AI’s Impact on Water and Energy. What’s Really Going On?

Most coverage of AI’s environmental impact gets two things wrong.

It treats water and power as the same problem. And it frames the choice as clean energy versus fossil fuels.

Neither is right. And the gap between the public conversation and the actual situation is wide enough to matter — because the decisions being made right now, by hyperscalers, utilities, and regulators, are going to shape the next decade of American infrastructure.

Let me lay out what is actually happening, point by point.

Water Isn’t an AI Problem. It’s a Where-You-Build-It Problem.

Start with the numbers. In the United States, thermoelectric power plant cooling accounts for roughly 41% of total water withdrawals, and irrigation for agriculture accounts for another 37%.[1] Together, those two activities use nearly four out of every five gallons of water consumed in this country.

Data centers, by comparison, consumed approximately 66 billion liters of water directly in 2023.[2] That works out to less than one-tenth of one percent of total U.S. water withdrawals. American golf courses, collectively, use roughly 2 billion gallons of water per day for irrigation — meaningfully more water than every data center in the country combined.[3]

So the headline framing — that AI is draining the planet’s water — is wrong at the national level. In aggregate, data center water use is a rounding error against the scale of agricultural irrigation and thermoelectric power generation.

But that is the wrong scale at which to evaluate the problem.

Drop a hyperscale data center in a small town in Arizona, and suddenly you are competing with someone’s sprinkler. Roughly two-thirds of new hyperscale campuses built since 2022 have been sited in counties already facing high water stress, and about one in five U.S. servers now sits in a water-stressed watershed.[4] The local impact is real, painful, and concentrated — even when the national impact is negligible.

The fix is not “use AI less.” It is “build them where water isn’t a constraint.”

This is a siting decision, not a technology decision. And it is one that can be made deliberately, by companies and by the regulators approving these facilities. The water question, properly framed, is solvable. It is a planning problem dressed up as an environmental crisis.

The Power Problem Is Real, and It’s Not Really About the Power Itself.

The energy story is different. The energy story is real, and it is significantly more serious than most coverage suggests — but for a different reason than people assume.

U.S. data center electricity demand is projected to roughly double by 2030, with the International Energy Agency estimating that consumption in the United States will rise by approximately 240 terawatt-hours over 2024 levels — a 130% increase in just six years.[5] By the end of the decade, data centers are expected to account for nearly half of all U.S. electricity demand growth, and to consume more electricity than the production of aluminum, steel, cement, chemicals, and all other energy-intensive goods combined.[6]

These are not speculative numbers. The capital is committed. The chips are ordered. The sites are selected.

But the binding constraint is not generation capacity. It is not chips. It is not capital.

It is the grid itself.

Right now, roughly 2,600 gigawatts of new power generation and storage capacity is built or ready to build but unable to connect to the U.S. grid — more than double the capacity of the entire existing American power system, which sits at about 1,280 gigawatts.[7] The median time to bring a new power plant online has roughly tripled, from under two years in 2008 to nearly five years for projects completed in recent years.[8]

The chips are ready. The money is ready. The wires aren’t.

This is why hyperscalers have started buying power plants directly. Microsoft signed a twenty-year, $16 billion deal to purchase the entire 835-megawatt output of a restarted reactor at Three Mile Island.[9] Amazon has acquired a nuclear-adjacent data center campus in Pennsylvania and signed offtake agreements for plants in Illinois. Meta signed a 1.1-gigawatt deal with Constellation. The pattern is unmistakable. The reason is not that these companies prefer to be in the power business. The reason is that the public grid cannot deliver power on the timeline AI requires.

The chips are ready. The money is ready. The wires aren’t.

Solar and Wind Aren’t the Answer on Their Own.

This is where the conversation gets politically uncomfortable, and where most coverage refuses to engage honestly.

Solar and wind are the cheapest sources of new electricity in most of the United States. We are building them faster than anything else, and that should continue. The IEA projects that renewables will meet roughly half of the global growth in data center demand by 2035.[10]

But data centers do not run during daylight hours. They run all twenty-four. The sun goes down. The wind stops. Even if we doubled solar capacity tomorrow, it would not help at 2 a.m. Until grid-scale storage technology catches up — and despite real progress, it is not close to the scale required — renewables cannot carry baseload power on their own.

This is a physics problem, not a political one. And pretending otherwise has consequences. It means we make policy on a fantasy of what the energy mix could look like, while the actual buildout proceeds in a different direction entirely.

Whether You Like It or Not, the Only Near-Term Answer Is Natural Gas.

Few in the industry will say this out loud, but everyone building knows it.

Active natural gas capacity in the U.S. interconnection queue increased by 72% in 2024 alone, with new gas interconnection applications breaking annual records by mid-2025.[11] Meta, Microsoft, and most of the other hyperscalers are signing gas deals, even those with public commitments to clean energy. The reason is simple: gas plants can be built and brought online in two to three years. New nuclear cannot. Grid-scale renewables paired with sufficient storage cannot. There is no other firm-power option that fits the timeline.

The IEA estimates that natural gas generation will expand by 175 terawatt-hours globally to meet data center demand, with most of that growth occurring in the United States.[12]

The clean-energy transition and the AI transition are running on different clocks. The AI clock is faster. For the next five to seven years, the marginal megawatt powering AI growth will be a gas megawatt. This is not a value judgment. It is what is happening.

Acknowledging this does not require giving up on decarbonization. It requires being honest about the bridge.

Over the Next Decade, Nuclear Quietly Comes Back.

The medium-term picture is more interesting, and more hopeful, than most people realize.

Three Mile Island Unit 1 is being restarted to power Microsoft data centers, with operations expected to begin in 2027 or 2028.[13] Palisades is being restarted in Michigan. Small modular reactor designs are moving from concept to construction, with the first units expected online around 2030.[14] Amazon, Google, and Microsoft have signed long-term power purchase agreements that effectively underwrite the next generation of firm clean power. Enhanced geothermal — the most promising new firm-clean technology — is moving from pilot to commercial scale, financed in part by hyperscaler offtake deals.

What is happening here is significant. The companies building AI are becoming the companies financing American energy infrastructure. That is a structural shift, and it is happening faster than most observers have registered.

For the first time in forty years, there are large, creditworthy buyers committed to twenty-year offtake contracts for firm clean power. That is the financing mechanism that makes new energy infrastructure get built. Utilities and merchant generators were not going to build this on speculation. The AI companies are.

The Bigger Opportunity Most People Are Missing

So the real story is more complicated than a battle between AI and the planet.

AI is forcing a long overdue conversation about American grid infrastructure. The hyperscalers are not waiting for permitting reform or transmission upgrades. They are writing the checks, signing the power purchase agreements, and building a parallel energy system in real time.

That works for them. But it raises a deeper question for the rest of us.

The United States has not built serious energy infrastructure in forty years. Transmission lines take ten to fifteen years to permit and construct. New nuclear licensing is measured in years and hundreds of millions of dollars per application. The interconnection process for new generation has tripled in duration since 2008. Meanwhile, the rest of the economy — manufacturing reshoring, electric vehicles, electrification of heating — depends on the same grid the AI buildout is straining.

AI is the first force big enough — and well-capitalized enough — to actually change that.

Done right, the energy investment driven by AI does not just power AI. It modernizes the grid for everything else. New transmission, new firm clean generation, new permitting frameworks, new partnerships between hyperscalers and utilities. The benefit accrues to the broader economy.

Done wrong, we end up with a two-tier system: hyperscalers with their own dedicated power infrastructure, and everyone else still queueing for the public grid. That is a worse outcome than the status quo, because it spends the political capital of this moment on a narrow private benefit rather than a broad public one.

A Leadership Choice

None of this resolves itself. The decisions being made now — about siting, about permitting reform, about which firm-power technologies get financed at scale, about how data centers participate in grid management — will shape the next decade of American infrastructure.

The companies that figure out how to secure power are going to win the next decade of AI. That is no longer an IT question. It is a board-level question. And it is one that intersects with public policy in ways most corporate leaders have not yet absorbed.

The framing of AI versus the planet is too small. The actual question is whether the United States uses this moment of capital, urgency, and political alignment to build the grid the next century requires — or whether we build a parallel system for a few companies and leave everyone else behind.

That is the choice. And it is being made right now, whether we engage with it or not.

I welcome you to join me in this conversation.

The track we lay now determines where it leads.

References

1. U.S. Geological Survey, “Estimated Use of Water in the United States in 2015,” reporting total U.S. water withdrawals of approximately 322 billion gallons per day, with thermoelectric power cooling at 41% and irrigation at 37%. usgs.gov.
2. Apstech Advisors, “Data Center Water Consumption in the US: Challenges, Trends, and Market Opportunities for 2025–2030” (January 2026). apstechadvisors.com.
3. AKCP, “The Truth About the Water Footprint of Data Centers” (September 2025), citing collective U.S. golf course water use at approximately 2 billion gallons daily, or 0.5% of national withdrawals. akcp.com.
4. Electric Choice, “U.S. Data Center Power Consumption Map by State (2026),” citing analysis that two-thirds of new hyperscale campuses built since 2022 are in high water-stress counties. electricchoice.com.
5. International Energy Agency, “Energy and AI” report, April 2025. U.S. data center consumption projected to grow by approximately 240 TWh (up 130%) from 2024 to 2030. iea.org.
6. International Energy Agency, “Energy and AI” executive summary: “In the United States, data centres account for nearly half of electricity demand growth between now and 2030.” iea.org.
7. Lawrence Berkeley National Laboratory, “Queued Up” interconnection queue analysis. U.S. interconnection queue reached approximately 2,600 GW of proposed generation and storage as of late 2024, more than double the country’s installed capacity of roughly 1,280 GW. emp.lbl.gov/queues.
8. Lawrence Berkeley National Laboratory, “Queued Up: 2025 Edition.” Median project duration from interconnection request to commercial operation has more than doubled, from under 2 years for projects built in 2000–2007 to over 4 years for those built in 2018–2024, with current wait times approaching 5 years. emp.lbl.gov/queues.
9. Constellation Energy, “Constellation to Launch Crane Clean Energy Center” (September 2024). Twenty-year power purchase agreement with Microsoft for approximately 835 MW from the restarted Three Mile Island Unit 1. Project cost estimated at $1.6 billion. sec.gov.
10. International Energy Agency, “Energy and AI” report: renewables projected to generate over 450 TWh of additional power to meet data center demand by 2035, representing approximately half of global growth. iea.org.
11. Lawrence Berkeley National Laboratory, “Queued Up: 2025 Edition,” reporting active natural gas capacity in U.S. interconnection queues increased 72% year-over-year in 2024. Wood Mackenzie reports new gas interconnection applications broke annual records by mid-2025. emp.lbl.gov/queues.
12. International Energy Agency, “Energy and AI” report: natural gas generation projected to expand by 175 TWh globally to meet data center demand, with most growth concentrated in the United States. iea.org.
13. U.S. Department of Energy Loan Programs Office authorized a $1 billion loan to Constellation in November 2025 to support the Three Mile Island restart, with operations expected in 2027 or 2028. techcrunch.com.
14. International Energy Agency, “Energy and AI” report: “The first small modular reactors come online around 2030.” iea.org.