AI in Your Backyard: The Water and the Wires

John Mauldin

Artificial intelligence is one of the defining topics of the day, but much of the public debate is driven more by opinion than by facts. I have been collecting research on AI for a long time, and in this letter I want to examine why so much opposition has formed around AI data centers.

Some of the resistance reflects the sweeping claims made by AI hyperscalers, who describe a future in which AI and robots do everything. Figures such as Sam Altman, Peter Diamandis, and Elon Musk speak of a post-work future and a Star Trek-like world of abundance, where people work only because they want to. On the other side are the pessimists who imagine a Blade Runner future, in which a small elite prospers while everyone else struggles.

The most negative view of any new technology is usually the easiest to make, and it is often wrong. In my view, neither extreme is realistic. Human beings are adaptable. AI is another tool, like electricity, cars, or the internet. It is up to us to decide how it is used. AI is the platform, not the artist.

Even so, there is significant anxiety around the infrastructure that supports AI. This week, I am focusing on the objections directed at AI data centers. Next week, I will address the question of whether AI kills jobs or creates them. The short answer is that it will do both, just as steam engines and automobiles did.

The reason this issue matters is clear. Seventy-one percent of Americans say they oppose a new data center being built near them. That is a supermajority, and it is a higher share than those who oppose a nuclear plant in their backyard. I tend to be skeptical of consensus views, especially when they form faster than the facts behind them. When I looked into the data, I found a large gap between sentiment and substance.

The two objections I hear most often are that data centers are draining water supplies and that they will overwhelm the power grid and raise electricity bills. In practice, these concerns are a mix of real engineering challenges and misinformation. Let’s take them in turn.

The Water Math Does Not Support the Panic

Water is the objection that is repeated most confidently and understood least well.

The best current estimate is that U.S. data centers directly consumed about 17 billion gallons of water on-site in 2023 for cooling. If indirect water use from the electricity that powers them is included, the total footprint rises to roughly 228 billion gallons a year. That may sound large, but it looks very different when compared with other sectors.

America’s golf courses consume an estimated 425 billion to 493 billion gallons of irrigation water each year, depending on the assumptions used. By different calculations, golf uses anywhere from 24 times to more than 100 times as much water as data centers. I do not say this to attack golf; I enjoy it like many others. I mention it because a hobby played on roughly 16,000 courses across the country, and one that almost no one protests at zoning hearings, uses far more water than the industry currently at the center of so many water-crisis op-eds.

Golf is also small beside irrigated agriculture, which consumes roughly 73 billion gallons of water per day, or about 27 trillion gallons a year, across the United States. At the high estimate of 228 billion gallons a year, data centers account for well under 1% of that total.

There is also a useful way to think about value per gallon. By one estimate, data centers generate about 92 cents of economic output per gallon of water consumed, compared with roughly 8 cents for golf. That is better than an 11-to-1 advantage.

None of this means water use is irrelevant. Some data centers, particularly in drought-stressed regions such as the Southwest, deserve close scrutiny regarding site selection and sourcing. But the technology is moving in the right direction. Direct-to-chip liquid cooling, which is becoming more common, can reduce a facility’s water use by as much as 95% compared with older evaporative systems. Immersion cooling can eliminate evaporative water loss almost entirely. More new facilities are also using closed-loop systems that recycle water rather than continuously drawing fresh supply.

That shift matters because the public debate often treats today’s cooling methods as if they are fixed forever. They are not. The industry most likely to improve its own water footprint is one built on $50 billion facilities and shareholders who read utility bills. If you want to worry about American water use, I would look first at almonds, alfalfa, and fairways before server farms.

A sidebar on alfalfa: large corporate farmers in California, including George Soros, grow alfalfa, which is highly water-intensive. The crop is then shipped in empty containers to China, where it is used to feed cattle. While estimates vary by year and source, California is thought to ship 70 billion to 100 billion gallons of water a year in the form of alfalfa, enough to supply a year’s water for about a million households. That is several times the direct use of AI data centers.

And almonds are even more notorious. One almond takes 1.1 gallons of water to produce. A serving of almonds can therefore represent about 20 gallons of water. California uses more than 1.6 trillion gallons of water annually to grow almonds, which is about 13% of the state’s water use. A typical first-class airline cabin uses 320 gallons of water or more serving almonds. Technology is improving data center water use; it is not fixing almonds, alfalfa, or other water-intensive crops.

The Grid Problem Is Real — But It Is Not What You Were Told

The power issue is more complicated, and the skeptics have a legitimate point.

Data center electricity demand is growing faster than the American grid was built to handle. The Electric Power Research Institute (EPRI) projects that data centers could increase from about 4% of total U.S. electricity consumption in 2023 to as much as 20% by 2035. It is also true that interconnecting new large loads has become painfully slow. A process that took about 15 months two decades ago can now take years, and in parts of Texas, some requests in ERCOT’s queue are not expected to receive power until 2031 or 2032. As one energy analyst put it, poles and wires are not exponential technologies.

That is the real issue behind the grid objection: the pace of physical infrastructure has not kept up with the pace of digital demand.

Where the popular narrative goes wrong is in assuming that data centers will simply pull power from the public grid while residential customers absorb the consequences. That is not what is happening.

First, a growing share of new data center capacity is not drawing from the public grid at all. Developers are increasingly generating power behind the meter, on-site, to avoid long interconnection delays. Florida and Georgia currently have more grid capacity than many people assume, and Texas has been able to supply grid power to data centers even as it builds substantial behind-the-meter generation in parallel. Where large gas turbines are unavailable — and 400-megawatt turbines are essentially sold out seven years in advance — operators are turning to repurposed jet engines. Boom Supersonic, the supersonic aviation company, now builds turbines for data centers, and Caterpillar reciprocating engines are being deployed at scale as well.

Second, utilities and grid operators are changing the rules to accelerate connections. In June 2026, ERCOT adopted a rule allowing data centers to accept interruptible status, meaning they agree to be curtailed when the grid is stressed in exchange for a much faster interconnection timeline. That cuts the wait from five to seven years to roughly 12 to 18 months. The Federal Energy Regulatory Commission has also sent letters encouraging six other regional grids to adopt similar rules. Tens of gigawatts of projects are already lining up to use these arrangements.

Third, nuclear power is seeing a real revival because of data center demand. Microsoft has already signed a power purchase agreement tied to restarting a reactor at Three Mile Island. France, by comparison, generates about 80% of its electricity from nuclear and exports the surplus to the rest of Europe, showing that the model works when a country commits to it. The United States is far behind.

Small modular reactors and fully factory-built reactors are still years away from commercial scale, but one energy forecaster said AI data centers may be the best thing ever to happen to the nuclear industry because, for the first time in decades, there is a large, creditworthy buyer willing to sign long-term contracts for firm, round-the-clock, carbon-free power.

But Will It Raise My Electric Bill?

This is the real question, and it deserves a direct answer.

The answer depends on the locality. Senator Elizabeth Warren made headlines this year by claiming that residential electricity bills near data centers had risen “as much as 267%” over five years. PolitiFact rated that claim Mostly False. The 267% figure referred to a cherry-picked wholesale price, not what consumers actually pay, and wholesale costs account for only 30% to 50% of a typical residential bill. The rest is transmission, distribution, and taxes, which do not move in the same way as data center demand.

Nationally, residential electricity prices did rise about 42% over five years, driven in part by higher oil and natural gas prices. In some data-center-heavy regions, the increase was larger: Washington, D.C. saw a 94% rise, and Maryland saw a 74% rise.

But a 2026 study by the energy consultancy E3 found something that cuts against the common narrative: states with the largest data center load growth, including Texas and Virginia, had the smallest electricity rate increases, while states with declining load growth, such as California and New York, had the largest increases.

That is basic economics. Electricity infrastructure is a large sunk capital cost. The first watt is expensive to deliver, but spreading that cost across gigawatts lowers the price per watt. If a state builds a large amount of infrastructure and then demand falls because population and business decline, customers still have to pay for the system. Rising demand can increase costs in the short term, but over time the effect is less straightforward. Inflation is also a major factor, and that is not a data center problem.

In the mid-Atlantic PJM market, E3 found that only about half of recent capacity price increases were attributable to load growth. The rest came from power plant retirements, market design changes, and supply constraints that predate the AI boom. In Virginia, now the data center capital of the world and soon to be challenged by Texas, E3 found no evidence that data centers historically shifted costs onto residential customers. In fact, the study found that individual large data center accounts generated millions of dollars in net surplus revenue for their utilities, helping hold down costs for everyone else.

That does not mean every data center deal is priced correctly or that every state has the right rate design. Some clearly do not, and regulators should keep pushing for better incentives. But the evidence does not support the broad claim that data centers are quietly raising bills for ordinary electricity customers nationwide. The states building the most data centers are, on average, seeing the least rate pain, which is the opposite of the popular story.

The AI Bottom Line

The grid does need to grow much faster than American infrastructure has grown in decades. That is a real and difficult policy challenge. I would rather see that solved through faster permitting, smarter interconnection rules, and more nuclear power than through a moratorium on the industry that is currently the largest source of private capital investment in the country and a key technology for the future.

The claim that data centers are draining aquifers and blowing up electric bills is not supported by the data. On water, data centers are a rounding error compared with golf and agriculture, and the technology is becoming more efficient every year. On power, the bottleneck is real, but it is about poles and permitting rather than an unsolvable physical limit. So far, the regions embracing data centers most aggressively are not the ones suffering the highest price increases.

Next week, in Part 2, I will address the objection that worries people most: jobs. Is AI taking jobs, creating jobs, or both? The headlines are not enough; the data will matter more.

At the end of the letter, I provide sources for some of the data above.

Geopolitics at the Speed of Light

AI is not the only area changing quickly. The pace of change in geopolitics is accelerating as well. A few months ago, Ed D’Agostino and I agreed that Mauldin Economics should add a weekly geopolitics letter that is free for all readers. We were fortunate to attract Jacob Shapiro.

If you are not already reading his weekly letter, The World Isn’t Ending, you should be. I like the title because it matches my own philosophy. Jacob was mentored by my friend George Friedman, first at Stratfor and later at Geopolitical Futures. He now has decades of experience and has become a trusted source for institutions navigating a multipolar world. I suggest giving his letter a try; it is free.

Austin, New York, Washington, D.C., and More

I am writing this morning from Cleveland, where I am with my daughter Abbi. This afternoon we are meeting with a top national neurosurgeon about a growing, hopefully benign tumor in the middle of her brain, in a very unfortunate location. Soon I will have blood drawn, and I am sure Dr. Mike Roizen will explain what it all means when we have dinner tonight with Abbi.

I will admit that I always get a little nervous when I have major blood work done, because the worrywart part of me wonders what they will find. So far, the results have mostly been good. Knock on wood.

Next weekend I fly to Austin to be with Pat Watson as we say a sad farewell to his wife Grace. Two weeks after that, I will be in New York for two nights for a few quick meetings. I will likely go to Washington, D.C. later in September and again in November. I have other tentative trips planned, but they are not yet booked.

As a sidebar, I am very aware of electricity costs because we pay about three times as much in Puerto Rico as you do on average, for what can only be described as intermittent power. When the power company, and Puerto Rico more broadly, went bankrupt, the lawsuits began. Nearly 10 years later, those lawsuits are still preventing a full rebuilding of the grid. The money is there, but the cooperation is not. That is why I have a 50,000-watt Kubota diesel generator in my golf cart garage. It is overkill, but it was the most cost-effective solution.

And with that, I will hit the send button and wish you a great week.

Your struggling to learn to use AI analyst,

John Mauldin
Co-Founder, Mauldin Economics

Sources: EPRI, Powering Intelligence (2026); Mark P. Mills, The Rise of AI: A Reality Check on Energy and Economic Impacts, National Center for Energy Analytics (November 2025); Ramez Naam via Peter Diamandis, Moonshots newsletter (2026); water-use analysis via napkinquest/AKCP data center water footprint research (2026); E3, Understanding the Drivers of Rising Electricity Rates and the Role of Data Centers (2026); PolitiFact fact-check of Sen. Elizabeth Warren (June 2026); Gallup; Pew Research Center.