Why is Everyone So Wrong About AI Water Use??

Main Argument

The video argues that AI “water use” numbers are easy to mislead because they depend on messy, subjective accounting choices and because different parts of the AI system—such as queries, model training, and the electricity that powers data centers—use different kinds of water.

Core Claims and Comparisons

• Small per-query figures can be “technically true” yet misleading. The video discusses Sam Altman’s claim that an average ChatGPT-style query uses about 0.000085 gallons (approximately 1/15 of a teaspoon). The presenter agrees that “small per-query” and “large annual” claims can both be true, but argues the framing can be dishonest: per-query estimates often count only the water used during the final inference step.

• Annual AI data-center water projections can look enormous depending on what’s included. A Morgan Stanley projection (as relayed in the subtitles) is described as reaching roughly 1 trillion liters per year by 2028, about an ~11× increase versus 2024. The presenter argues this can’t be derived just by multiplying the tiny per-query number, implying that these larger projections likely include additional lifecycle components.

• The main reason the debate is unreliable: “resource use analysis” is hard. Water/electricity accounting requires decisions about system boundaries, assumptions, and allocation rules—so numbers can be made to look small or large by including or excluding certain processes.

What’s Missing from “Per Query” Water Accounting

• Inference isn’t the only thing using water—model training can be the biggest omission. The presenter argues that the largest gap in small per-query figures is model training. Training runs for weeks/months on large GPU clusters, requiring substantial electricity and water for cooling. Since conversations depend on training, the “water cost” should be allocated across queries. However, OpenAI allegedly does not provide inspectable source data, making transparent attribution difficult.

• Even “one query” can involve many behind-the-scenes steps. The video notes that complex questions can trigger internal “reasoning,” follow-ups, and sometimes internet searches—so “one user query” may translate into multiple backend queries. This could increase the true inference footprint beyond a simplistic “one query = one cost” view, though the presenter emphasizes that lifecycle omission (especially training) is the more important distortion.

What Inflates Some “Big Water” Claims

• Electricity-linked water use from thermoelectric power plants. Some high AI-water projections are described as attributing water used by power plants to AI, including water withdrawn and then returned to rivers/lakes/oceans after condensing steam. The subtitles cite U.S. electricity generation (via USGS) as accounting for about 40% of freshwater withdrawals.

• But the type of water matters. Power-plant water use differs from municipal drinking water systems: it’s often surface water intake/return rather than treated potable water. There may also be relatively small evaporative losses (subtitles suggest ~2–3%). Still, impacts like thermal pollution (waste heat) are treated as real and regulated concerns.

• Methodological flaws exist on both sides—but water impacts aren’t dismissed. The presenter emphasizes that water constraints are local and depend on watersheds, temperature/flow stress, and allocation limits—not just total gallons.

“Hidden” Lifecycle Piece: Chips and Materials

• Chip manufacturing uses a different, harder-to-produce water source. The video claims AI hardware production (e.g., GPU-related processes) requires large quantities of ultra-pure water, which is more difficult and energy-intensive than normal drinking water. It may be smaller in volume terms than cooling, but could be more impactful due to scarcity and production complexity.

The “What About Corn?” Argument (Extreme Comparison)

• The presenter claims AI data centers’ global water use is vastly smaller than irrigated corn in the U.S. (subtitles list: ~20 trillion gallons/year for U.S. corn production vs ~260 billion gallons for global AI data centers).
• The video further claims corn is mostly used for livestock feed and especially ethanol for transportation, with only a small fraction directly consumed by humans.
• The point is not that agriculture is “fine,” but that AI water debates can mislead when they ignore the scale and context of existing water-intensive industries (including other uses like lawn irrigation described as municipal).

Policy/Environment Focus: Power Demand as the Bigger Issue

• While water is treated as salient and potentially serious in some regions, the presenter argues that increased power demand from AI is likely the larger threat—affecting carbon budgets and household/economic costs more broadly.
• The presenter also expresses skepticism that projected AI buildouts will necessarily happen on the assumed timeline, warning about the risk of overinvestment.

Overall Conclusion

• The video’s main conclusion is that AI water-use debates are structurally easy to manipulate because lifecycle boundaries, attribution methods, and water categories vary.
• Even if per-query numbers and annual projections can both be “true,” they answer different questions—and without consistent accounting, public discussion becomes misleading.
• The presenter emphasizes that water is locally constrained and hard to substitute, but also worries more about electricity growth and broader economic/political risk.

Presenters/Contributors

• Sam Altman (referenced)
• Hank (the video’s narrator/presenter; “Hank from the future” appears as a self-reference in the subtitles)