Masters of Measurement°

The 55th in the Egan Pattern Language

Just how much water does AI use? And once we find the answer, how can we tell what it means?

The wonderful Andy Masley, author of the substack “The Weird Turn Pro”, recently wrote a piece zooming into the specific question of AI and water usage:

The AI water issue is fake

Frankly, this is a ballsy essay, and I love it. There’s a certain lily-liveredness that creeps into rationalist environmental writing — authors more afraid to critique doomers than denialists — and it’s a delight to see Masley just stating what we can be confident of, and showing the data that lead to it. (I, of course, am an interloper in these debates: if you spot an error in the essay, please let me know!)

And he uses the typical tools of science communication to make his argument: bar charts, cool dot diagrams, and so forth. It’s good stuff! But by the end of it, I was still struggling to hold the whole picture of data centers & water usage in my head. And that’s because, of course, I’m basically innumerate. I’m not ashamed — this is normal! As Andy writes:

People are easily alarmed by contextless large numbers, like the number of gallons of water a data center is using. They compare these large numbers to other regular things they do, not to other normal industries and processes in society.

He helps us go beyond that — but I suspect that, by using the tools that we’re forever exploring in this newsletter, we can go further. I think we can get a much clearer picture of this that both fits in our heads and stays there. And I think that, in doing this, we open the door to how we might help every middle school student in our new approach to education banish their innumeracy, see our real environmental problems sharply, and quantify their imaginations.

(Note: I came up with the first version of this while working on a Science is WEIRD lesson about toilets. Usually, I don’t share those lessons here — they’re, um, how I pay the bills — but I think the lesson has enough moral oomf to justify linking to it in this post. You’ll find it at the bottom of this.)

---

How thirsty is AI?

Imaginary Interlocutor: Why is AI using water, anyway?

Computers get hot. A whole building full of computers — like a data center — can get really hot. And while it’s considered unwise to open up the frames and spray circuitry with a fire hose, piping cold water around in tubes is a pretty effective way to keep the building chilly. After it’s warmed up, the water is put back into the water cycle: either by evaporation, or by flowing into a local river or lake.

Behold the cloud!

I.I.: How much water does AI use?

All data centers in the United States use approximately 3.4 billion liters of water per day.¹

I.I.: That sounds like a lot!

It really does — but there’s a reason why “a lot” isn’t a scientific term. Our struggle to imagine truly huge numbers is why the makers of the metric system developed all those prefixes:

• 1,000 liters is a kiloliter,

• 1,000 kiloliters is a megaliter,

• 1,000 megaliters is a gigaliter,

• 1,000 gigaliters is a teraliter,

• 1,000 teraliters is a petaliter,

• 1,000 petaliters is an exaliter, and

• 1,000 exaliters is a zetaliter.

The long and short of it is that, using the Systèm Internationale², we can now say that the all the data centers in America use 3.4 gigaliters of water per day.

I.I.: In no way was that helpful.

Indeed. Kieran Egan might point out that this is an old story in science — we’ve created numerical abstractions to help us understand the world, and while they’ve succeeded at helping engineers actually do stuff, very few other people actually understand what they mean. The unwashed masses (which is almost all of us) only experience quantities through our bodies — one might say SOMATICALLY (🤸‍♀️).

What we need is a way to translate between the two.

I.I.: How can you translate technical units to felt quantities?

Link every unit to a particular thing that has approximately that measure. So, for example:

• a milligram ≈ a bee’s brain

• a gram ≈ a chocolate chip

• a kilogram ≈ a cantaloupe

I.I.: I happen to have recently weighed a single chocolate chip, and I’ve found that they weigh a bit more than 1 gram.

It’s wonderful to get these precise, and if you’d like to propose anything more accurate, please do in the comments! What’s necessary is that they be things that you experience in the real world, and can recall easily. (Close your eyes and try to feel the mass of a cantaloupe in your mind.)

I.I.: That’s mass, but what does this look like for volumes like “four gigaliters”?

I spent some time fleshing this out for my class on toilets:

• 1 liter ≈ a big water bottle³

• 1 kiloliter ≈ a hot tub⁴

• 1 megaliter ≈ a swimming pool⁵

• 1 gigaliter ≈ a small lake⁶

• 1 teraliter ≈ a big regional lake⁷

• 1 petaliter ≈ a Great Lake⁸

• 1 exaliter ≈ a sea⁹

• 1 zetaliter ≈ the sea, the world ocean¹⁰

So each day, data centers in America use about four small lakes.

I.I.: That’s helpfully visualizable!

And the nice thing is you can do a lot with a system like this. You might know that you drink 2–3 liters a day. (We can exploit that range, and use either “2” or “3” when it’s convenient.) For example:

How much did my elementary school drink per day?

• ~500 people in my elementary school

• ~500 people × ~2 liters = ~1,000 liters

• ~1,000 liters = a hot tub

How much did my high school drink per day?

• ~2,000 people in my high school

• ~2,000 people × ~2 liters = ~4,000 liters

• ~4,000 liters = 4 hot tubs

How much do I drink in a year?

• more than 300 days in a year

• ~300 days × ~3 liters ≈ ~1,000 liters

• ~1,000 liters = a hot tub all to myself…

How much does Minneapolis drink in a day?

• ~½ million people in Minneapolis itself (not including the burbs)

• ~½ million × ~2 liters = ~1 million liters

• ~1 million liters = a swimming pool

Being able to visualize the results gives a “punch” to finding an answer. Suddenly, Fermi estimates are fun again!

I.I.: Enough of your frivolity; we have the environment to think about. Is 4 small lakes a day a lot?

Well, let’s use an appropriate comparison: how much does America drink in a day?

• ~⅓ billion Americans

• ~⅓ billion × ~3 liters = 1 billion liters

• 1 billion liters = 1 small lake

If all Americans collectively drink about one small lake per day, then data centers are four times as thirsty as humans.

I.I.: That seems bad.

It does! But look at the bigger picture. Everyday, Americans…

• flush about 17 small lakes down the toilet (source), and

• spray about that much down the shower, and

• soak 30 small lakes onto our lawns (source).

I.I.: That seems terrible.

As someone who is much too excited about the low-flow, dual-flush toilet he’s preparing to buy, I won’t disagree. But note that all America’s data centers use only about a tenth as much water as do our bathrooms. If you really want to use less water, take shorter showers, and buy a better toilet.

But also, put this in the still-bigger picture: the food that we eat. Each day, our crops drink something in the ballpark of four hundred small lakes.

I.I.: Well, eaters gotta eat.

Agreed! But when we eat meat (and I am an enthusiastic, even if rather guilty, carnivore), we’re doubling the amount of water we use — roughly half goes to the animals, and half goes to the plants they ate. If you want to lower your water consumption substantially, go meatless on Mondays. If you’re already doing that, then take shorter showers and buy the new toilet. If you’re doing all of those things, then we can invest time talking about AI.

Also, remember that “four small lakes a day” is what all America’s data centers use. Only something like a fifth of that is for AI — much of the rest is “the internet” (source). And of the amount used for AI, only a fraction is used for the AIs that my friends seem particularly worried about: LLMs like ChatGPT.

---

A new possibility in education

Do I seem over-excited about this method of linking units? If I am, it’s because I don’t just see this as being useful here. This can be the anchor to all environmental education — and to the rest of science beyond it.

The environmental journalist Quico Toro — my favorite environmental journalist? — recently excoriated would-be environmentalists to whom quantities are just an abstraction:

Our brains melt on contact with unfamiliar scales.

We can sort of imagine a ton: it’s about what a small car weighs. Multiply that by a thousand and we’re lost. Kiloton, megaton and gigatons all sound like just fancier and fancier ways of saying “a lot.”….

This does bad things: smart engineering-types are investing themselves in projects that feel like they’re making a dent in combatting climate change, but are barely scratching the paint:

I get pitches about carbon dioxide removal techniques that claim to scrub carbon out of the atmosphere “at scale” all the time.

Ask “at what scale” and 99 times out of 100 it turns out they mean kiloton scale.

Press them a little, and you often get an answer in the form “if x, y and z happen, we could take this to megaton scale” — where x, y and z are some combination of political sorcery and technological miracle-making.

It enrages me.

Look, we’re adding the equivalent of 53 gigatons of CO₂ to the atmosphere each year, and each year we add a bit more than the year before.

A recent report from researchers at Oxford University shows we’ll need to remove 7-9 gigatons per year from the air to have a meaningful impact. We’re currently removing 0.0023 gigatons per year….

Carbon dioxide removal technologies with no plausible path to gigaton scale are draining talent, attention and funding from proposals that do have a credible path.

People naturally suck at thinking about big quantities. It’s not our fault; we evolved for the savanna, and envisioning “gigatons” helped precisely no one pass along their genes. Lacking a quantitative feel for the world, people are forced to fall back on their qualitative feelings… which aren’t up for the challenge of making sense of the modern world.

An inability to feel huge quantities is one of the main things standing in the way of having democracies champion policies that would actually lead to environmental flourishing.

And, look — we can fix this.

We can create these unit stand-ins for lots of different measures. Above, I shared what I’ve created for the masses of a gram, milligram, and kilogram. I’ve already created a few others:

• 1 megagram = 1 million grams = 1 ton ≈ a black rhinoceros

• 1 meter = a lightsaber¹¹

And as your body is the physical object you know best, for variations on the meter (millimeter, micrometer, nanometer, kilometer…), see our pattern The Shrink Ray°.

But we shouldn’t stop with just those quantities (length, mass, and volume). To let science truly colonize your imagination, you need to have a feel for lots of other units — like watts, joules, lumens, sieverts, and bytes.

These are easy to learn, and powerful to use. And they make Fermi estimates — which is going to be a mainstay of our approach to middle school math¹² — delightful! Used consistently, they can help form a generation of kids who can have a profound, quantified understanding of reality.

---

Questions, answered

Imaginary Interlocutor: Are you aware that you’re not the first to use these sorts of stand-ins for quantities?

Quite! They’re commonplace, especially among excellent science & math communicators. See, for example, Randall Munroe’s famous chart on radiation exposure, or any of Ben Orlin’s delightful charts near the start of his book Math for English Majors:

©2024 Ben Orlin. Ben, we haven’t met, but your work is amazing and I need to buy you a coffee.

What an Egan education has to offer here is to recognize their power, and make them a foundation of the middle and high school curriculum. (This is a theme of ours — see the pattern Cumulative Content°.) This requires time and planning to be taught, committed to memory, and regularly used so they stay remembered.

If anyone knows about a school or curriculum that already does something like this, please tell us in the comments.

Imaginary Interlocutor: When would you say these should be taught?

As soon as measures are introduced in the curriculum. Some (like length) are introduced in first and second grade; others (like electrical energy) come in far later. Whenever they come in, they should be given these links.

I.I.: You mentioned “Fermi estimates”. What are those, again?

Quick calculations that help you get a loose grasp of the big world. (The “how much did my school drink?” questions above are good examples.) Along with Bayesian reasoning, they’re going to be one of the signature pieces of our middle school math curriculum — I’ll write about this more soonishly.

I.I.: How do you get kids to remember these?

We already take memory quite seriously — see our previous posts on Memory Boxes° and Spaced Repetition Everywhere°. But these metric links are so important we’re going to bust out the big guns, and bring in tricks from The Art of Memory°. And after they’re learned, we may want to post them on the wall, just so our memories can always be double-checked.

I.I.: Above, you wrote that we want science to “colonize” our minds. That’s… a not-entirely-pleasant metaphor.

Indeed, because I’m using it to describe a not-entirely-pleasant process! Egan was big on this: you lose something when you develop a modern mind. His whole paradigm — the one that we’re developing — aims to build both qualitative and quantitative ways of understanding the world, while reducing the amount of interference they have on each other.

---

Want to help us?

If quantifying the world tickles your fancy, there are a few projects we could use your help in.

What to call these

The first is coming up with a title for these “quantity stand-ins”. I’ve been playing around with a few:

• felt measures

• magnitude markers

• embodied units

• reference magnitudes

• yardsticks

• metric analogs

But I haven’t come up with anything I’ve fallen in love with. (Anything with “units”, actually, runs the risk of being confused for the many other uses of “units” in schooling.)

More ladders

Above, I shared a ladder for units of volume a liter and above. And I mentioned some of the other quantities we’ll need — like watts, joules, lumens, sieverts, and bytes. If you’d like to try your hand at picking one of those, and fleshing it out (1 watt = a Christmas light, 1 kilowatt = a toaster, 1 megawatt = a neighborhood…), we’d love it!

Completed ladders

The volume ladder above doesn’t go below a liter. It should! We could use milliliters, microliters, and, heck, maybe even nanoliters.

Finer-graded ladders

So far, our ladders use powers of 1,000 — we jump right from “liters” to “kiloliters”. But it’d be helpful to create some stand-ins for all the powers of 10, at least in some parts of the ladder. (That is, it’d probably be more useful to have an anchor for “10 liters” or “100 liters” than it would be for “10 exaliters”.)

If you’re eager to do a bit of research, and would be willing to share what you come up with, we’d love it — and might use it to help lots of kids understand the world. Just throw them in the comments below. (And, when appropriate, if you could also include the precise measures of any of your stand-ins, that’d be useful — it’d help us judge between multiple options.)

---

The toilets lesson

All of the above is just a sliver of the Science is WEIRD lesson I gave recently about toilets, which also includes where your wastewater actually goes, how a toilet works, and an interesting twist at the end about the future of society. Feel free to enjoy it with kids, or by yourself.

Special special thanks to Amy Teegarden, our Science is WEIRD science advisor, who helped me research that lesson, and then did the extra corrections for this blog post. There are a few estimates that we’ve improved between the two; we’ll make an addendum on the class video shortly. (We’ve also improved the estimates of desalination at the end. We’ll be publishing our notes on that soon; if you’d like them now, just ping me.)

1. This is a high estimate, taken from a 2021 article in the journal Nature, “Data centre water consumption”, and doubled to account for growth of the sector in the last four years. However, there are multiple ways to measure water usage, and 3.4 billion liters is a high estimate. For more, see the big 2024 report on data center electricity and water usage from Lawrence Berkeley National Laboratory (LBNL), especially pages 55–58. For some hemming and hawing over that, see the post “I Was Wrong about Data Center Water Consumption” on Construction Physics. For more, I don’t know, get an engineering degree.

2. The modern metric system, known as the “S.I.” to its friends, but pronounced with an over-the-top French accent in the U.S.

3. I was ashamed to discover, at the age of 43, that a liter is almost precisely the same as a quart! Somehow, I never knew this; my wife tells me she’s always known this.

4. A very small hot tub might be only 1/2 kiloliter; a very big one might be 2 kiloliters.

5. Again, a range: an above-ground backyard pool might be only 1/10th a megaliter, while an Olympic-sized pool is 2.5 megaliters. If you want to be precise, you might want to link this to a particular pool you know of something close to the right size.

6. Think “lake which, if I swam across it as a kid, I’d boast to my friends about.” My link here is one of my two lakes in my town: ChatGPT’s estimates of my local Cascade Lake system range from 0.4 to 1.0 to 1.6 gigaliters.

7. Think “lake which, if I swam across it now, I’d make a custom t-shirt to brag about.” Having lived for a decade in and around Seattle, I use Lake Washington for this. (Recommended for kayaking!) The only trouble is that it’s nearly three teraliters. I try to fix this by only imagining the middle third, between the two freeways.

8. Specifically, Lake Ontario, that measliest of the Great Lakes, which weighs in at 1.6 petaliters. The other four are much larger.

9. I’ve had some trouble here: the Black Sea is too small (~½ exaliter) and the Mediterranean is too big (~4 exaliters), so if you can find a Goldilocks Sea that’s just right, let us know in the comments.

10. Not quite 1.4 zetameters.

11. I believe there was a hot minute where “1 meter” was actually canon for the length of a lightsaber; as with so many situations, this was complicated by the arrival of Darth Maul.

12. More to come about this soon.