The Almond Defense: Inside Sam Altman’s Claims on AI Water Footprints

Tech & Science

Facing mounting community friction over infrastructure demand, OpenAI CEO Sam Altman offered a memorable defense on the Sources podcast: running 38,000 ChatGPT queries, he argued, consumes less water than growing a single California almond.

Dismissing claims that individual prompts drain bathtubs as an unfounded "meme," Altman asserted that in modern, closed-loop designs, computing facilities can consume roughly the same water volume as an ordinary commercial office building. While published direct-cooling estimates support his broader point that a single prompt uses tiny amounts of water compared to heavy agriculture, the specific 38,000 figure is not an independently verified constant. More importantly, evaluating a facility's environmental toll solely through on-site evaporation leaves out the largest slice of the resource ledger: off-site thermoelectric power generation.

1. Auditing the Analogy: Representative Benchmark vs. Hard Science

Comparing computational loads to crop cultivation blends two disparate sets of baseline data:

Comparison Metric	Disclosed Baseline Figure	Standard Conversion Value
ChatGPT Text Query	~0.32 milliliters (mL) per interaction	Direct on-site consumption baseline
Single California Almond	~3.56 liters (1.1 gallons) per nut	≈ 11,100 queries (3.56 L ÷ 0.32 mL)
Altman's Claim (38,000 Queries)	38,000 queries cited from memory	~12.16 liters (equivalent to ~3.4 almonds)

At 0.32 mL per query, 38,000 interactions consume roughly 12.16 liters of water—equivalent to the water needed to grow about 3.4 almonds. While claiming that "thousands of queries" equal an almond falls safely within the ballpark, citing 38,000 specifically stretches the ratio by more than threefold. Altman acknowledged during the broadcast that he was recalling the figure from memory without releasing his underlying model parameters. The comparison functions as a memorable conversational metaphor rather than an experimentally controlled measurement.

2. The Iceberg Effect: Direct Cooling vs. Indirect Power Water

The primary technical flaw in evaluating AI through single-prompt micro-metrics is the separation of direct site operations from the broader energy grid:

Water Category	Operational Boundary	Grid & Environmental Profile
Direct Water Usage	Cooling towers, chillers, and on-site facility evaporation.	Directly impacts municipal water reserves and local aquifers.
Indirect Water Usage	Thermoelectric power plant cooling to generate server electricity.	Often accounts for roughly two-thirds to three-quarters, and in some grid mixes over 80%, of total operational water.
Total Water Footprint	Direct on-site evaporation + upstream utility water draw.	Determined by regional grid energy mix, hardware, and climate.

Because thermoelectric plants (coal, natural gas, nuclear) require vast volumes of cooling water to spin steam turbines, indirect water consumption often accounts for roughly two-thirds to three-quarters—and in some grid mixes well over 80%—of a data center's total operational water consumption. Disclosing only direct on-site evaporation cuts out the majority of the environmental bill.

Furthermore, applying a static 0.32 mL metric overlooks the architectural evolution of artificial intelligence. As systems pivot from single-turn chatbots to multi-step reasoning models and recursive autonomous agents, computational workflows expand by orders of magnitude. Complex inference jobs executing iterative code runs and tool lookups cannot be evaluated using legacy text-prompt benchmarks.

3. Local Realities: The Virginia Audit and Regional Disparities

Is the claim that data centers consume water like an ordinary office building valid? The answer depends heavily on geography and mechanical infrastructure:

• The Virginia JLARC Findings: A landmark 2024 study by Virginia’s Joint Legislative Audit and Review Commission (JLARC) examined the world's largest data center hub. The audit highlighted growing strain on regional electrical and water infrastructure, while noting that facility design drives massive divergence: some modern closed-loop campuses can use water comparable to or less than a typical large commercial office building, though facility-to-facility variance is large.
• Facility-to-Facility Variance: The Virginia audit also revealed that low water use is not universal. In 2025 disclosures, direct water efficiency varied by tens of times across different sites operated by the same cloud providers, heavily influenced by whether older evaporative towers or modern closed loops were installed.
• Regional Balancing: In drought-prone regions, deploying dry or low-water chillers protects scarce drinking aquifers. Conversely, on carbon-intensive electrical grids, cutting kilowatt-hour load delivers far greater reductions in overall water consumption than tweaking on-site plumbing.

4. Global Policy Shifts: Mandatory Caps and Grid Moratoriums

As community resistance grows and power reserves tighten, governments worldwide are moving past voluntary corporate estimates to impose binding regulatory standards:

Key Regulatory Interventions:

• Singapore's Roadmap Target: Under Singapore’s Green Data Centre Roadmap, the Infocomm Media Development Authority (IMDA) targets a Water Usage Effectiveness (WUE) cap of 2.0 cubic meters per megawatt-hour (m³/MWh) for sustainable facilities, accelerating the shift toward advanced water-recycling and air-cooled designs.
• Texas Infrastructure Scrutiny: Facing mounting strain across regional utility networks, in some jurisdictions approvals for prospective data centers have been paused or subjected to additional joint reviews of power and water demands.

Sam Altman’s comparison succeeded in contextualizing the micro-scale direct consumption of everyday digital queries against water-heavy industrial agriculture. However, because it bypasses the dominant indirect footprint of power plants, overlooks the computational demands of agentic AI, and ignores localized aquifer stress, the "almond defense" remains a clever public relations shorthand rather than an exhaustive sustainability framework.

---

Official Research & Regulatory Documentation

• Virginia General Assembly (JLARC): JLARC Study — Data Center Resource and Water Impacts in Virginia
• U.S. Geological Survey (USGS): USGS California Water Science Center — Agricultural Water Use Data
• Infocomm Media Development Authority (IMDA): Singapore IMDA — Green Data Centre Roadmap & Sustainable Standards

Data Methodology & Environmental Accounting Notice: This analysis references statements from the Sources podcast, agricultural statistics from the USGS, data center legislative audits from Virginia JLARC, and sustainable computing frameworks. Specific query-level water estimates in academic literature range from roughly 0.3 mL to several tens of mL per query depending on accounting boundaries, model scale, runtime ambient temperatures, and whether indirect power plant generation water is factored into the models.