By 2028, AI data centers are projected to consume roughly 1.068 trillion liters of water annually — an 11-fold increase from 2024 levels, according to Morgan Stanley research. That number is so large it stops feeling real. Every time you ask ChatGPT to rewrite your cover letter, you’re one tiny drop in an ocean that’s filling up fast.
The Numbers Behind the Hype
Direct cooling is only a fraction of AI’s water story — the bigger cost is hiding inside your electricity bill.
That 0.000085-gallon figure, cited in Altman’s 2025 blog post, covers only direct cooling. It ignores the water used to generate the electricity powering the server. A CBS News synthesis puts indirect water use at roughly 92.54% of AI’s total footprint. Independent researchers found 10–50 GPT-3 queries could consume around 500 mL when the full lifecycle is counted — a meaningfully different picture than one-fifteenth of a teaspoon.
The scale, in numbers:
Generative AI infrastructure compute power is expected to expand 8.5× between 2024 and 2028 (Morgan Stanley)
- U.S. data centers currently account for roughly 4.4% of electricity consumption; projections point to around 12% by 2028
- AI electricity use could hit 165–326 TWh per year by 2028 — enough to power roughly 22% of U.S. households, per MIT energy research
“We now expect AI data centers to drive annual water consumption…to approximately 1,068 billion liters by 2028 — an increase of 11 times from 2024 estimates.” — Morgan Stanley researchers
More than half of the world’s top data center hubs already sit in “medium” or higher water-risk zones. Meanwhile, 68% of data centers are located near protected areas or Key Biodiversity Areas, according to a UK government evidence review. The aggregate number is staggering. The local concentration makes it tangible.
Data centers cluster in Phoenix, Northern Virginia, and Singapore — regions already managing drought and ecological stress. There’s also the Jevons paradox problem. As AI gets more efficient per query, total usage climbs because more people use it more often. The UN flags this explicitly in its AI electricity projections through 2030. Efficiency alone won’t close the gap.
What the Industry Is Promising
Microsoft and Google have made “water positive” pledges — but auditable outcomes are still catching up to the commitments.
Microsoft has committed to a 40% improvement in water-use intensity by 2030. Google has issued a similar water-positive pledge. Closed-loop cooling systems can reduce data center water use by 50–70%, according to the World Economic Forum — though they come with engineering headaches like bacterial growth and corrosion that require ongoing treatment.
Roughly 74% of U.S. consumers are at least somewhat worried about AI’s environmental impact, per Exploding Topics survey data. Yet 27% say they don’t have enough information to make sustainable choices about the AI tools they use daily, according to Kantar. Sixty-three percent of current AI users say they’d choose products with strong sustainability credentials if those options existed.
That gap — between consumer appetite for transparency and actual product labeling — is where regulatory pressure is building. Per-query water and carbon labels for AI services, modeled on appliance energy ratings, are a logical next step. Whether the industry moves first or waits to be pushed is the question worth watching.
