Are AI data centers stealing all your water?

Looking at the facts (and fictions) on data centers and their impact on our world

Jake Handy·

Two years ago I wrote “Energy Gluttony in the AI Age.” I quoted the bottle-of-water-per-query stat and I drew the Bitcoin comparison. Looking back, I was both confidently wrong and confidently right (the best way to be, in my opinion). Looking forward, I’m seeing that the discourse around AI’s environmental concerns is only grown louder and more hostile.

So, I gathered reliable papers and reports from the past and present, and I’m finding that the honest picture is messier than either camp wants it to be. The viral “AI drinks a bottle of water” panic is wrong by orders of magnitude. The “data centers are an existential threat to your neighborhood” stories are real, sitting in legitimate court filings, and badly under-reported.

Below are seven claims you’ve probably seen on your feed more than once. I’ll tell you which are fact, which are fiction, and what we can do about all this.

Claim 1: “ChatGPT drinks a bottle of water every time you ask it a question”

Fiction.

The viral version says 500 ml per query. This traces back to a 2023 preprint by Shaolei Ren at UC Riverside titled “Making AI Less Thirsty,” later published in Communications of the ACM. Real paper with real, useful research. But Ren never wrote that a query takes 500 ml.

What he actually wrote was 500 ml per 10 to 50 medium-length GPT-3 responses. The “per query” part was a citation-chain mutation that escaped onto Twitter and never came back (The AIAAIC database has it filed as a discrete misinformation incident).

Then the debunking expanded.

Software engineer Sean Goedecke retraced Ren’s math in a careful write-up and argued the original Ren estimate compounded a per-page power figure misread as per-request, then applied it to GPT-3 (2020), a model roughly 10x less efficient than what’s serving you today. Goedecke’s reconstruction puts a current ChatGPT response closer to 5 ml. Google reports a median Gemini text prompt at 0.26 ml of water and 0.24 Wh of energy (their methodology, their disclosure, but it’s the only first-party number out there).

Sam Altman, when pressed on the bottle stat at the India AI Impact Summit in February, called it “completely untrue, totally insane, no connection to reality.” He’s a self-interested party (who also said really odd things about human energy use at this same summit), but he’s also right.

Andy Masley, who has spent an unreasonable amount of time digging through local AI water claims, puts the narrow version cleanly: “There are no places (so far) where it seems like data centers have raised water costs at all or harmed local water access.” Whether you trust his framing or not, the receipts are public.

One more data point I find darkly funny. Karen Hao’s 2025 book Empire of AI, one of the most cited critiques of the industry, contained a Chile data center water figure that was off by roughly 1,000x. She corrected it publicly. The original viral version of that number is still circulating on LinkedIn.

The bottle-per-query thing is a fiction that we should stop quoting.

(Hold that thought, because the upstream story, what data centers do to the power plants and aquifers that supply them, is a different conversation. Keep reading.)

Claim 2: “Every ChatGPT prompt is a climate sin”

Fiction.

The math on this is just bad. A single ChatGPT text query costs roughly 0.3 watt-hours. That’s the number Sam Altman put on his personal blog in June 2025 (0.34 Wh, technically) and the same number Epoch AI got independently in a February 2025 analysis using GPT-4o on H100 GPUs (for those that, understandably, doubt Altman). Hannah Ritchie at Our World in Data lined up Google’s published Gemini number (0.24 Wh median) and confirmed the order of magnitude.

To translate this into something tangible: 0.3 Wh is roughly two minutes of an LED bulb, or four feet of driving a sedan, or about five seconds of streaming Netflix in HD. Simon Willison summed it up in November: “A ChatGPT prompt equals about 5.1 seconds of Netflix.”

If you sent a thousand ChatGPT prompts a day, every day, for a year, you’d raise your personal energy footprint by less than 1%. Masley calculates a single prompt at roughly 1/150,000th of an average American’s daily emissions.

There is some credible dissent on this one, and her name is Sasha Luccioni at Hugging Face. Luccioni has been the most rigorous voice on the alarm side. She claims, reasonably, that reasoning models (5.5 Thinking, Opus 4.7 High, etc) burn around 30x more energy per query than chat-style queries. Long-context queries can hit 40 Wh.

Video generation is in another universe (a five-second AI-generated clip clocks in around 3.4 million joules per the MIT Tech Review investigation). Luccioni’s point isn’t that the 0.3 Wh number is wrong, it’s that the industry is sliding toward query types where 0.3 Wh stops being the average. And she’s probably right, given what we’re seeing with AI psychosis and slop cannons.

But on the question “should I feel guilty asking Claude to summarize a PDF” the answer is still no. The individual-user guilt framing has been the most wrong part of the climate-AI discourse for two years running; a bit reminiscent of big corporations insisting individuals be held responsible for not recycling while they’re off slugging down oil by the barrel full.

Claim 3: “AI is the new crypto, useless computational waste”

Fiction.

I made this comparison in 2024, and I take it back.

Bitcoin mining alone consumes roughly 138 TWh per year, comparable to the annual electricity use of a mid-sized country, doing math that (charitably) secures a payment network. The energy-per-utility ratio is unfavorable, to put it politely.

The comparison to AI was attractive in 2023 because both involved a lot of GPUs and breathless valuations, but the productivity story has since diverged hard. Stanford HAI’s 2025 AI Index found that inference cost for GPT-3.5-class performance dropped 280x between November 2022 and October 2024. Frontier-tier inference is on roughly a 40x-per-year cost decline. DeepSeek shipped R1 at a reported 20-50x cheaper inference price than OpenAI’s o1 (Altman’s own framing), and DeepSeek V3’s reported training compute cost was around $5.6 million (which is training compute alone, not total R&D or capex) against Llama 3.1 405B’s hundreds of millions on the same metric.

Bitcoin’s hashrate doesn’t get more useful when it gets cheaper, but AI tokens do. Every efficiency gain on the inference side translates directly into more capable models doing more work per watt. NVIDIA and SemiAnalysis report Blackwell delivering up to 50x throughput per megawatt over Hopper, with corresponding drops in cost per token.

You can also read this on the Handy AI Substack.

There’s also the actual outputs question. DeepMind’s RL cooling controller cut Google’s data-center cooling energy 40%. Google’s GNoME identified 380,000 new stable crystal structures, equivalent to nearly 800 years of accumulated materials science. Microsoft’s MatterGen, working with PNNL, narrowed 32 million candidate electrolytes to a viable lithium-reduced battery material in under a week. DeepMind’s tokamak controllers with EPFL and the open-source TORAX plasma simulator are showing up in Commonwealth Fusion Systems’ SPARC work. None of that is hypothetical.

I’m not arguing that AI is going to single-handedly solve climate. The IEA’s claim that AI adoption could enable 1.4 gigatons of CO2 reductions by 2035, 3-4x larger than data centers’ own emissions in their base case, is in the IEA’s own words contingent on adoption that “currently [has] no momentum.”

But Bitcoin produces a worse Bitcoin. AI produces a better grid forecast, a faster battery, and (somewhere down the line) maybe fusion. Those aren’t the same thing, and pretending they are is lazy.

Claim 4: “Data centers are jacking up your electricity bill”

Fact.

Now we’re in the part that’s true. The cleanest evidence on this comes from PJM Interconnection, the grid operator covering 13 states from Virginia to Illinois, and from Monitoring Analytics, its statutorily independent market monitor.

PJM’s capacity prices (what utilities pay to guarantee power will be available when demand peaks) went from $28.92/MW-day in the 2024-25 auction to $329.17 in 2026-27, and cleared at the FERC-approved price cap of $333.44 in the December 2025 auction for 2027-28 (an 11x increase in three years). Monitoring Analytics, PJM’s independent market monitor, attributed 63% of the price spike in the 2025-26 auction to data center load growth: $9.3 billion that gets recovered from ratepayers. Cumulatively across the last three auctions, data center demand has added $23.1 billion to PJM system costs.

The cleanest state-level case is Virginia. The State Corporation Commission issued an order on November 25, 2025 in Dominion Energy Virginia’s biennial review that adds about $11.24/month to a typical residential bill in 2026, and created a new “GS-5” rate class effective January 1, 2027 for customers above 25 MW. The new class makes hyper-scalers cover at least 85% of transmission/distribution demand and 60% of generation demand under 14-year contracts.

Basically, Virginia regulators officially decided that residential ratepayers shouldn’t subsidize Amazon’s GPU clusters anymore.

The Harvard Electricity Law Initiative documented that Virginia and Maryland ratepayers are footing “the lion’s share” of regional transmission built to serve “just a few of the world’s wealthiest corporations.”

This is the cleanest harm in my entire debate here. There is no serious argument against AI driving electric bills higher, it’s just a reality.

What you can actually do

Bills typically go up because state-level rate cases get rubber-stamped while no one watches. A few things can move the needle:

  1. Show up to your state Public Utility Commission’s rate-case comment process. Every state PUC runs a public comment process for major rate filings, and they are sparsely attended. The Virginia SCC GS-5 rate class only happened because consumer advocates kept pushing at hearings.
  2. Back cost-allocation reform. The fight is whether data centers pay for the grid upgrades they trigger or whether the rest of us do. Public Citizen, the Citizens Utility Board, and Sierra Club run state-by-state campaigns on this. Find yours and push on it.
  3. Push for large-load tariffs. Texas SB6 (signed June 2025) defined any load ≥75 MW as “large” and made it pay for transmission screening and accept curtailment. Data centers and electric bills won’t be in the public zeitgeist forever; now is the time to push your state legislator to act.

If your bill is going up, your neighbor’s ChatGPT habit isn’t the villain (well, depending on what they’re using it for). It’s the rate filing your PUC approved last quarter while your community didn’t show up.

Claim 5: “Big Tech is poisoning Black neighborhoods to build AI”

Fact (and it’s worse than you’ve heard).

Elon Musk’s xAI brought its Memphis “Colossus” supercluster online in 2024 inside a former Electrolux plant in South Memphis. Surrounding it on three sides are majority-Black neighborhoods including Boxtown, a community already on the EPA’s smog non-compliance list, with cancer rates roughly 4x the national average (and this was before any data center showed up).

To meet Colossus’s power demand, xAI installed gas turbines on-site. According to filings by the Southern Environmental Law Center, the company did so without the Clean Air Act permits that environmental groups argue were required. Aerial imagery cited by SELC on March 31, 2025 documented 35 turbines and roughly 420 MW of combined capacity (comparable to a TVA power plant) operating on-site, allegedly before the company had filed for permits. A second cluster at Colossus 2 in Southaven, Mississippi added 27 more. xAI still disputes the legal characterization to this day, while the Memphis turbines are the subject of active litigation.

The numbers from the EmPower Analytics study commissioned by SELC and led by Harvard-trained Dr. Michael Cork: the 33 turbines covered by the analysis have a potential to emit roughly 2,507 tons of nitrogen oxides per year, which would make xAI’s installation likely the single largest industrial NO₂ source in greater Memphis. The study models $30-44 million in annual regional health damages from premature deaths, asthma exacerbations, and cardiac events. Sensors installed by Memphis Community Against Pollution (MCAP) recorded peak NO₂ readings up 79% since xAI began operations.

The NAACP, SELC, Young Gifted & Green, and Earthjustice filed a formal Clean Air Act notice of intent to sue in June 2025. The Mississippi suit was filed in 2026.

KeShaun Pearson, executive director of MCAP (and probably the most articulate voice on Memphis air quality), put it this way: “We are, unfortunately, a cautionary tale about what will and possibly can happen if you don’t have the right rules and guardrails in place.”

Memphis is the headliner but it’s not the only story. Loudoun County, Virginia residents have filed complaints about a planned Vantage data center with 51 diesel backup generators and 8 natural gas turbines, modeled to cause 3.4-6.5 premature deaths annually. Iowa state regulators found 40 unpermitted water wells at a Cedar Rapids data center site in 2025. Google’s facility in The Dalles, Oregon used 25% of the entire city’s water in 2021, a fact Google tried to keep secret as a “trade secret” until The Oregonian sued and won.

The pattern seems to be fairly consistent: A hyperscaler shows up in a community without the legal muscle to push back, builds first, asks for permits second, and gets retroactive approval because the facility is already operating. No good.

What you can actually do

This is the area where individual action has the highest leverage, partly because it’s the area least covered by national media.

  1. Find out who’s building near you. Most data center proposals go through county planning commissions, not state legislatures. They get approved on consent agendas at meetings with three attendees. Search “[your county] + data center + planning.”
  2. Donate or volunteer with frontline groups. Memphis Community Against Pollution, Southern Environmental Law Center, and Earthjustice are doing the actual litigation while local NAACP chapters with environmental justice committees are doing the organizing.
  3. Read the air permits before they’re approved. Most state environmental agencies post draft permits with public comment periods of 30-60 days. The xAI permit likely got approved partly because almost no one wrote in.
  4. Pay for the journalism. Inside Climate News, Prism Reports, Democracy Now, and SELC’s comms team have been doing the work, and they survive only on subscriptions and earned attention.

The companies are betting nobody who reads about Memphis lives in Memphis. Disprove the bet.

Claim 6: “Data centers pay for themselves with jobs and tax revenue”

Fiction.

Every press release about a new data center includes the same three claims: massive capex, lots of construction jobs, and “permanent positions.” The first two are real but the third is a rounding error.

Specific cases, sourced:

The state-level subsidy ledgers are similar:

Good Jobs First, the nonprofit that tracks economic-development subsidies, ran the ROI math: states that actually compute it (most don’t) admit they lose between 52 and 70 cents on every dollar of data center subsidy.

The construction jobs are real, but they’re temporary. The “permanent positions” pitch is looking more and more like a bait-and-switch.

Claim 7: “AI is going to use as much electricity as Japan”

Fact (but the framing is doing some work).

The IEA’s April 2025 Energy and AI report projects global data center electricity consumption rising from roughly 485 TWh in 2025 to about 945 TWh in 2030 in its base case. That’s just under 3% of global electricity, growing to roughly Japan-scale by the end of the decade. The headline “AI is going to use as much electricity as Japan” is a 2030 claim.

Lawrence Berkeley National Lab’s December 2024 report, the most authoritative US-specific source, found US data centers used 176 TWh in 2023 (4.4% of total US electricity) and projects 6.7-12% of US electricity by 2028. McKinsey’s 2025 estimate is 11.7% of US electricity by 2030. Morgan Stanley’s high-end scenario tops out at 18% by 2030.

So, the aggregate story is real. But like the water bottles, the viral headlines are doing some heavy lifting:

  1. AI is not all of data centers. AI workloads are a growing subset (around 70% of the 2030 capacity buildout in McKinsey’s model), but the IEA’s headline 945 TWh by 2030 refers to total data center electricity, including all the cloud computing, video streaming, and corporate IT that was already there. Conflating “AI” with “all data centers” inflates the AI-specific number.
  2. 3% of global electricity is not an apocalypse. Global air conditioning is around 7-8% of global electricity and rising. Cement is roughly 7% of CO2 emissions. Aviation is 2-3%. Data center demand is real and growing, but it’s also not categorically different from other industrial loads we’ve already absorbed.
  3. The growth rate is the issue, not the level. Tyler Norris and the Duke Nicholas Institute’s “Rethinking Load Growth” work found that the existing US grid could integrate up to 98 GW of new flexible load with only ~0.5% annual curtailment (and 126 GW at 1.0%). That doesn’t mean the grid crisis evaporates (interconnection rules, reliability standards, and individual plant economics still matter), but it does mean that the disastrous “we need new gas plants because data centers can’t flex” framing is a choice the industry is making and not a real, physical constraint.

Where the alarm is warranted: the buildout of new natural gas peaking plants justified specifically by AI loads. Goldman Sachs estimates 15-30 GW of new US gas capacity through 2030 attributable to data center demand. That’s the part that locks in fossil emissions for decades and that the clean PPA announcements don’t actually offset on the timeline that matters. We need to avoid this at all costs.

Where the alarm is overblown: per-query, per-user, per-day terms. Hyperscaler clean-energy procurement is also real. Amazon, Meta, Google, and Microsoft together accounted for 49% of all global corporate clean energy procurement in 2025. Microsoft is single-handedly responsible for restarting Three Mile Island. Meta signed a 20-year PPA with Constellation for the Clinton nuclear plant. Google’s Kairos SMR deal is the first multi-SMR corporate fleet contract in history. None of this happens without the AI buildout.

It is genuinely possible for the AI industry to be driving the biggest expansion of US nuclear since the 1980s and locking in 20 GW of new gas peakers in the same five-year window.

What you can actually do

  1. Push for flexible interconnection, not bans. The Duke work shows the grid could integrate substantially more large load if data centers accept modest curtailment (~0.5% per year). That doesn’t make the rest of the integration problem trivial obviously, but it does change the policy conversation from “build more gas” to “design better tariffs.” Demand-response programs and large-load flexibility tariffs are the lever. Most state legislatures don’t have one. They should.
  2. Track gas plant approvals in your state. Every new gas plant justified by AI demand locks in 30-40 years of fossil generation. State PUCs approve these and, don’t forget, they run public hearings!
  3. Subscribe to a green tariff or community solar. Hyperscalers are buying up the clean energy supply that should be cheap and available for households. Demand-side pressure on utility green tariff offerings matters more than you realize.
  4. Don’t let “AI is bad for the planet” become an excuse for status quo gas. The honest fight isn’t AI vs. no-AI, it’s which energy gets built next. We need to make that fight harder for the gas-plant developers than for the solar and nuclear ones.

Where this leaves me

The prevalence of AI services in my workday isn’t going anywhere anytime soon. The discourse where one camp says “you murder a tree every time you ask Claude to help you draft a Slack message” and another camp says “the environmental concerns are made up by Luddites” is exhausted and wrong on both ends.

A single query is trivial. The aggregate buildout is real, is costing everyone money on their power bill right now, is concentrated in communities that can least defend themselves, and is producing the largest expansion of US nuclear since the 1980s (alongside an unconscionable wave of new gas plants). There is no neat narrative, there is only what is happening.

Stop quoting the water bottle and start showing up to the rate case.

Originally published on the Handy AI newsletter →