AI Geopolitics

The Water Gap: US AI Stalls Where China Builds

As of 20 June 2026, both the United States and China face real water scarcity. Only one of them is letting it slow the AI buildout. The difference is not water. It is governance.

By AI Race Facts  ·  20 June 2026  ·  9 min read

In October 2025, OpenAI told the White House that “electrons are the new oil” and asked for 100 GW of new US power capacity per year (our coverage here). Energy is the constraint that gets the headlines. Water is the one quietly stalling projects on the ground — and it exposes a structural gap between how the US and China build.

Two water-scarce superpowers

The instinct is to assume China, with more than a billion people, must be drowning in water and the US must have plenty. Both assumptions are wrong.

China’s problem is geographic mismatch. Roughly 80% of its water sits south of the Yangtze, while the dry north holds about a third of the population, two-thirds of the cropland, and only around 14% of the supply, per Stanford’s SPICE program. Beijing’s per-capita water sits below 100 m³ — under one-fifth of the 500 m³ line for acute shortage. Northern aquifers are draining fast: Hebei’s deep aquifer drops about three meters a year, and the Yellow River has run dry in stretches in most years since 1985. ● CHECKED-PRIMARY

The US is no oasis. As of early May 2026, roughly half the country was in drought, with the West worst hit — the same arid states where cheap land and power are pulling data centers in. Neither country has abundant water where the AI buildout wants to go. ● CHECKED-SECONDARY

What a data center drinks

Servers throw off heat, and the cheapest way to remove it is evaporating water. A typical data center uses about 300,000 gallons a day — roughly 1,000 households’ worth — while a large one can use up to 5 million gallons a day, equal to a town of 50,000 people, per the Brookings Institution and the Lincoln Institute of Land Policy. That is only the direct use; more water is consumed off-site generating the electricity and manufacturing the chips, as the Environmental Law Institute notes — and that indirect footprint is the larger share. ● CHECKED-PRIMARY

At national scale, US data centers directly consumed about 17.4 billion gallons of water in 2023, a figure federal researchers project could climb to between 38 and 73 billion gallons by 2028, per analysis citing the EPA and a 2024 Lawrence Berkeley National Laboratory report. ● CHECKED-SECONDARY

Projection, not a measured figure. Source: EPA / LBNL via MOST Policy Initiative. ● PROJECTION

Where the US gets stuck

The US constraint is not the reservoir. It is the permit. In 2025, at least 48 data center projects representing $156 billion in investment were blocked or stalled by local opposition, according to the Data Center Watch tracker, reported by Fortune. Cancellations rose from six in 2024 to 25 in 2025, and water ranks among the most common grievances, cited in more than 40% of contested projects. A 2026 Gallup survey found 71% of Americans would oppose a data center in their community. ● CHECKED-SECONDARY

The grievances are concrete, not abstract. In Newton County, Georgia, Meta’s $750 million data center uses about 500,000 gallons of water a day — roughly 10% of the entire county’s use, per the county water authority in a New York Times investigation. Local water rates are set to rise 33% over two years, and in May 2026 Rep. Alexandria Ocasio-Cortez held up jars of discolored water from nearby Morgan County at a congressional hearing. Meta commissioned a well study and said it was “unlikely” its campus affected the supply, citing a separate watershed. The dispute, not the verdict, is the point: water turns siting into a fight. ● CHECKED-PRIMARY

The deeper issue is structure. The US runs on more than 50,000 local water systems, each a potential veto point, with no federal authority that can route water to a site or override a county that says no. That fragmentation is a feature of the system, not a bug — but it makes the buildout slow and uncertain.

Where China just builds

China faces the same scarcity and builds through it. A 2025 facility-level study in Applied Energy found that about 72% of the country’s computing capacity sits in severely water-scarce regions, with the sector’s total water footprint reaching roughly 15.7 billion m³ in 2022 — about 2.7% of national water use, per the peer-reviewed analysis. That figure counts both direct cooling and indirect electricity-generation water, so it is not comparable to the US direct-only number above; the takeaway is the siting, not the size. The state put the compute there on purpose. ● CHECKED-PRIMARY

Two policies do the work. The first is “Eastern Data, Western Compute,” a national plan that steers data centers to cold, cheap inland provinces — Ningxia, Gansu, Inner Mongolia — where sub-zero winters allow year-round free-air cooling and facilities can hit efficiency (PUE) near 1.2 without water-hungry cooling towers, per trade reporting and market analysis. Hohhot can air-cool for up to eight months a year. Beijing caps new sites at 2.5 liters of water per kWh, and the central planner reimburses up to 30% of capital costs for centers that recycle process water. ● CHECKED-SECONDARY

The second is raw water logistics. The South-North Water Transfer Project — the largest water-diversion scheme on Earth, more than $70 billion spent and at least 330,000 people relocated — is designed to move about 44.8 billion m³ of water a year from the wet south toward the dry north when its routes are complete, and has become a major source of Beijing’s tap water (estimates of its share range from roughly a third to about 70%). When the state can move rivers, siting compute on scarce water becomes a planning problem, not a veto fight. ● CHECKED-SECONDARY

The real gap is governance

Put the two systems side by side and the lesson is not that China has water and the US does not. It is that China can command its water, and the US cannot. One system routes scarce water to strategic compute by decree; the other negotiates it county by county, and often loses.

Water capacity = AI capacity = national power.

It is the same shape as the electron gap: a resource that is physically scarce on both sides, where the advantage goes to whoever can mobilize it fastest. Energy and water planning are now inseparable from AI strategy, and on both, the US system trades speed for local consent.

The honest caveats

China’s position is not as strong as the contrast implies. Experts have called the South-North Water Transfer unsustainable — it treats symptoms, the south is increasingly stressed itself, and much of China’s supply depends on shrinking Himalayan glaciers. Centralized control buys speed and certainty, not abundance.

Nor is US opposition irrational. Residents are reacting to real rate hikes, real water-quality scares, and tax breaks that shift costs onto them. “Slow” is partly the price of a system that lets people object.

And water may not be the binding constraint yet. Moody’s identifies semiconductor shortages — not local opposition — as the more acute near-term limit on the buildout. As with energy, water is best read as a ceiling a few years out: the gap that widens if nothing changes, not the wall everyone is hitting today.

Last updated 20 June 2026. Corrections, if any, will be logged here with the date and what changed.