Centralized AI data centers are draining our water, dirtying our air, and jacking up our power bills. The "decentralized" alternative is just the same problem in a smaller box, backed by the same people.

I want to talk about the madness that is currently unfolding in backyards, in rural farmland, along aquifer-fed communities, and in the server rooms of billionaires, because I don't think most people have connected all the dots yet. And once you do, you can't unsee it.

There are two parallel stories happening in AI infrastructure right now. One involves mega data centers consuming as much power as small nations and draining community water supplies dry. The other involves a shiny "decentralized" alternative that sounds like a solution until you notice who's funding it and who quietly benefits. Spoiler: it's the same cast.

The Scale of What's Actually Happening

Let me start with the numbers, because they're the kind that make you set down your coffee.

In 2024, a single data center in Iowa consumed one billion gallons of water, which is enough to cover all residential water use across the entire state for five days. One facility. One year. According to research from Lawrence Berkeley National Laboratory, U.S. data centers collectively consumed around 17 billion gallons of water directly through cooling in 2023, plus an estimated 211 billion gallons indirectly through the electricity required to run them. By 2028, those figures could double, or even quadruple.

Google reported its data center water consumption grew from 4.3 billion gallons in 2021 to 6.1 billion gallons in 2024. That 24% year-over-year jump isn't a fluke, it's a trend line pointed in one direction. And these are the companies that at least publish some data. The industry as a whole is not exactly rushing to put this information on the front page of its sustainability reports.

It gets more specific, and more troubling, at the community level. The city of Joliet, Illinois, has historically relied on an aquifer for its water. Recent industrial development and data center expansion in the Chicago region have depleted that aquifer to the point where researchers estimate it will be gone by 2030. In Fayette County, Georgia, a data center was recently discovered to be secretly drawing 29 million gallons of water through two connections the county didn't even know existed. Residents only found out because their water pressure dropped so dramatically they couldn't ignore it anymore.

Two-thirds of all data centers built or in development since 2022 are located in water-stressed areas: southern Arizona, the Colorado River Basin, and Texas. These aren't coincidences. Land is cheap there. Regulation is thin. And apparently, so is accountability.

On the electricity side, data centers consumed about 4.4% of total U.S. electricity in 2023 and are projected to exceed 9% by 2028. To generate a single 100-word AI response, researchers at UC Riverside estimated the equivalent of roughly one bottle of water is consumed. Not a dramatic number on its own, until you multiply it by hundreds of billions of queries a day, every day.

Erin Brockovich Shows Up Again, For Good Reason

If you know anything about Erin Brockovich, you know she doesn't show up for small problems. She showed up in Hinkley, California, when PG&E was poisoning groundwater with chromium-6. She showed up in East Palestine, Ohio, after a train derailment poisoned an entire community. She doesn't make noise unless there's something worth making noise about.

In late May 2026, she launched brockovichdatacenter.com, an interactive map tracking operational, under-construction, and proposed AI data centers across the United States, overlaid with community-reported concerns. Within the first week, over 1,600 reports came in from 47 states. By early June, the total had surpassed 10,000 submissions, and she said in interviews that even that number represents only a fraction of the people who reached out.

What struck her wasn't any single incident. It was the pattern. Hinkley was one town, one aquifer, one chemical. This is Hinkley repeated in multiple counties and cities across every state, and nearly all of it happening in secret, protected by non-disclosure agreements between developers and local officials. A review of 31 Virginia municipalities with existing or proposed data centers found that 80% of them had NDAs in place. People weren't told. Councils weren't consulted. The water started disappearing, and then somebody finally made a phone call.

She's been particularly vocal about what's happening in rural areas, where the old data center playbook of suburban industrial parks has given way to something more aggressive. These facilities are now being planted in agricultural communities where residents rely on private wells. The first sign something is wrong isn't an environmental study. It's that the water pressure in their kitchen sink drops.

A Gallup poll from March 2026 found that 7 in 10 Americans oppose the construction of an AI data center in their local area, including 48% who said they're strongly opposed. Seattle passed a one-year moratorium. Multiple Kentucky counties followed suit. Cities in Georgia banned them outright. Maine's legislature passed an 18-month ban before the governor vetoed it.

The people are not confused. They know what they're looking at. The question is whether the people with the money and the political access are listening, and based on the NDAs and the fast-tracked permits, the answer so far is mostly no.

Elon Musk's Air and Water Problem in Memphis

Elon Musk built his xAI "Project Colossus" data center in south Memphis in 122 days. That's faster than most cities can issue a zoning permit for a gas station, which tells you something about what happens when you have enough money and the right political connections.

The facility, housed in a 750,000-square-foot former Electrolux plant in the Boxtown neighborhood, requires approximately 150 megawatts of power, enough electricity to power 100,000 homes, and roughly 5% of Memphis Light, Gas & Water's total daily load. Memphis residents didn't find out about any of this until it was already announced. There was no public review, no environmental impact study, no community input.

The power situation got creative fast. xAI installed gas turbines, lots of them, to generate electricity behind the meter. According to the Southern Environmental Law Center, 26 of those turbines were initially operating without proper environmental permits under the Clean Air Act. Research from the University of Tennessee found that average nitrogen dioxide concentrations around the facility increased by 3% after xAI began operations, with peak concentrations rising by 79% in the immediate surrounding area, and 9% in nearby Boxtown. These are not abstract statistics. These are the lungs of the families who live there.

As Tennessee state representative Justin Pearson put it: "Our lives and our lungs are being sacrificed on the altar of their capitalistic exploitation."

Musk's xAI is now expanding. A second facility in Whitehaven is reportedly double the size of Colossus. A third facility, called MACROHARDRR, has been announced in Southaven, Mississippi, with a $20 billion investment. When complete, xAI's combined Memphis-area infrastructure could require close to 2 gigawatts of electricity, enough to power about 1 million homes. To supply that kind of power, Musk has reportedly purchased an entire power plant overseas and is having it shipped to the U.S., because you simply can't get a new one built in time domestically.

On the water side, Musk's team did take one notable step: rather than drawing from Memphis's famous pure underground aquifer, xAI committed to funding an $80 million wastewater treatment plant. That's real, and worth acknowledging. But it doesn't erase the air quality impact, the secretive process, the unlicensed turbines, or the billions in ratepayer-subsidized infrastructure that residents are being asked to quietly absorb.

Kevin O'Leary and the 40,000-Acre "Opportunity" in Utah

If Musk is the impulsive billionaire who builds first and asks for forgiveness later, Kevin O'Leary is the investor who frames environmental destruction as a national security imperative.

O'Leary is the force behind the "Stratos Project," also being called "Wonder Valley Utah," a proposed AI data center campus in Box Elder County. The scale is almost hard to process: 40,000 acres, roughly the size of Washington, D.C. Up to 9 gigawatts of power demand, more than double the total annual electricity consumption of the entire state of Utah. A projected $100 billion in total investment.

On May 4, 2026, the Box Elder County Commission voted unanimously to advance the project. The meeting, attended by hundreds of protesters, grew contentious enough that one commissioner told the audience to "grow up," after which the commission withdrew to a private room and approved the project while residents watched on a livestream.

Robert Davies, a physics professor at Utah State University, warned that the facility could generate the thermal equivalent of 23 atomic bombs per day in waste heat. The Great Salt Lake, already at record-low levels and in what Davies calls "a fairly advanced stage of collapse," sits a few miles from the proposed site.

O'Leary, in response, called concerns about the Great Salt Lake "ridiculous," dismissed the project's water impact as a non-issue, and suggested on Fox News that the protesters might be paid operatives of China. He claimed without evidence that more than 90% of demonstrators were bused in from out of state. This is how a Shark Tank celebrity investor handles democratic opposition: by questioning whether the democracy is real.

His parallel project, Wonder Valley Alberta, has already run two years behind schedule. The Utah project doesn't have a tenant yet. But the land has been approved, the county has been convinced, and another rural community gets to absorb the risk.

The Decentralized "Solution" and Its Hidden Architecture

Now let's talk about the other side of this story, because it's being sold as a solution, and that framing deserves some scrutiny.

Earlier this year, a San Francisco startup called Span, in partnership with Nvidia and homebuilder PulteGroup, announced a product called XFRA. The concept is a small, liquid-cooled compute node, roughly the size of an air conditioning unit, installed beside a home. The node packs 16 Nvidia RTX Pro 6000 Blackwell Server Edition GPUs and 4 AMD EPYC CPUs. In exchange for hosting it, the homeowner gets discounted electricity, free internet, and a backup battery. Span plans to pilot 100 homes in the southwestern U.S. in late 2026 and scale to 80,000 nodes nationwide by 2027, delivering over 1 gigawatt of distributed compute capacity.

The pitch is genuinely clever. The mega data center is the villain, right? It consumes enormous amounts of power, depletes aquifers, poisons air, and shows up in your community without permission. The distributed node is the antidote: smaller, quieter, spread across existing residential grid capacity rather than demanding enormous new infrastructure buildouts.

But here's what the marketing doesn't lead with.

Span is backed by Nvidia. Not just partnered with them, backed by them. Nvidia has been building an investment empire in AI infrastructure companies with a pattern that critics have called "circular financing": Nvidia invests in startups, those startups use the capital to buy Nvidia GPUs, Nvidia collects both investment returns and hardware revenue from the same transaction. In 2024 alone, Nvidia participated in 50 funding rounds, and its portfolio now includes xAI, OpenAI, CoreWeave, Lambda Labs, Scale AI, and dozens of others. Span fits the pattern precisely: Nvidia backs the company, the company deploys Nvidia's GPUs at scale, Nvidia wins on both ends.

And then there's the Musk thread. Nvidia was a named investor in xAI's $6 billion funding round. Musk's xAI is building facilities that will run one million Nvidia Blackwell GPUs, the same GPU architecture that Span is now proposing to install in 80,000 American homes. The ecosystem here is not diverse. It is tightly interconnected, and many of the same financial interests sit at multiple nodes within it.

When Span says it's "closing the speed-to-power gap" for AI infrastructure, what it means is that traditional data center development is too slow because grid interconnection can take up to seven years. The distributed approach bypasses that bottleneck by tapping residential electrical capacity that already exists. Which sounds efficient, until you ask whose problem is actually being solved. The homeowner gets subsidized utilities and a battery backup. Nvidia sells more GPUs. Span collects fees from the hyperscalers renting compute time. And the AI compute buildout that communities across the country have been desperately trying to slow down continues at scale, now distributed invisibly across suburban backyards.

The XFRA node is explicitly designed to run AI inference workloads, the phase of AI operation that happens after a model is trained, when actual queries come in and get answered. Security researchers and grid experts have flagged real concerns: physical tampering with nodes at residential locations, side-channel attacks on compute hardware hosted in unsecured environments, and grid instability if large numbers of nodes cycle on and off simultaneously. The pilot locations are reportedly in Nevada or Arizona, both already in the southwestern water-stressed region that researchers keep flagging as the worst possible geography for AI infrastructure expansion.

What's Actually Going On Here

When you step back and look at the full picture, a clear structure emerges.

The centralized model operates on extraction: find the cheapest land with the least regulatory oversight, drain whatever water and power you need, protect yourself from accountability with NDAs, and call it national security when anyone complains. The communities that bear the cost, air quality degradation, water depletion, rising utility rates, grid strain, are not the communities where the profits accrue.

The decentralized model is presented as the democratic alternative, compute for the people, in the people's backyards. But it's being architected and funded by the same companies that are also building the centralized model. Nvidia has a financial stake in xAI and in Span. The distributed network of residential nodes doesn't replace the mega data centers. According to Span's own announcement, XFRA is explicitly designed to "augment" centralized facilities, not replace them. The distributed layer handles inference, the big centers handle training. You need both. So the homeowner hosting a node in Nevada isn't opting out of the centralized infrastructure race. They're just adding capacity to it, while the companies that stand to gain the most remain exactly the same.

What Erin Brockovich understood in Hinkley, and understands again now, is that the pattern of harm follows a predictable shape: a corporation identifies a resource it can consume cheaply, secrecy gets baked into the approval process before community members can organize, and by the time the public finds out what's happening, the infrastructure is already in the ground. Or bolted to the side of a house.

The Joliet aquifer will be gone by 2030. The Great Salt Lake is collapsing. Memphis neighborhoods are breathing nitrogen dioxide spikes 79% higher than pre-xAI levels. And 10,000 people have sent Erin Brockovich their stories from 49 states.

The question isn't whether we need AI infrastructure. I'm writing this on a device powered by it, and I have no illusions about my own complicity in this system. The question is whether the people who bear the environmental and financial cost of that infrastructure get any meaningful say in how it's built, where it's built, and who profits from it. Right now, the answer is largely: only if you organize loudly enough, for long enough, before the NDAs get signed and the turbines go online.

The communities showing up at county commission meetings in Utah, the residents in Memphis watching nitrogen dioxide readings tick upward, the rural farmers watching their well pressure drop, they're not anti-technology. They're asking for the thing that every citizen deserves: a transparent accounting of what something costs and who pays.

That seems like a reasonable thing to ask.