Two pieces of analysis landed this week that, taken together, undercut the most comforting story the AI industry tells about its own resource use. The first is a wave of follow-on commentary on last week's United Nations University report, much of it built around a Conversation essay that introduces a piece of nineteenth-century economics: the Jevons paradox. The second is a Guardian analysis, picked up by Tom's Hardware, finding that 517 of 809 planned US data centers are slated for areas that have been in drought over the past year.
Start with Jevons. William Stanley Jevons noticed in 1865 that as steam engines got more efficient, Britain burned more coal, not less, because cheaper energy unlocked new uses. The UNU report applies the same logic to AI: every gain in tokens-per-watt makes inference cheaper, cheaper inference invites new product surfaces, and total demand climbs. Researchers quoted in the new coverage are blunt that "AI models will need less in the future" is the trap, not the rebuttal. The report's headline number, that AI could consume three percent of global electricity by 2030 and pull more water for cooling than the world drinks each year, assumes efficiency gains have already been priced in.
That makes the siting analysis matter more than it would in isolation. The Guardian's count of 809 planned US data centers, with 64 percent of them on land NOAA classifies as drought-stricken in the last twelve months, is not just a story about cooling towers. The Tom's Hardware writeup unpacks a useful breakdown borrowed from a January report by Xylem and Global Water Intelligence: by 2050, direct data center cooling will account for only about 4 percent of the additional water AI demands. Power generation eats 54 percent, and semiconductor fabrication another 42 percent. The drought map gets bigger once you include the fabs feeding the chips and the thermal plants feeding the fabs.
TSMC's three Phoenix fabs, once complete, are projected to draw a combined 16.4 million gallons a day in the fourth-driest state in the country. The company reclaims roughly 85 percent of that on site, with a target of 90, but the input still comes out of an aquifer that is already short. A fab and the data center it supplies can sit on top of the same groundwater and only one of them will show up in the industry's "cooling efficiency" charts.
The contrast between these two analyses is what makes them worth reading together. Operators routinely cite improving power usage effectiveness and direct-to-chip liquid cooling, including Nvidia's claim that its GB200 NVL72 system is up to 300 times more water-efficient than air cooling. Those numbers are real, and they apply to a sliver of the total footprint. The Jevons argument is that even a sliver of efficiency, multiplied by the volume of new uses, ends with the absolute number going up. The drought-zone build-out is what that abstraction looks like on a map.
The UNU report also points to a second tension that the data-center map quietly confirms. Only 32 nations host AI-specific cloud infrastructure, and roughly 90 percent of that capacity sits in the United States and China. The countries paying the environmental costs of mineral extraction, manufacturing, and electronic waste are very often not the countries hosting the workloads. The drought zones in the American Southwest are one version of that bill. The lithium pits and tailings ponds in lower-income nations are another.
What the UN authors are asking for is not a halt. It is what they call "efficiency by design," paired with transparency on water and land, lifecycle responsibility for hardware, and international coordination on siting and reporting. The Guardian's map and the Jevons trap suggest those guardrails will need to land soon, because the build-out is already deciding the answer on the ground.