I hope this email finds you parched.

You HAVE TO send that email today. Yeah, that one. The one you put off. The one you don’t want to think about.

But you know, nobody is looking.

You could just use ChatGPT.

Or—hear me out—you could suck it up, write the email yourself, and light an apartment for an hour with the power you saved.

That was the roughly the figure that the Washington Post published last year, in partnership with researchers Pengfei Li, Jianyi Yang, Mohammad A. Islam, and Shaolei Ren. I was reminded of the Post piece (and the academics’ related paper) last week when I heard someone say that a single Sora 2 video clip costs OpenAI $5 to generate (which seems plausible given the pricing for the Sora 2 API).

Five. Dollars. For a minute of AI slop that virtually nobody but the person requesting it will want, and that literally nobody—including its requester—needs.

ZDNet reported earlier this year that the original Sora generated 600 videos per minute. If OpenAI's new free Sora app is that busy, at $5 a shot, it costs $3000 a minute. Over $4 million per day.

OpenAI could shut the servers down, and use the savings from five months of Sora 2 video output to pick up the tab for solving Flint, Michigan’s water crisis (estimated at $626 million).

That kind of Eat This, Not That comparison gets my attention. Scarce drinking water is a timely topic, too, as hyperscalers are in the news this month for trying to conceal how much water their data centers need. Which is a lot, apparently: generating a 100-word email consumes about a half-liter of water (for power generation and data center cooling), according to the previously-mentioned WaPo article.

As an aside, it is worth noting (barely) that this figure is contested. OpenAI CEO Sam Altman said that an “average” ChatGPT query used about a fifteenth of a teaspoon of water. But what’s an average query? And what is counted as “using?” Is it just the amount of water consumed by data center cooling that is directly attributable to ChatGPT inference? Or is it inference AND model training? What about water consumed by power generation? Sammy, predictably, did not show his work, so we will stick with the half-liter figure.

I also accept the “100-word email” frame of reference for large language model output. As illustrated by headlines in the “Scientists Discover Asteroid Measuring Over 50 Football Fields Across” style, Americans will do anything to avoid learning the metric system. Well, here I am, an American, measuring things in terms of “emails.” But a brief email is a good unit of work, if we are talking about LLMs. In 2024, when a former boss had me testing and demoing Copilot for other executives, one of the best demos (by which I mean the only one that worked) was summarizing and drafting emails.

So, let’s spend some water and see how much email we can make.

I could have outsourced the email I wrote when I got back from my morning walk to ChatGPT, in lieu of cooling off with a bottle of water.

I sent thirteen emails Tuesday afternoon as part of a conversation discussing a current project; I could have had ChatGPT handle those for me, at a cost of over six liters, and just skipped flushing the toilet in the office restroom.

What if I sent around 40 emails a day? I send fewer than that, but back when I was a project manager, 40 would have been a calm day. At that rate, I could outsource an entire week of email while consuming 100 liters of water, which I could make back by skipping one morning shower.

Now, both my office and I stink, but let’s take it further. If I was still a project manager, I could use ChatGPT to draft all 10,000 of the emails I would write in a year, instead of spending those 5,000 liters of water doing 50 loads of laundry or running the dishwasher a few hundred times.

And yes, I could also just, I don’t know, RUN THE DISHWASHER LESS OFTEN. And I’m sure AI boosters who hate this article and eco-conscious people who are otherwise on my side will come together to say, “You could also take shorter showers, Max.”

And yeah. I could. Shut up. This is my newsletter, so we aren’t going to discuss my individual choices.

But I’m also not here to yell at you for your individual choices (not today, in any case).

In the short run, many of us aren’t even aware that we’re making a this-or-that choice. You can spend all day gabbing with ChatGPT and writing code with Claude and letting the mask slip with Grok, while still taking thirty-minute showers and leaving the water running when you brush your teeth. But I bring up the resource cost because while each of us, individually, do not yet feel the pain of making that choice, we are collectively making it. And when I make such choices, I want to be conscious about what I am getting, and what I am giving up to get it. Even—especially—if the bill hasn’t yet come due.

And the currency that bill is paid in is, among other things, WATER. Without water, we die. In many places, we are running out of it. And we are trading it for, what? For email? For brainstorming help? For code that a real programmer is going to have to fix later, anyhow?

Water is finite. Available potable water is running short. Meanwhile, we are in no danger of running out of email. We face exactly zero problems, as a species, for which the solution is "more email."

Using generative AI is trading the very stuff of life for something we already have too much of.