AI's Energy & Water Demands: Sorting Fact from Fiction with Andy Masley

What It Covers

Andy Masley, director of Effective Altruism DC, analyzes AI's actual energy and water consumption using back-of-envelope calculations to counter widespread misconceptions. The conversation establishes concrete heuristics comparing ChatGPT prompts to microwaves, car trips, and flights, concluding that AI represents a 1-2% increase in global energy use even at full projected build-out scale.

Key Questions Answered

• •Personal prompt footprint: A median ChatGPT prompt consumes approximately 0.3–0.6 watt-hours of energy — equivalent to running a microwave for one second. Reaching 1,000 prompts in a single day increases personal emissions by roughly 1%. Since that volume requires ~10 hours of continuous prompting, any activity it displaces almost certainly emits more. Computing is so energy-efficient that substituting AI for nearly any physical activity produces a net emissions reduction.
• •Single car trip offset rule: One 20-mile crosstown car trip emits 3–4 kilograms of CO₂, equivalent to approximately 10,000 median ChatGPT prompts. A full 20-gallon tank of gas equals roughly 250,000–500,000 prompts. A transcontinental flight reaches 1–2 million prompt equivalents. Avoiding a single car trip therefore offsets an entire year of typical AI usage, making behavioral substitution the dominant variable in any personal emissions calculation involving AI.
• •Global build-out scale check: The full $7 trillion AI infrastructure build-out, projected to consume approximately 80 gigawatts of power over time, represents a 1–2% increase in global energy usage. This is smaller than the energy increase expected from general global economic growth over the same period. Powering all 80 gigawatts with solar panels would require roughly 800 square miles — less than 1% of Nevada's land area.
• •Water consumption reality check: The widely circulated claim that one ChatGPT prompt consumes a bottle of water is off by a factor of roughly 200. Actual consumptive water use per prompt is closer to 2 milliliters. The bottle-of-water figure originated from a Washington Post analysis that assumed 10–20 prompts per 100-word email and used pre-commercialization efficiency data. A single additional pair of jeans, by contrast, requires water equivalent to approximately 1 million prompt-equivalents due to irrigated cotton production.
• •Consumptive vs. withdrawal water distinction: Up to 90% of water figures cited in AI environmental coverage refer to non-consumptive withdrawal — water taken by power plants and returned to the source. Actual consumptive water use by data centers themselves represents roughly 3% of headline figures. The more meaningful metric is consumptive use in high water-stress regions. In 2023, total U.S. AI water consumption was comparable to roughly 10 towns of 15,000 people distributed across the country.
• •Chip energy economics: An 8-GPU H100 server node costs approximately $300,000 to purchase but only $35,000 in electricity over a four-year lifespan — a 10:1 ratio favoring purchase cost. However, carbon emissions flip in the opposite direction: electricity consumption generates roughly 20 times more carbon than the embodied emissions of manufacturing the chips themselves. This means operational electricity, not hardware production, is the correct focus for any emissions analysis of AI infrastructure.
• •Air pollution as the primary local risk: Climate impact and water use are secondary concerns compared to localized air pollution from data centers drawing on fossil-fuel-heavy grids. U.S. air pollution already causes an estimated 30,000–100,000 deaths annually. Data centers in areas with existing poor air quality — such as the Memphis Colossus site, which had pre-existing Grade F air ratings — concentrate additional pollution harm on already-burdened communities. Local leaders negotiating data center agreements should prioritize air quality monitoring and permitting compliance above water or climate concerns.