water is not a national commodity. Google consumed roughly 550 million gallons at The Dalles in 2025, close to 40% of that city's entire municipal use, and Bloomberg found about two thirds of data centers built since 2022 sit in water stressed regions. Calling a county level belief a misconception because the national aggregate is small is a category error, not a correction.
The essay's tour of the persuasion research is its strongest section and its most quotable one, which is why it is worth reading the papers. Most of the citations hold. Four do not hold the way they are used.
| Cited work | How the essay uses it | What it says |
|---|---|---|
| Kalla & Broockman 2018, 49 field experiments | Persuasion research comes up empty handed | Accurate, but the scope condition is general elections, where party ID anchors everything. The authors say persuasion does appear to work in primaries and ballot measures. A county siting fight is closer to the second thing. |
| Kalla & Broockman 2020, interpersonal conversation | Effects are meager | It is a positive result. Argument alone did nothing. The same conversation with non judgmental exchange of narratives durably reduced exclusionary attitudes for at least four months. Small, around d = 0.08, and durable. |
| Broockman et al. 2024, shared demographics | Persuasion shortcuts do not work | The paper's conclusion is optimistic. Persuasion travels across demographic lines, so you do not need a matched messenger. It presupposes that conversation works. |
| Coppock, Green & Porter 2022, digital ads | Effects small and they fade fast | There is no over time measurement in the paper at all, so it cannot speak to fade. The estimate is 0.04 points with a standard error of 0.85 on precinct vote share, which is a precise null rather than a small effect. |
| Gerber et al. 2011, televised ads | Small and short lived | The paper's own summary is "strong but short lived." Maximum ad volume moved standing about 6 points and it was essentially gone within a week. Large and decaying implies persuasion works and needs sustaining, which is a different argument. |
| Monkkonen & Manville 2019, LA apartments | 63% raised support with community benefits; 23% of participants were pure NIMBY | The 63% is the best of three benefit framings, which ran 63, 57 and 50. The 23% is 23% of respondents who opposed the building on their own block, not 23% of participants. |
| Doberstein, Hickey & Li 2016, Kelowna | Unmoved by expert views, moved by neighbors' views | There is one combined treatment, not two arms, and its effect was weak. The essay splits a single treatment in half and gives the halves opposite verdicts. Also n = 202, a 7.1% response rate. |
One omission is worth naming, because it is by the same author the essay leans on for the null result. Alexander Coppock's book Persuasion in Parallel argues that factual information does move issue attitudes, that it moves nearly everyone in the same direction, and that the effects are durable. His distinction is between candidate choice in a partisan general election, which is immovable, and issue attitudes, which are not. A data center is the second category.
A four hour hemodialysis treatment draws roughly 500 litres of municipal water. About a third of that becomes dialysate. The rest leaves as reverse osmosis reject and goes down the drain at drinking water quality. A metered 11 machine hospital unit in North Carolina used 1,717,000 litres across 4,200 treatments in 2023, of which 927,000 litres were reject. Per patient the figure runs near 78,000 litres a year, which is roughly one and a half times what an ordinary person uses at home over the same year. Global hemodialysis consumes something on the order of 265 million cubic metres annually.
Multiply the per patient figure by the US hemodialysis census and you land somewhere near 45 billion litres a year. American data centers consumed 66 billion litres directly in 2023. Those are the same order of magnitude, and I have never heard of a county hearing about a dialysis unit.
My field has not been quick here either. A 2019 survey of public dialysis facilities in Victoria found that 25% recycled their reject water. The rest of us poured it away.
None of which makes dialysis wasteful, and the comparison is not a defense of anybody's cooling tower. It is a point about salience. The volume of a water use turns out to predict almost nothing about how much anybody minds it. What predicts it is whether the benefit is legible to the person absorbing the cost. A dialysis unit's benefit is a neighbor who is alive on Thursday. A data center's benefit is a line item in somebody else's budget, in a currency the neighbor does not hold.
The closing recommendation is to give people what they think they want, run roughshod over them with any available permitting workaround while maintaining the appearance that locals were heard, make sure building costs them nothing, and then talk up what the essay itself calls "the implicit and actual bribery" of construction jobs, permanent jobs, and abatements.
Half of that is well supported. Liebe, Bartczak and Meyerhoff found that perceived involvement in planning was the single strongest lever they measured, larger than local ownership or regional consumption. Perception really is the mechanism.
The other half is at war with the paper cited three entries earlier. Walker, Wiersma and Bailey ran exactly this test. Community benefits raised support for an offshore wind farm from a mean of 4.59 to 5.22 on a seven point scale. Add a counterframe noting that some people see such benefits as an attempt to bribe the community and the boost vanishes, landing at 4.62, statistically indistinguishable from telling people nothing at all. And the largest opposition effect in Monkkonen and Manville, about 20 points, came from telling people the developer stood to make large profits.
So the strategy is a campaign of manufactured consultation and undisclosed payment whose failure condition is any local reporter using the word the essay used to describe it. That is not a plan. It is a wager that nobody reads the plan.
Public, shared, visible benefits. Genuine procedural involvement. Local ownership. Regional consumption. Compensation delivered as services and infrastructure rather than private payments, which García et al. found need to be far larger to buy the same acceptance.
What it does not supportSimulated consultation plus quiet money. Knauf found that even where benefits bought assent, they did not produce any perception of fair distribution. The money changes the vote without changing the grievance, which means it has to keep being paid.
The essay asks whether the truth can convince people to like data centers, treats a four point nudge as the answer, and concludes that the honest path is closed. The poll it cites ran a better experiment and reached a different place. Six arguments were tested. The winner was public money for schools and roads. The loser, by fifteen points, was the one explaining that the scary number had been a math error the author has since owned.
That is not a finding about whether people can handle facts. It is a finding about what a fact is for. The industry keeps mailing corrections in answer to a question nobody asked. The question was never whether the number is right. It was what the neighbors get.
Sources
The essay: Cremieux. "Can the Truth Convince People to Like Data Centers?" Cremieux Recueil, 19 August 2026. cremieux.xyz
The poll: Echelon Insights for Puck. June 2026 Voter Omnibus, topline. Fielded 11 to 14 June 2026, n = 1,012 likely voters, margin of error 3.7 points. Questions 30 to 43. puck.news
Prior WAiR coverage: "Seven in Ten Americans Don't Want a Data Center Next Door," 22 July 2026, on Gallup's first survey of the question. wair.ajwein.com
Water, national: Shehabi, A. et al. 2024 United States Data Center Energy Usage Report. LBNL-2001637, December 2024. lbl.gov
Water, correction: Potter, B. "I Was Wrong About Data Center Water Consumption." Construction Physics, 30 August 2025. construction-physics.com
Capacity market: Monitoring Analytics, LLC. 2025 State of the Market Report for PJM and October 2025 capacity analysis. monitoringanalytics.com. PJM, "PJM Auction Procures 134,479 MW of Generation Resources," 17 December 2025. pjm.com
Retail bills: Synapse Energy Economics for the DC Office of the People's Counsel. Drivers of PJM's Capacity Market Price Surge and Its Impacts, May 2025. opc-dc.gov. PolitiFact, "How much have data centers increased electricity prices?" 12 June 2026. politifact.com
Virginia: JLARC. Data Centers in Virginia, Report 598, 9 December 2024. jlarc.virginia.gov
Georgia: Georgia Department of Audits and Accounts. Tax Incentive Evaluation: Data Center Sales and Use Tax Exemption, December 2025. audits.ga.gov
Subsidies: Good Jobs First. "Data Center Subsidies Surge as States Lose Billions," 12 June 2026. goodjobsfirst.org
Property values, negative: Chia, L.E., Hu, S., Wang, Q. & Fan, M. The Spatial Incidence of Hyperscale Data Centers. SSRN 6500238, April 2026. ssrn.com. Rubinovitz, M. Here Comes the Cloud. SSRN 6136790, February 2026. ssrn.com
Property values, positive: Alvarez, F. et al. Data Centers and Local Economies in the Age of AI. NBER Working Paper 35194, May 2026. nber.org. Brookings, "New evidence on data center employment effects," 2026. brookings.edu
Persuasion: Kalla & Broockman, APSR 112(1), 2018; Kalla & Broockman, APSR 114(2), 2020; Broockman et al., BJPS 54(4), 2024; Coppock, Green & Porter, Research & Politics 9(1), 2022; Gerber, Gimpel, Green & Shaw, APSR 105(1), 2011; Coppock, Persuasion in Parallel, University of Chicago Press, 2022.
NIMBY messaging: Monkkonen & Manville, Journal of Urban Affairs 41(8), 2019; Walker, Wiersma & Bailey, Energy Research & Social Science 3, 2014; Liebe, Bartczak & Meyerhoff, Energy Policy 107, 2017; Vuichard, Stauch & Dällenbach, ERSS 58, 2019; Knauf, Energy Policy 165, 2022; García, Cherry, Kallbekken & Torvanger, Energy Policy 99, 2016; Muñoz & Tormos, ERSS 127, 2025; Doberstein, Hickey & Li, Land Use Policy 54, 2016.
Dialysis water: Haddad, S., Kittner, N. & Flythe, J.E. "Thinking Globally, Acting Locally: Water Use in a Hospital Hemodialysis Unit." Kidney360 5(11), 2024. doi.org. Ben Hmida, M. et al. "Water implications in dialysis therapy." Kidney International 104(1), 2023. doi.org. Moura-Neto, J.A., Barraclough, K. & Agar, J.W.M. J Bras Nefrol 41(4), 2019. doi.org. Barraclough, K.A. et al. "Green dialysis survey." Nephrology 24(1), 2019. doi.org
Context: Silver, N. & Sun, J. "Why does everyone hate data centers?" 17 August 2026. natesilver.net