I Drove Through a Car Wash and Used More Water Than 500,000 AI Prompts
On amortized cost, the discovery dividend, and the strange math of moral consistency.
I drove my Tesla Model Y through a car wash on Sunday. Or more accurately, I spent three frustrating minutes in the queue trying to figure out where Tesla moved the car wash mode button this time — they seem to inconveniently relocate it with every update.
Once I successfully engaged the “free roll” setting, off we went through the wet, sloppy, spinny automated cleaning machine. Five minutes, maybe six. That satisfying moment where the dryer hits and the water sheets off the hood. Ooooh! Shiny! Ahhhh.
Now, how do I turn off the damn Car Wash mode?
Later that night, after taking a photo of my friend’s adorable long-haired chihuahua, I found the sudden urge to see what Jasper might look like in a Weyland- Yutani spacesuit. (Because, let’s be honest, who doesn’t want to see that?)
This sparked a conversation between Jasper’s human and me: is generating an AI image of an adorable long-haired chihuahua in a Weyland-Yutani spacesuit a crime against the environment?
The argument goes something like this: training a large AI model costs the equivalent of 50 transatlantic flights in carbon emissions. Therefore, every AI-generated image is a small withdrawal from an account we can’t afford to overdraw.
That number gets repeated often. It is also, when examined with any rigor, a profoundly incomplete way to think about cost.
A single pair of Levi’s 501 jeans — not a factory, just one pair — consumes 3,781 liters of water and emits 33.4 kg of CO₂ over its lifetime. I own several pairs. Nobody is writing op-eds about dad jeans destroying the environment.
Jeans Are Not a One-Time Cost
The AI energy debate consistently ignores this: my Levi’s 501’s didn’t stop consuming resources when I bought them. Every wash cycle adds to the ledger. The cotton required roughly 10,000 liters of water to grow. The “stone-washed” dyeing process generated chemical waste. The shipping involved fossil fuels at every leg. And I wash them every few weeks, for years, in a machine that uses water and electricity each time.
AI training, by contrast, is an infrastructure cost — not a consumable one. You pay it once. Then you build on top of it.
Training GPT-3 consumed an estimated 700,000 liters of water for cooling. That is a lot. But spread across hundreds of millions of queries, the per-interaction cost becomes statistically negligible. And 700,000 liters is the lifetime water footprint of just 185 pairs of jeans. I don’t own 185 pairs, but between me and everyone I know, we’re getting there.
The Car Wash
According to the International Carwash Association, a professional tunnel car wash uses approximately 150 liters of water (40 gallons). A driveway wash with a garden hose runs north of 450 liters.
A single AI text prompt consumes roughly 0.3 milliliters of water when you account for data center cooling on a per-query basis. Which means my Sunday car wash consumed more water than approximately 500,000 text prompts.
Nobody is writing headlines about my trip through the car wash. Yet we have an abundance of headlines about pets in spacesuits.
A car wash asks nothing of us philosophically. It doesn’t claim to reshape the world. It just makes my hood shiny (until it rains.) AI claims more , so it receives more scrutiny. Scrutiny should scale with ambition. But it should also scale with evidence — and the evidence, when we do the math, points more at us than at the data centers.
The Discovery Dividend
The 50-flights framing omits what has been coming out of the laboratory using today’s leading AI technologies.
In 2023, Google DeepMind’s GNoME project — a materials science AI — predicted 2.2 million new crystal structures, of which 380,000 were found to be stable. The paper, published in Nature, described this as the equivalent of roughly 800 years of prior human discovery, compressed into a single research effort.
Microsoft and the Pacific Northwest National Laboratory used AI to screen 32 million candidate materials in 80 hours — identifying 18 viable battery candidates in a process that would have taken human researchers an estimated 20 years manually.
And Insilico Medicine ran a drug entirely conceived, designed, and progressed by AI into Phase II clinical trials with human patients — the first time that milestone had ever been reached by a machine-discovered compound.
My Sunday car wash produced a shiny hood.
AI training produced the capacity to synthesize 2.2 million new materials, collapse drug discovery timelines by decades, and find battery chemistry that could reshape the energy grid.
These are not hypothetical.
The Real Question
AI energy consumption matters. Data centers consume real resources, and that consumption will grow. But having the conversation seriously means using the right unit of analysis. The relevant question is not how much energy training cost — it’s the amortized cost per unit of value created.
By that measure, the 50 flights to build the model look very different when those flights also paid for a tool that can optimize global shipping routes, potentially eliminating thousands of actual flights; a tool that can design more efficient pumps, saving billions of gallons of water; a tool that can screen millions of drug candidates in the time it once took to screen dozens.
My Levi’s will never do any of that. They’ll just keep accumulating water and energy costs every time I run a wash cycle.
The Math I Didn’t Do Until Sunday
I’m not here to let AI off the hook. I’m here because I drove through a car wash this weekend and then did the math on it, and the math made me uncomfortable — not about Jasper in a spacesuit, but about how selectively I’d been applying my outrage.
I own Levi’s. I run a dishwasher. I take hot showers. I have eaten more cheeseburgers than I care to admit. Each carries a resource cost that dwarfs a text prompt by orders of magnitude. I never once felt guilty about any of them.
(Well, maybe the cheeseburgers.)
My car wash was an environmental expense that ended with a shiny car. AI training is an environmental investment that ends with the capacity and potential to solve the very environmental problems we’re most worried about.
And, because I teased it already, here’s Jasper in all his Nano Banana-generated spacesuit glory:
Sources and methodology notes available on request. Key references: Levi Strauss & Co. Life Cycle Assessment (2015); Luccioni et al., “Power Hungry Processing” (ArXiv 2023); International Carwash Association WaterSavers data; Google DeepMind GNoME (Nature, 2023); Microsoft/PNNL battery research (2023); Insilico Medicine Phase II trial (Nature Biotechnology, 2023).