Sam Altman Claims AI Doesn't Use That Much Water. Does the Math Hold Up?
When OpenAI CEO Sam Altman sat down on Alex Heath’s Sources podcast, he offered a vivid analogy to quiet the growing anxiety around AI’s environmental impact: asking ChatGPT 38,000 questions consumes about the same amount of water as growing a single California almond.
By framing individual AI prompts as practically weightless, Altman attempted to push back against viral online claims that a few minutes of chatting with an AI consumes gallons of fresh water. Modern hyperscale data centers, he argued, have largely phased out thirsty evaporative cooling towers in favor of closed-loop systems. In his view, today’s facilities draw no more water than a standard corporate office building, where usage is mainly confined to breakroom sinks and restrooms.
However, a closer look at both the back-of-the-envelope math and the realities of grid-level energy usage reveals why researchers, policy analysts, and lawmakers remain unconvinced.
Running the Numbers
Altman himself admitted on air that he was pulling figures from memory, and the arithmetic bears that out.
Agriculture industry standards estimate that producing one California almond requires roughly 1.1 gallons (4,164 milliliters) of water. Based on OpenAI’s previously cited baseline of roughly 0.32 milliliters per query, 1.1 gallons of water would cover roughly 13,000 queries—not 38,000. For Altman’s 38,000-query benchmark to work, an individual prompt would have to draw a negligible 0.11 milliliters, or less than two drops of water.
Metric / Source
Per-Query Consumption
Equivalent Queries per Almond (1.1 gal)
Methodology & Scope
Sam Altman (Podcast Quote)
~0.11 mL
38,000
Stated on air from memory; assumes minimal closed-loop draw.
OpenAI Direct Baseline
~0.32 mL
~13,000
Direct on-site cooling draw for standard text inference.
Independent Research (UC Riverside)
~10.0 mL
~416
Full lifecycle footprint, including power grid and supply chain cooling.
Why the "Office Building" Narrative Misses the Mark
Focusing strictly on on-site plumbing misses the broader environmental mechanics of modern computing. While an individual text output requires very little direct cooling, multiplying that unit footprint across billions of global interactions turns tiny fractions into significant municipal loads.
- The Off-Site Shift: Phasing out evaporative cooling towers at the server rack doesn't eliminate water use; it often shifts the burden onto the electrical grid. Running high-powered mechanical chillers demands massive amounts of electricity, which in turn consumes water at regional power plants.
- Geographic Strains: Unlike farm acreage with historical water rights, data centers are often built near major suburban populations or in arid regions where municipal water grids are already stretched thin during peak summer months.
- The Compute Expansion: Basic text prompts are no longer the primary standard. As AI shifts toward autonomous agents, multi-step reasoning, and image generation, the compute—and energy—required per task grows exponentially, stretching early per-query baselines.
The Push for Verification
The debate highlights a fundamental issue in tech infrastructure: transparency. Because major cloud providers treat facility metrics and energy arrangements as guarded proprietary data, independent researchers must rely on indirect models.
That lack of clear insight is driving legislative efforts. In states like California, proposed bills aim to require data center operators to publicly report their exact water and power consumption. For industry observers and local communities, moving past casual podcast analogies to verified, audited disclosures is the only way to measure the true cost of the AI boom.