Both sides of the data center argument repeat numbers that do not survive a look at the source. We checked the ones that travel furthest. Some are badly overstated. Some are true and worse than people think. We take no position on whether any project should be approved.
Checked September 23, 2026. Everything below is either traced to the source named beside it or marked as unresolved. Where a figure only appears in one analyst's blog or in vendor marketing, we left it out rather than repeat it. Our working notes, including the figures we refused to publish and why, are kept with the research.
OVERSTATED
“Every AI prompt uses a bottle of water”
This traces to real research. Shaolei Ren at UC Riverside published an estimate in April 2023, and in September 2024 the Washington Post, working with Ren, restated it as about 519 milliliters to write a 100 word email. That is the version that went around the world. Ren's group now puts a GPT-4 era prompt at roughly 15 milliliters in total, and only about 5 milliliters of that is used on site at the data center. That is around a thirtieth of the viral figure. The first number came out of real work. It modeled an older, less efficient system on a worst case electricity grid (coal heavy, 2023), and the caveats fell off in the retelling.
A bottle, or a teaspoon Water per AI prompt, as first estimated and as the same researchers now put it.
What to ask: If someone quotes you the bottle, they are quoting a 2023 estimate of a 2023 system. Ask for the current one.
“Closed loop cooling means the site uses no water”
Closed loop is a real and meaningful difference. An evaporative system deliberately boils water off into the air to carry heat away, and that water is gone. A closed loop circulates a sealed fluid and sheds heat into the air on the same principle as a car radiator, so it uses far less water and more electricity. Some water is still consumed: the loop has to be filled and topped up. The national laboratory that studies this describes a real trade-off, less water in exchange for more electricity.
What to ask: The way to check a closed loop claim is the site's actual water withdrawal permit, not the brochure. Ask for the permit.
This is where the skeptics are wrong to wave it away. In March 2026 Amazon Data Services agreed to pay $20.5 million to settle a class action over nitrate contamination in the Lower Umatilla Basin in Oregon, an area of 562 square miles with about 45,000 residents and more than 600 domestic wells carrying nitrate at levels unsafe to drink. The claim was that water passing through the data centers came out more concentrated and went into disposal systems that had exceeded their permits. Amazon denies the allegations and settled without admitting them. Most of the documented harm in cases like this comes from construction and wastewater disposal rather than from day to day cooling.
What to ask: Routine operation and construction are different risks. Ask about stormwater, wastewater disposal and well monitoring, not just cooling.
The construction workforce is large and temporary. The permanent one is small. These buildings are heavily automated, and the gap between the two numbers is the most consistent pattern in the reporting. ProPublica found that Microsoft's entire data center footprint in Washington State employed 417 people as of 2023, after years of tax breaks written to encourage exactly that employment.
The two job numbers are not the same number Ask a developer which one they are quoting you. Permanent headcount is the one that stays.
What to ask: When a developer gives you a jobs number, ask which one it is: construction, or permanent after the builders leave.
SOMETIMES TRUE, AND THE BEST EXAMPLE COMES WITH ITS OWN WARNING
“The tax revenue makes it worth it”
Loudoun County in Virginia has the most mature data center cluster in the country, and the county says on its own website that the revenue has let it lower the residential property tax rate almost every year for a decade, giving it the lowest rate in northern Virginia. That is a real benefit, stated by the government that collects it. The same county page also warns that it cannot become overly dependent on a single revenue source that may change. Both halves come from the same official source, and anyone quoting one half should quote the other.
What to ask: Revenue can be real. Ask whether the projection in front of you rests on an assessed value the county has actually agreed to.
“The assessed value in the projection is the value the town will get”
Data center equipment loses value fast, and owners routinely go back and argue the assessment down after the building is up. This is common enough that the Cook County Assessor's Office in Illinois has publicly described defending its data center assessments at appeal hearings. We could not find a national count of how often this happens or by how much, so we are not going to claim a pattern beyond the jurisdictions where it is documented.
What to ask: Ask whether the tax projection is based on a value the assessor has agreed to, and what happens to the town's budget if it is appealed down.
Good Jobs First, which tracks subsidies and argues for limiting them, examined eleven large data center deals and calculated an average of about $1.95 million in public subsidy for each permanent job created. That is their calculation from public disclosures rather than a government audit, and they hold a clear policy position, so we label it as theirs. We have seen no comparably detailed calculation from the other direction.
What to ask: Subsidy per permanent job is the honest way to compare deals. Ask for it in writing.
“If it meets the noise limit you will not hear it”
This is the most useful thing on this page for anyone who lives nearby. Data center noise is concentrated at low frequencies, roughly 100 to 500 hertz, and the A weighted meters written into most municipal noise ordinances are poor at capturing that range. So a facility can genuinely pass a compliance reading while neighbors genuinely hear a constant hum. Both can be telling the truth, because the ordinance measures the wrong thing.
Why it passes the meter and still wakes you Most municipal noise limits are written in A weighted decibels, which are modeled on how people hear speech, not low frequency hum.
What to ask: Ask that the noise condition be written to cover low frequency sound specifically, and that measurements be taken at night, not just as an A weighted average.
The evidence does not currently support a confident answer in either direction. A 2026 survey of real estate agents by the National Association of Realtors found 25 percent reported a positive effect on nearby home values and 22 percent reported a negative one, and the association itself cautions against reading that as proof of anything. Academic work so far is short window, because this building wave is recent. The uncertainty cuts against a developer promising no impact and against a neighbor promising ruin, in equal measure.
What to ask: Anyone who tells you confidently what this does to your home's value is ahead of the evidence.
“Data centers are using a huge share of the country's electricity”
The national laboratory report commissioned for this question found data centers used 176 terawatt hours in 2023, which was 4.4 percent of all electricity in the United States, and projected somewhere between 6.7 and 12 percent by 2028. That range is wide on purpose; the people best placed to forecast it are saying plainly that they cannot yet narrow it.
Share of all US electricity used by data centers The 2028 figure is a range on purpose. Anyone quoting one number picked an end of it.
What to ask: Treat any single confident 2028 number as a choice someone made inside a range, and ask why they chose that end of it.
A resident said this at the September 8 public meeting and it has travelled since. The figure we can source is different: Parkland School District reports the current plan keeps 356 backup generators, upgraded at the district's request from Tier 2 to Tier 4, the strictest federal tier for non-road engines. We could not find any document supporting 2,000, and we could not find one that definitively settles the count either. We are including this one because a page that only corrects the developer is not a neutral page. The number on the opposition's side has to survive the same check.
What to ask: Ask for the generator count on the plan sheet, the tier, and the testing schedule, all in writing. That ends the argument.
What data centers add to your specific monthly electricity bill. Credible estimates differ by several times over, and nobody has a settled way to separate the data center share from gas prices, plant retirements and grid spending that would have happened anyway.
Whether the new state rules making large customers pay for their own grid upgrades actually hold. Pennsylvania's is four months old and has not been tested by a cancelled project.
Long run effects on nearby property values, because this building wave is too recent for anyone to have measured a full cycle.
Whether the health complaints residents report near operating facilities, mostly sleep disruption and headaches, hold up under a controlled study. We could not find one either way.
We would rather leave a gap here than fill it with a number we cannot stand behind. If you have a source that closes one of these, send it.
Closer to home
Six projects in the Lehigh Valley, with their sources.