Drive through almost any Sun Belt suburb at 3 a.m. and you will see it: rotors spinning across an 18-hole layout, soaking fairways that will be mowed, sprayed, and soaked again tomorrow. Nobody is protesting that sprinkler head. Meanwhile, a mile away, a county board is voting down a data center campus because residents are convinced it will drain the aquifer dry. This irony around the water panic around data centers is driving votes, but it is often disconnected from the scale of actual water use.

What Data Centers Actually Drink

U.S. data centers consumed roughly 228 billion gallons of water in 2023 (about 626 million gallons per day), according to the Lawrence Berkeley National Laboratory (LBNL). Most of that is indirect: only 17.4 billion gallons went to on-site cooling, while roughly 211 billion is attributable to the water consumed generating the electricity they run on.  

Either way, data center water use remains small relative to the roughly 322 billion gallons per day of total U.S. withdrawals, according to the USGS. The data center share is estimated at roughly 0.015% to 0.2% of withdrawals, versus agriculture’s 43% and thermoelectric power’s approximately 42%.

The comparison gets more intuitive at an annual scale. The EPA WaterSense program estimates U.S. households use close to 9 billion gallons per day for landscape irrigation, roughly 3.3 trillion gallons a year. U.S. golf facilities use approximately 759 billion gallons annually for irrigation, while California almond orchards use roughly 1.3 to 1.8 trillion gallons a year.

The point is not that water is irrelevant. It is that a national narrative built around data center water usage ignores the much larger, longstanding uses that are rarely treated as existential threats.

Fake News Goes Viral

Much of the panic traces to a viral claim that every ChatGPT question consumes a full bottle of water. That claim is a misreading of a 2023 UC Riverside study, which estimated that a 500 milliliter bottle of water could be consumed across roughly 10 to 50 medium-length GPT-3 responses, depending on when and where the system operated

Where basin level water stress is real, the hyperscalers are increasingly changing their designs. Microsoft says it cut data center water use by 25% from its 2022 baseline, introduced zero water cooling designs for new AI optimized facilities beginning in 2024, and replenished more water than it withdrew globally in fiscal year 2025. Google says it replenished more than 7 billion gallons in 2025. It also shifted the first building at its Botetourt County, Virginia campus to air cooling after local pushback and has described a Texas site designed to use water primarily for kitchens. These changes are not merely public relations initiatives: they are practical responses to permitting, community trust, and local resource constraints. For systems where water is still required due to the extreme density of chips, new closed-loop water solutions are attempting to nearly eliminate water consumption.

Foreign Influence Driving Misinformation

Researchers at Alethea identified roughly 700 references to U.S. data centers in Russian, Chinese, and Iranian state media over a six-month period. The content frequently connected AI infrastructure to water scarcity, higher electricity bills, environmental damage, and economic insecurity, seeking to turn legitimate local concerns into a broader indictment of the U.S. AI buildout. Chinese state outlets including CGTN and the Global Times amplified the claim that the AI boom was causing American electricity prices to rise, while Russian messaging portrayed local resistance as evidence that the industry was an unsustainable bubble.

The more troubling finding was the use of deceptive content intended to resemble local reporting or authentic grassroots activism. Storm-1516, a Kremlin-linked influence operation, used fake domains designed to resemble TechCrunch and Gizmodo to distribute false claims about a data center in Armenia. Alethea also anonymously ran Facebook groups that presented themselves as local American communities, including pages with names such as “Life in Oklahoma” and “Only in Montana,” but were operated from Bangladesh and shared AI-generated anti-data-center material.

Water should not be used as a universal veto on AI infrastructure. It is one variable in a site-specific question, and one that should be assessed alongside power availability, land, cooling architecture, wastewater options, and permitting timelines. The winning operators of data centers will publish credible, site specific water metrics early, use lower water designs where conditions demand them, and distinguish real community concerns from claims that collapse under basic comparison against other use cases of water (i.e. agriculture, lawns, almonds, and golf courses).

Takeaway: The U.S. data center buildout will be constrained first by power, land, and permitting, not national water scarcity. Water still matters profoundly at the local level, but the public debate needs proportion and to be sensitive to our adversaries attempting to slow us down through misinformation campaigns (which have proven to be highly effective in the past). Lawns, almonds, and golf courses (just to name a few) are the much larger water users. The most durable data center projects will pair thoughtful cooling and water design with fast, transparent, evidence based community engagement.