Notes on Hank Green's Video on AI Water Usage

The Notes

• water use is very easy to mislead about because the term can mean lots of different things.
• most folks don't have a grounding in the nuance so vastly different numbers end up in the conversation which makes it hard to figure out what's actually going on.
• two numbers quoted at the start of this video are that each query takes 1/15 of a teaspoon of water and that 1 trillion liters of water per year will be used by 2028.
• those end up being very different numbers but could both be true because resource consumption analysis for any industry analysis is very tricky with multiple possible approaches.
• there are different sources of water that data centers can use. Fresh water from municipal sources (i.e. the same stuff that comes out of your tap), non-potable water coming from a sewage treatment plant, etc... but most are currently using the municipal stuff.
• the misleading part of the 1/15 of a teaspoon number (which came from sam altman) is that that's just the query and not all other parts of the process (e.g. the training that went into making the model. hank actually labels this as a lie).
• there's also different ways to use those numbers. e.g. should you include the cost for just training cpt5 and spread it across the queries that go to gpt5. since gpt5 is a bunch of models rolled into one, should you also include all the models that go into it. what about all the models that were trained that aren't used in gpt5, etc....
• U. of California estimate puts training at 50% of the resource use of running an AI service.
• It seems like altman is not including the training when talking about resource usage.
• One reason for all the new data centers is the expectation that training requirements will go up a lot in the future.
• Microsoft, Amazon, and Google are collectively spending $100Billion+ per year on new data centers.
• Aside from direct use of water, data centers also use a lot of power and lots of that power comes from thermo-electric systems (that use water that makes steam that spins turbines and is also used to cool the steam back down to re-condense it so it move through properly).
• this usage in energy systems accounts for 40% of "fresh water withdraws" in the u.s. but, a lot of that water gets put back. also, none of that water is municipal water. It's usually taken directly from a body of water and released back into it. If you throw that energy generation number in for water usage the number gets way bigger.
• regardless of the source, every location has a finite hydrological budget.
• "a data center drawing from a lake is not competing with houses for tap water, but it is drawing from the same watershed. And, in a lot of places, that watershed is already fully allocated. But, in a lot of cases, that's not the case."
• "so, the question isn't what kind of water is it (the data center) using, it is also: does this place actually have the water to spare. Some places do. Some places do not."

• talking about the types of water, the water that goes into making the GPU chips is another type. It doesn't take much of it by comparison, but it has to be super-duper distilled to remove impurities. So, even though was less of it is necessary, it costs a ton of energy to produce.

  "and that's not something anybody needs to know unless they are experts working on the problem."

• big take away at this point: the issue is really complex. That, in turn, makes it easy to mislead about.
• "this is the reason why expertise is so important. Because, the discourse is terrible at expertise. But, there are people who are actually involved in this conservation, who are actually involved in trying to mitigate this, to decrease the water use, and they are aware that AI is just one of many industrial uses of municipal water. And, that that is very different from water evaporating at a power plant that isn't coming from municipal water. "

• Plot twist: Resource managers will tell you that use is one part of the story, but often the smaller part.

  What matters most is where you're using it.

  (e.g. a bunch of data centers in the deserts is a totally different game than a bunch in the pacific north west)

• water is not like electrons, it's hard to move around. If you use it where it's scarce, it matters a lot more than if you use it where it's not.
• All that said, as much water as AI data centers are going to use, it's not going to be nearly as much as the amount we use to grow corn.
• It takes between 600K and 1MM gallons of water to grow an acre of corn.

• According to U.S. dept of Agriculture corn take about 20 trillion gallons of water per year in the U.S. alone.

  compared to current total global AI use of ~260 billion gallons

  i.e. U.S. corn alone takes ~80x more than the global AI footprint

• While corn is food which accounts for a lot of it, 40% is burned in automobiles as fuel. The acre of corn that took 600+K gallons of water produces about 500 gallons of ethanol.

• So, before processing, each gallon of ethanol carries an irrigation footprint of ~1,500 gallons of water.

  It's not municipal water, but see about for that whole conversation.

• We also spend trillions of gallons of water each year watering lawns.

• Wrapping up:

  • There's a lot of complexity to all this and the nuance is often lost (or never even enters) the general discourse.
  • The big take away: some areas are right up against their water budgets. Some have more room. In areas where there's stress, it totally makes sense to not increase the use and save as much as possible.
  • Probably there will be folks who know what they are doing who do an okay job of it most of the time when it comes to planning.
  • The projected increase in water use is relative small compared to existing industrial and agricultural use. What's not small is the increase in power use.

  • The other big question is if all this compute that a trillion dollars of money have gone into making GPUs for doesn't turn out to be all that necessary (because of technical, political constraints or just a miss in the expectation of demand) is if this is a bubble that's popping.

    "I think our entire economy is being wagered by not very many people making very strange choices based on an imagining of the future that is, honestly, I don't think likely to occur."

    "I remain fairly unconvinced that people are going to continue inviting giant machines that raise their electricity bills while eating jobs into their communities. "