The AI Data Center Argument Is Being Had With Numbers Nobody Can Check

There are now at least 321 data center moratoriums on the books across 32 states. New Jersey alone has 72. Ohio and Michigan have 35 each. New York passed the first statewide moratorium in July. In the second week of September, the governor of New Hampshire said she would seek a multi year moratorium and the governor of Connecticut said his state “will not host massive AI data centers.” Data Center Watch counts 64 billion dollars in projects blocked or delayed.

Something real is happening here, and it is happening fast. Moratorium bills have been spreading through state legislatures all through this year, and the two governors above moved within the same fortnight.

I run a hosting company. We incorporated in 2008 as Innovative Scaling Technologies and we have owned and operated our own infrastructure ever since, which means I have spent seventeen years answering the questions those councils are only now learning to ask. Where does the power actually come from? What does it cost the people already on the grid? Who picks up the phone when it goes wrong? So when the moratoriums started stacking up, I went looking for the numbers underneath them. I expected to disagree with some of the conclusions and concede others.

What I found was not a disagreement about conclusions at all, which is the argument I had prepared for and would have enjoyed having. It was something further down, in the material everyone on both sides is building on.

The numbers are mostly untraceable. Not wrong, necessarily, and not invented. Untraceable: quoted in article after article, attributed to institutions that appear never to have published them, and impossible to follow back to anything you can open and read for yourself.

The report everyone is passing around

The document circulating hardest right now is the AI Data Center Ecology Report, published this year by Above Phone. They give it away free: their own announcement of it is abovephone.substack.com/p/the-ai-data-center-ecology-report” rel=”nofollow noopener” target=”_blank”>here, and the report lives at aidatacentermap.org. If you are going to read the rest of this, go and get it, because I am about to make specific claims about what is in it and you should be able to check them. It is 261 pages. It is well organized, it is clearly the product of real effort, and parts of it are genuinely good. The chapter on how a data center actually gets sited and permitted taught me things I did not know, and I am in this industry.

It also says this, on page 241, in its own AI Statement: writing and graphics are “Human-Only.” Citations are “AI-Generated.”

I want to be fair about what that does and does not mean. It does not mean the report is wrong. It means the attributions cannot be trusted as attributions, including the ones that turn out to be correct. When a number in that document is credited to “Bloomberg” or to “Senator Elizabeth Warren’s office,” that credit was produced by a language model, not by a person who read the source.

I had the three most quotable numbers in the report checked against primary documents.

The claim that consumers near data centers are seeing a 267 percent jump in wholesale electricity prices, credited to Bloomberg, has no URL anywhere in the 261 pages. The chapter it appears in has no links section at all.

The claim that tech companies deducted 152 billion dollars in 2022, credited to Senator Warren’s office, appears nowhere in the report’s Resources chapter. That chapter has eight subsections. It is not in any of them.

The claim that the Joint Committee on Taxation projects roughly 178 billion in corporate tax savings for 2026 names no JCT publication, table, JCX number or date. The report’s Government Reports list contains the IEA, the EIA, Lawrence Berkeley twice, and the USGS. No JCT.

Those are the three that travel. They are short, they are shocking, and they fit in a post. I have seen all three quoted in news write ups, in council testimony and in social threads, always with the same institutional names attached and never once with a link.

That is not really a coincidence, and it is not a failure of anyone’s diligence either. A number built for sharing does not need a source in order to travel, it only needs to sound as though it has one, and an institutional name in the sentence does that work perfectly well without ever being checked.

What actually held up, and why it matters more

One of the six came back clean, and it is worth dwelling on, because it is the single claim in the document that genuinely supports the case people are making with it. It is also stronger than the report made it sound.

Peer reviewed modeling published in Environmental Research Letters in June 2026 projects that data center and cryptocurrency growth through 2030 raises wholesale electricity costs by a national average of 6 to 29 percent depending on scenario, and by as much as 57 percent in the hardest hit regions. Twenty six interconnected power regions were modeled. The paper has a DOI. You can open it yourself: 10.1088/1748-9326/ae6c3d.

Buried in that paper is the sentence that should reframe this entire debate. Distributing new data center demand more broadly across the grid sharply reduces the regional price spikes, but barely moves the national average.

That cuts against almost every argument being made on either side of this, so it is worth sitting with for a second. Spreading the same total demand more widely across the grid substantially fixes the local damage while barely touching the national figure, which means the harm being done is not really a function of how much we are using in aggregate. It is a function of how much of it lands in one place at once.

The problem is not that we are using more electricity. The problem is where we are using it, how fast, and all in the same small handful of places. This is a concentration problem wearing a consumption problem’s clothes. Every policy argument that treats it as a national supply question is aiming at the wrong target, and so is every defense that points at national averages and says the panic is overblown.

One correction to my own side

When you see a project described as 380 megawatts, or a gigawatt, that is nameplate capacity. It is the theoretical maximum the facility could draw, not what it draws.

Lawrence Berkeley National Laboratory estimates these facilities run at about 60 percent of their operational capacity, and that they are running about 75 percent of the time. Multiply those two together and you get roughly 45 percent. So a project announced as 380 megawatts is drawing something closer to 170 megawatts on an average day, and a gigawatt project is closer to 450 megawatts.

Call it what it is. The headline figure is a ceiling, not a reading. It describes what a building is permitted to pull on its worst possible day, which is exactly the right number for an engineer sizing an interconnection, and the wrong one for describing what a town is actually living beside.

That correction cuts against my own argument, which is why it is here. It is also why you will not find a megawatt figure anywhere else in this piece. The ones in circulation are nameplate, and for the specific projects I would have quoted I could not get the underlying filings to check them, so I left them out rather than repeat a ceiling as though it were a measurement.

When you see one of these numbers, halve it before you pass it on. The honest figure is still alarming. The inflated one is easy to dismiss, and handing the other side an easy dismissal is how you lose an argument you should win.

The part that should end the argument

Here is what I could verify, and it is worse than the scary numbers.

Lawrence Berkeley National Laboratory estimates that fewer than one third of US data center owners measure their water consumption at all. An Uptime Institute survey found that as of 2021, half of surveyed facilities tracked water in some form and fewer than one in ten tracked it across their whole fleet. Operators told Uptime why. There was no business justification for measuring it.

That Uptime figure is five years old and I want to flag that rather than let it sit there looking current. But the direction is corroborated by Berkeley, and by the industry’s most transparent participant.

Google publishes its water numbers, which genuinely does put it ahead of most of this industry. In 2024 the company consumed roughly 8.1 billion gallons of water across its operations, and its data centers accounted for about 7.7 billion of that. Consumed, in Google’s own footnote, means water not returned to the source. So data centers are roughly 96 percent of the water Google takes and does not give back.

That share is not a one year artifact. In 2023 the same split was 6.1 billion out of 6.4 billion, again about 95 percent. The volume is climbing and the data center share of it is not moving. Whatever else is true about Google’s water, this part is settled: the offices are a rounding error.

I am quoting consumption here and not withdrawal on purpose. Google reports both, and published accounts of the withdrawal figure disagree with each other, most likely because some are counting freshwater only and others all water. I could not resolve which from outside the report, so I am not going to print a number I cannot stand behind in an article about printing numbers you cannot stand behind.

And in its own appendix, Google discloses that some facility water use is estimated from square footage and engineering principles rather than metered.

That is the most transparent company in the sector. Every efficiency ratio, every replenishment percentage, every reassuring comparison in this debate rests on numbers that in many cases were never actually measured.

The absences

Google, Microsoft and Meta publish global power and water totals. They are imperfect numbers, published late and on the companies’ own terms, but they exist and you can argue with them.

After those three it simply stops, and the drop is steeper than most people covering this seem to realise. Amazon, the largest of them by infrastructure, publishes a sustainability report containing neither a power figure nor a water figure. Efficiency ratios only, which tell you how well it uses what it uses and nothing at all about how much that is. xAI has published nothing since going private in 2022. Neither OpenAI nor Anthropic publishes environmental reporting of any kind, and both are now building capacity of their own.

I am naming the AI labs deliberately, and not because I think they are villains. The point is narrower and harder to argue with. Four of the seven largest builders of this infrastructure will not tell you what they use. You cannot regulate, price, or plan around a number that does not exist.

What it looks like when the process does not stop

South Memphis is the case with the full arc.

xAI’s Colossus was announced in 2024 with the usual pitch. Investment, jobs, a city at the center of the AI era. To power it, xAI ran methane gas turbines on site. In April 2025 the Southern Environmental Law Center flew a thermal imaging survey over the facility and documented turbines running well in excess of the fifteen the Shelby County Health Department had permitted.

The legal fight that followed is worth describing precisely, because the precision is the point. SELC, acting as counsel for the NAACP and the local group Young, Gifted & Green, appealed the health department’s permitting decisions. Separately, the NAACP sued xAI over turbines operating without an air permit at a second site across the state line in Southaven, Mississippi, represented by SELC and Earthjustice. TIME commissioned an air study and the University of Tennessee Knoxville ran the nitrogen dioxide analysis.

South Memphis is a historically Black neighborhood that already carries a disproportionate industrial load. That is why the NAACP is an appellant and not a bystander.

On water, Memphis Light, Gas and Water initially greenlit 1.3 million gallons a day. After sustained pressure from a local group called Protect Our Aquifer, xAI committed to using recycled wastewater. That concession is real and I will state it as one. I will also state its size. The recycled supply covers about 20 percent of the facility’s need. The other 80 percent comes from municipal supply drawn on the Memphis Sand Aquifer.

The company operating that facility publishes no environmental reporting at all.

Why I am writing this as a hosting operator

I should be straightforward about my position. I sell hosting. InnoScale owns and runs its own infrastructure rather than reselling someone else’s cloud, and has since 2008. I have a stake in how this conversation goes.

So let me be precise about what I am not claiming. I am not claiming my footprint is virtuous or that ordinary hosting is green. I am claiming that the workload driving this buildout is a specific one, and it is worth separating from everything else that runs in a rack.

For Google’s AI infrastructure, inference, meaning actually serving the models to users, is about 60 percent of AI energy use against 40 percent for training. A 2025 preprint benchmarking LLM inference puts it as high as 90 percent, and I flag that it is a preprint and not peer reviewed.

What that means is that the gigawatt buildout is not being driven by a few dramatic training runs. It is driven by AI features switched on by default across billions of devices, serving billions of requests a day. That is a genuinely new kind of load. It is not what your company’s website, your database, or your email is doing, and conflating the two helps nobody.

What I actually want

Not a moratorium. I want the thing that is missing underneath all 321 of them.

Measure the water. Publish the power. Disclose the contracts that local officials are currently permitting against without being allowed to see. Right now a groundwater district or a city council is asked to approve a facility whose actual draw is covered by a nondisclosure agreement, using projections nobody can audit, against a plan written before this class of facility existed.

You do not need to believe data centers are bad to believe that is a broken way to make a thirty year decision.

And if you are going to argue this in public, on either side, trace your numbers. The 267 percent figure is going to keep circulating because it is a great number. It just does not appear to have a source. We are going to lose arguments we should win by repeating things we cannot back, at exactly the moment when the peer reviewed evidence finally showed up and turned out to be stronger than the hype.

Sources and status

• Johnson, J.X., Wade, C., Blackhurst, M., DeCarolis, J.F., de Queiroz, A.R., Posen, I.D., & Jaramillo, P. (2026). Power system costs and emissions from data center and cryptocurrency mining expansion in the United States. *Environmental Research Letters* 21(11): 114007. DOI 10.1088/1748-9326/ae6c3d. Peer reviewed.
• Google. (2025). *Google Environmental Report 2025*, FY2024 data. Water table and appendix.
• Shehabi, A., et al. (2024). *2024 United States Data Center Energy Usage Report* (LBNL-2001637), Lawrence Berkeley National Laboratory.
• Bizo, D., Ascierto, R., Lawrence, A., & Davis, J. (2021). *Uptime Institute Global Data Center Survey 2021* (UI Intelligence Report 51). Flagged in text as 2021 data.
• Jegham, N., et al. (2025). How hungry is AI? arXiv:2505.09598. Preprint, flagged in text.
• Above Phone. (2026). *AI Data Center Ecology Report*. Free. Publisher’s announcement: abovephone.substack.com. Report: aidatacentermap.org. Cited here as an advocacy report, never as the basis for a figure. The AI Statement quoted above is on page 241. Go and check it.
• Amazon Web Services (2024). *2024 Amazon Sustainability Report: AWS Summary*.