Empire of AI Overestimated Datacenter Water Usage by 1000x (twitter.com)

A recent reanalysis has found that the “Empire of AI” account/report overstated datacenter water use by roughly three orders of magnitude — about 1,000x. What had been presented as massive, industry-scale consumptive water usage was based on incorrect assumptions and conflated metrics (for example, using peak evaporative cooling capacity or water withdrawal figures as if they were continuous consumptive use). When corrected for real-world cooling architectures, closed-loop systems, water reuse, and the difference between withdrawal and consumptive loss, the implied water footprint for modern hyperscale datacenters falls far lower than the sensational claim. This matters because exaggerated figures shape policy, investor sentiment, sustainability debates, and downstream academic estimates of AI’s environmental cost. The episode highlights key technical points: datacenter water impact depends on cooling type (air, evaporative, closed-loop liquid, or immersion), operational practices (make-up water, blowdown, reuse), and appropriate metrics such as Water Usage Effectiveness (WUE) and consumptive vs. non‑consumptive reporting. For the AI/ML community, the takeaway is to use standardized, transparent measurements when estimating infrastructure environmental costs, and to distinguish withdrawal from true consumption to avoid misleading conclusions about AI’s sustainability footprint.