The era of building AI infrastructure without environmental accounting just ended. A peer-reviewed study published today quantifies what's been theoretically known but operationally ignored: AI data centers consumed between 32.6 and 79.7 million tons of carbon this year—roughly equivalent to New York City's entire annual emissions—while guzzling 312.5 to 764.6 billion liters of water. For the first time, enterprises have hard numbers to plug into cost models. The research shifts environmental impact from rhetorical concern to capital allocation variable, forcing a reckoning across three constituencies: companies building infrastructure, investors pricing risk, and regulators drafting policy.

The numbers hit different when they're in the peer-reviewed literature. Today's study from Alex de Vries-Gao at VU Amsterdam's Institute for Environmental Studies quantifies something the industry has been dancing around for years: AI isn't just power-hungry in the abstract. It's consuming resources at city-scale. Between 32.6 and 79.7 million tons of carbon annually. Up to 764.6 billion liters of water. These aren't estimates. They're built on earnings call data, analyst hardware projections, and the power demand research showing AI hitting 23GW globally this year—already surpassing Bitcoin mining's electricity consumption.

The inflection point isn't the environmental impact itself. It's the moment impact becomes quantifiable enough to change buying behavior.

Here's what shifts right now: Enterprises planning data center deployments can no longer hide behind "sustainability reports that don't break out AI specifically." The research gives them the conversion formula. Want to build a training cluster in Phoenix? You're looking at roughly 1.2 to 2.9 tons of water per megawatt-hour depending on the local power grid's fuel mix. That number plugs directly into total cost of ownership calculations. It changes which regions make financial sense. It changes the math for on-premises versus cloud vendor lock-in. For companies evaluating 2026 capex, this research just became a required input.

Microsoft, Google, Meta, and Amazon have been transparent about total data center emissions and water use in annual reports—but they've deliberately avoided breaking down what AI specifically consumes. De Vries-Gao worked around that opacity. He used hardware shipment estimates, power consumption models, and publicly available information to reverse-engineer the numbers. The study essentially proves that if tech companies wanted to be transparent, they could be. They're choosing not to, which itself is now data: opacity signals cost avoidance.

The water angle deserves specific attention. Data center cooling is where the bulk of water use concentrates. A single large training run can require water volumes that rival small city consumption. The US, hosting more data centers than any other country, is already seeing surge in local opposition driven explicitly by water concerns. That opposition exists today; now it has peer-reviewed justification. The first major project denial—and there will be one—happens when a municipality cites this research in a zoning decision. That's the regulatory inflection moment.

Investors are watching this sequence play out. Data center operators face a simple calculation: current cost of environmental impact is zero in their P&L. Future cost, post-regulation, is massive. The companies building with water efficiency and renewable power NOW are positioning themselves as the lower-risk infrastructure plays. Companies betting on continued opacity are taking on regulatory tail risk that's now quantifiable.

The study's most important finding might be its own limitation. De Vries-Gao notes his analysis captures only the operating emissions of AI equipment—not the supply chain impact, manufacturing footprint, or end-of-life recycling of chips. The real environmental cost is higher. Possibly significantly higher. Which means every number in this study is a floor, not a ceiling. The actual impact is larger than what regulators will see in their first policy drafts.

For teams building AI products: the constraint you'll hit isn't compute availability or model capability anymore. It's where you're allowed to put the compute. Regions with water scarcity, aggressive renewable mandates, or anti-data-center political momentum are becoming infrastructure deserts. The geographic distribution of AI workloads is about to become as strategic as chip availability.

The 2023 study projecting 600 billion liters of water use by 2027—de Vries-Gao's 2025 numbers are already in that ballpark. Which means the trajectory is baking in even before regulatory requirements land. Market forces alone are starting to constrain expansion. When policy arrives (and it will, probably within 18 months), it's going to codify what the market is already forcing.

What's not in this story but matters: the companies that built environmental monitoring into their infrastructure decisions six months ago now have competitive advantage. They know their actual impact. They can model scenarios. They can position for the post-transparency era. Companies still operating on the assumption that this stays invisible are about to have a very bad quarterly earnings call.

The environmental impact of AI infrastructure just stopped being hypothetical. With peer-reviewed data quantifying carbon at NYC-scale and water at global-bottled-consumption-scale, decision-makers face a new variable: where you build matters as much as what you build. For enterprises, the window to factor environmental costs into infrastructure planning—before regulatory requirements force it—closes in Q1 2026. For investors, data center operators without water efficiency and renewable power strategies face emerging tail risk. For builders, geographic constraints on compute location are arriving faster than most teams realize. Monitor the next regulatory announcement or major project denial; that's the inflection point where environmental impact becomes policy. Until then, every infrastructure decision made without accounting for this research is bet-on-opacity going to zero.