Guest contributor: Jonah Lipsitt, PhD, MSc is a researcher at Forward Future. Before joining, Jonah spent nearly seven years at UCLA earning his PhD in Environmental Health Sciences, publishing 13 peer-reviewed papers across sustainability, public health, and geospatial applications. He's held research and engineering roles at Zipline and Deloitte.
The Jar of Brown Water
Recently, Alexandria Ocasio-Cortez (AOC) held up a jar of brown, dirty water on the floor of Congress and said it was the drinking water of a community living next to a Meta data center in Georgia. She was pressing an EPA official about it during a May 2026 hearing, and the moment went viral. It did something I think we’ve badly needed: it forced a real conversation about what AI and data centers are doing to the environment, and to environmental health, the messier question of how the environment around us affects our bodies.
Her demonstration was not quite right, even if powerful.
Where AOC Misleads
Three things.
First, it isn’t true that the whole county’s water turned brown. The affected homes are likely a handful of properties near the construction site running on private wells, their own localized systems, drawing straight from the ground with no utility or treatment plant in between. The most-documented case is Beverly and Jeff Morris, whose house sits about 1,000 feet from Meta’s Stanton Springs campus on the Newton–Morgan county line. That’s a real hardship for those families. It is not the same as “the county’s drinking water is now brown.”
Second, there is nothing about running AI that turns water brown. Serving inference doesn’t produce sediment. The brown water is almost certainly from construction: clearing land, blasting, disturbing soil, dewatering, dropping the local water table. The tell is the timeline. The Morris family’s water trouble started back in 2018, when crews first began clearing the land, years before a single server switched on. Meta, for its part, commissioned its own groundwater study and concluded its work was unlikely to have affected the well, citing watershed and topography. But residents don’t buy it, and the EPA officially agreed to look into it. But notice what’s actually being argued over: the construction, not the AI or the data center operations.
Third, that construction point is bigger than data centers. Any large build (a warehouse, a highway interchange, a stadium) done close enough to a well without proper mitigation can foul someone’s groundwater. There is nothing uniquely AI about a backhoe.
So when AOC calls the jar “drinking water” destroyed by a data center, she’s compressing a construction story, a private-well story, and an AI story into a single image. They aren’t the same story.
Where She’s Right
Even if AOC’s theatrics are a bit problematic or misleading, was she right to drag this into Congress and the cultural spotlight? Yes. I think so.
We are far too late to this conversation about energy, carbon, materials, and the local cost of AI. That last one on local impacts leads to the asymmetry I can’t get past. When a town gets a new bridge, or even an Amazon warehouse, the people who live there get something back: they drive across the bridge, they get their packages a day sooner, some of them get jobs. A data center is different. Unless you happen to be a power user of AI who benefits from living a few milliseconds closer to the compute, the building next door is mostly imposition. (At least for now while AI is still nascent.)
Now imagine it’s your town. In a matter of months, in rural America, one of the largest buildings on the planet goes up next to your community.
What This Infrastructure Actually Costs
Humans aren’t going to stop building things, and building things has always cost the environment something. These projects devour concrete and steel. The brown-water phase AOC is pointing at is the construction phase: the loud, dusty, dirty part.
But the operational phase has its own bill. To run, data centers are packed with chips made from rare earths and other materials, much of it pulled out of the ground by mining that is brutal on the environment in its own right. Then, once it’s live, the data center burns energy to serve inference, and if that energy, or the cooling, throws off emissions, you can get a global cost (more warming from climate change) and a local one (worse air quality).
Data centers were about 1.5% of the world’s electricity in 2024, and the IEA expects that to roughly double by 2030. That is not the whole grid. But locally it can feel weird if the compute gets used in Palo Alto, and the infrastructure shows up in Georgia.
Memphis Is the Cleaner Example
If you want the sharper version of this story, don’t look at the brown jar in Georgia. Look at Memphis.
To power its Colossus supercomputer in South Memphis, Elon Musk’s xAI installed a fleet of gas turbines, essentially an on-site methane power plant. The trick was calling them “temporary,” which let them run without the air permits a normal power plant needs. By the spring of 2025, environmental groups counted as many as 35 turbines on the site, many operating with no permit at all. Then, in July 2025, the Shelby County Health Department gave xAI the clearance it wanted, a permit for 15 gas turbines, granted after the fact. By 2026 the NAACP, with the Southern Environmental Law Center and Earthjustice, had sued, arguing the broader power plant (dozens of unpermitted turbines across the Memphis area) is among the largest sources of smog-forming NOx in a region that already fails federal air standards.
So Why Pay the Cost?
So why build any of this?
Because a lot of people (including me) believe AI can be a net positive for humanity. That is AI’s environmental bargain: humanity and the planet absorb real costs now for a future benefit we can imagine but cannot yet prove. And I want to be honest about that: it is still a belief, with hints of proof, not a known outcome.
I don’t think humans, or Americans specifically, will change our individual behavior enough to head off the worst of climate change. The consumption we’d have to give up to bend the heating curves is, realistically, not happening. Much of the global impact is already locked in. We are headed for a hotter, more volatile climate, with all the uncertainty and strife that comes with it. A lot of people in AI believe this technology might be one of the only tools fast enough to claw some of that back: that even though AI emits carbon now, something ahead of us, maybe superintelligence, maybe just every climate scientist on earth working alongside a capable model, helps us reverse the trend faster than we otherwise could. In a way, we are accepting real environmental and community costs now in the hope that future technological progress helps reduce larger costs later. It’s like a vaccine. It hurts a little now, to be better protected later — given AI can help us both mitigate and adapt to future climate.
My favorite example of how AI can bring hope is fusion. When Lawrence Livermore’s National Ignition Facility finally hit fusion ignition (getting more energy out than the lasers put in), they leaned heavily on machine learning to design the implosion and dial in the lasers on the fuel. Although not Large Language Models (LLMs) the lab calls AI a key factor in getting there.
In May 2026, an OpenAI reasoning model disproved a longstanding conjecture in discrete geometry, on a unit distance problem that Erdős first posed in 1946, turning up a new construction that a group of outside mathematicians then verified. By OpenAI's account, it was the first time a general-purpose model cracked a prominent open problem on its own. If AI can keep chipping away at problems like that, the upside to science and the rate of scientific improvement is enormous. If you believe AI can 10x, or 100x, or 1000x our productivity or scientific progress, then you might think we can get to some of these environmental solutions faster too. We just need to get these AI tools into the hands of our brightest minds.
The other reason to pay the cost lies in the other costs we are already paying, that may be better cut first. If you take more than one international flight a year, that single trip almost certainly out-emits your entire year of ChatGPT or Claude use (for the average user): a transatlantic round trip runs on the order of a tonne of CO2 per passenger, while a heavy year of personal chatbot use is measured in kilograms. Aviation was about 2.5% of global energy-related CO2 in 2023, larger than every data center on earth combined. So is AI really the thing to give up? Would flying more help us reverse climate change? Yes, we absolutely need to reduce consumption in general if we are to stave off the worst impacts of climate change, but what goes first?
Who Bears the Brunt, and Who Benefits
So let’s give AI the benefit of the doubt. Say it’s worth building, even with the costs. And say the optimists are right that we can shrink those costs over time, with more efficient chips, cleaner mining, and hardware that lasts longer. (The chips really are improving fast: today’s AI hardware gets roughly 34–40% more efficient every year. The catch is that we keep buying so much more compute that total demand climbs anyway.) Fine, grant all of it.
The real question still stands, and it’s the one AOC is actually pointing at: who bears the worst of the cost, and who captures the benefit? While data centers bring tax revenue, economic deals, and jobs to rural areas, is it fair for Silicon Valley companies to reshape the local air, water, and ground in Georgia and Tennessee to achieve their ends? And when those same companies cut corners (skipping environmental review, taking carve-outs to move faster, running turbines without permits), of course it stops feeling fair.
We have to protect the most vulnerable people from the sharpest edges of this. We cannot let AI become an environmental justice problem. We cannot claim AI is for the good of humanity while being careless with the humans who need protection the most right now, in the present, before any of the promised future arrives.
To be clear - many leading AI companies are not ignoring these concerns. Data center construction and operations are getting cleaner and more efficient. Closed-loop water-cooling systems are possible. Frontier labs have publicly committed funding toward health, economic opportunity, AI resilience, and community programs – including paying for household utility costs around the datacenters they build. The labs have specifically funded outside research examining who benefits from AI and who may be harmed by it. Some of the largest studies on Universal Basic Income (UBI) have been funded by frontier labs. It is very hard to see around the corner of this massive technology, and independent research is likely key in understanding what lies ahead. If AI creates extraordinary value, some meaningful share of that value should flow back to the people and places absorbing its costs, and we need to better understand what those costs are.
The Race We’re Actually In
There’s a national-security version of the “go fast” argument that I actually buy: I would rather the world be shaped by AI developed in democratic societies than by AI developed under authoritarian systems. But if speed is the justification, let’s be honest about the race we’re actually in – one for not just AI and energy, but renewable energy. While we fight over gas turbines in Memphis, China is out-building the entire planet on energy: in recent years it has installed more new wind and solar than the rest of the world combined, several times what the U.S. manages in a year, and it already makes around 80% of the world’s solar panels and three-quarters of its batteries. China is still the biggest emitter on earth and still approves new coal, but its coal generation is now actually falling as clean power soaks up almost all of its new demand. They are quietly winning the cheap, clean power that AI will ultimately run on.
If we’re going to win, let’s win on the thing nobody has cornered yet: doing it right. Lead with renewables instead of diesel and methane turbines. Lead with closed-loop cooling and honest environmental review instead of waivers and carve-outs. Lead by taking care of the towns we build in and the people who get hit first. And lead with AI that’s genuinely meant to benefit all life, not just the sliver that can afford a subscription. A race to power the future on dirty energy, won over the backs of the communities hosting it, isn’t the kind of win that protects a way of life worth protecting.
The Solution Isn’t To End AI
The solution was never to stop AI development. The solution is to do and be better, faster, and that obligation falls on everyone holding a piece of this: the companies, the regulators, the local governments handing out the permits, and us. I believe AI will significantly benefit humanity, including those most vulnerable. But we have to be careful as we ‘go fast and break things’. It is not worth building a robot utopia if the cost is staying inside 24/7 because the air outside is too hot or gross to breathe. The physics of the climate and the environment are too large a system for superintelligence to fix overnight. And most importantly, AI isn’t the only thing to consider as stewards of the earth – many more of our behaviors and endeavors matter just as much.
On AOC: she was wrong to mislead. She was right to start the conversation. I hope the theatrics end up helping more than they hurt.
 | Jonah Lipsitt, PhD, MSc is a researcher at Forward Future. Before joining, Jonah spent nearly seven years at UCLA earning his PhD in Environmental Health Sciences, publishing 13 peer-reviewed papers across sustainability, public health, and geospatial applications. He's held research and engineering roles at Zipline and Deloitte. |
