AI Water Use

A reader asked “What is the deal with AI and water usage?”.

This broadly speaks to ongoing discussion around AI and resource usage. The popular debate generally centers around claims that AI datacenters are using too much water and that this resource usage is harming people.

The argument is then positioned as an example of the upper-class putting finite resources toward a luxury at the expense of the lower-class. Rich versus poor, the haves versus the have-nots, bourgeoisie versus proletariat, etc.

In the interest of being productive, instead this will be positioned as a risk analysis. How much water is actually used? Why does this matter (or not)? What another analysis can be gained from this?

Central to this challenge ought to be determining how much water is actually being used. There is little point in debating a topic such as this without some minimum understanding of the basics that are being discussed. As this topic surrounds too much water being used, the central claim ought to establish at least one of three things:

1. Is there a threshold of ‘too much water’ usage in total. For example, where so much water is used the entire natural reservoir is drained dry. This is agnostic to AI, but speaks to the absolute limit of all water usage in total in a given reservoir.

2. Is there a threshold of too much water usage to AI specifically. This is an opportunity cost issue. The idea being that past a certain point, the water could be better allocated to something else.

3. That AI is a net negative to society regardless of resource usage. This makes the above irrelevant. If this is established that any resource usage would be too much. Consider, this argument also makes the entire resource argument pointless. If the issue is that AI is ‘evil’, ‘bad for humankind’, etc. irrespective of resource usage then the resource argument is irrelevant unless AI is actively generating resources and not using them.

To address the first two, one should first figure out how much water AI is using. This is complicated, as AI companies have gone to some effort to not publicly disclose their total water use, but one can approximate.

There are three broad categories for AI water usage:

• Direct – Water consumer onsite by the data center.

• Indirect – Water consumed to generate electricity for the data center.

• Supply Chain – Water consumed to manufacture the equipment to run the data center (like building the servers).

To compare the water usage of AI datacenters and include the third category would require comparing any other comparison industry with a full life-cycle against full life-cycle analysis. 

This would mean that to compare water usage against a retail clothing store, one would have to also include everything, including all the water consumed by all the workers making the clothes across the world. For the purposes of simplicity this third part will be ignored.

For direct and indirect water consumption a study published by the Association for Computing Machinery titled “Making AI Less Thirsty” provided the following table:

Model training and continued use have been separated, and direct versus indirect have also been separated.

One thing that immediately stands out is that each of these data centers has different environmental conditions, different cooling equipment, different models, and a few other variables. Each of these things effect how much computation is required per use and how much water would be required for cooling. Thinking about it another way:

A more complex computation needs more water.

Hot environments are harder to cool and require more water.

Now looking at offsite versus onsite it becomes immediately clear that the majority of the water usage is actually not at the data center. Most of the water is used at the power generation facility to generate the power the data center needs.

Indirect water usage (the power plant)

Addressing water usage and power generation very quickly leads into overall conversations surrounding climate change.

The debate effectively becomes a reframing of the standard climate change formula:

Impact = Population x Affluence x Technology

Water Usage = (The number of people) x (the amount of water a given person uses) x (technology’s ability to improve water use efficiency)

To summarize where this goes as this has already been previously addressed in our article and YouTube video around the “Maximal Approach to Climate Change”:

• Population control is a topic that is generally considered so distasteful, no one tends to engage on it with this topic.

• People tend to prioritize short-term gain for the majority over.

  • Long-term gain (in general)

  • The interest of the minority (in general)

• Technological advancement is generally the only practical route that is palatable to the public.

For simplicity’s sake: in general people are going to demand that power be build, the consequences be damned.

Now while this is the sort of behaviour that tends to infuriate risk managers while simultaneously justifying their existence, the reality is what it is.

It also means that the water use for power generation does not really matter. They still need to pay the market price for power, and they need to provide a product or service that consumers are willing to pay for. 

In the long-term if they don’t provide a profitable product or service, they essentially subsidize the creation of power infrastructure that the public will end up using anyways.

The power company is going to build power, the public tends to vote for more power, these companies still have to survive in the market to justify the resource use, this generally is not going to change.

They could use 90% of the power on the planet but would have to provide a product or service that is sufficient to justify the absolutely colossal resource expense that would incur. Unless they are curing all disease and fixing the majority of all the worlds problems, the resource cost would bankrupt them. Even if they did this, it still might bankrupt them.

The direct water usage is far more interesting. For that, here are some comparison numbers:

Context: An AI model is the thing one sends the prompt to. “Hello AI, please tell me how to bake a cake” The AI model is trained to know how to answer this, this takes water. The AI also has to process the individual request, this also takes water. 

The model training is before use, a fixed cost; the per request is during use, a variable cost.

Water per request ranged from 0.08mL – 6.52mL. The amounts for facilities under construction were omitted. One of these is zero, the other is slightly higher than the upper limit.

Water for model training ranged from 0.026 million L – 2.098 million L. Same as above, under construction was omitted.

To help visualise these amounts:

A standard sized water bottle is usually 500ml.

An official Olympic-size swimming pool holds approximately 2.5 million litres of water.

This means that one 500mL water bottle is about 6250 – 76 requests.

To train an AI model it is 0.0104 – 0.8392 Olympic-size swimming pools.

Training the AI model in water bottles is 52,000 – 4,196,000 water bottles.

This may sound like either a lot of water or a little water without further context. This cuts directly to the opportunity cost issue, the 2^nd threshold mentioned at the beginning.

Either way, now there is a number range that can be worked with.

Threshold #1, Too Much Water Usage in Total.

Addressing this requires analyzing the areas available water reservoir/watershed.

If the addition of an AI datacenter means exceeding the ability for the area to support it, the resource is not sustainable. This is not only impractical for the local population, it is also impractical for the AI datacenter itself.

This problem is location specific, seasonally specific, relative to what other strains additionally exist on the watershed, and have several other relevant factors.

There is also the consideration that all water is not technically the same, sometimes water needs to be used to purify water or alter it into a specific state, but this all becomes location specific considerations.

Threshold #2, Opportunity Cost

This is going to be the bulk of what can realistically be explored here.

The first comparison is going to be done against Almonds in California. Almonds are a cherry-picked edge case. This is going to be one of the cases that is most favourable for AI datacenters. It is fairly frequently brought up, so it is useful to cover.

Almond water use on-site is broken down into three categories:

Green water (rain water used)

Blue water (irrigation water used)

Grey water (fresh water used to dilute pollution)

There are debates as to how many of these numbers should be used in comparison to AI. Some argue that only Blue water should be considered as that is direct irrigation.

Others argue that all three should be used as the Green water use would have otherwise gone in the reservoir and is used on-site, and the Grey water use was made necessary because of the almond pollution.

Both Blue only and all three combined will be provided as a range to be fair for comparison.

Blue water is 0.8 – 1.3 gallons per almond (3-5 litres)

All water is approximately 3.2 gallons per almond (12 litres)

This provides a range of 3 – 12 litres per almond.

Converted into water bottles, that is 6 – 24 water bottles per almond.

Comparing AI requests to almonds, this would mean that the ratio is:

Low almond water use, High AI water use (best case for almonds): 

1 almond / 6 bottles = 76 requests / 1 bottle  

1 almond = 456 requests

High almond water use, High AI water us (best case for AI):

1 almond / 24 bottles = 6250 requests / 1 bottle

1 almond = 150,000 requests

To compare against the AI model training:

A 1lb bag of almonds has 368-400 almonds. This translates to 2208 – 9600 water bottles. (lower bound x lower bound) – (higher bound x higher bound).

From above: Training the AI model is 52,000 – 4,196,000 water bottles.

Using the same method just used:

(Best for almond bags)

1 almond bag / 2208 water bottles = 1 AI model / 4,196,000 water bottles

1,900.4 almonds bags = 1 AI model

(Best for AI model training)

1 almond bag / 9600 water bottles = 1 AI model / 52,000 water bottles

5.4 almond bags = 1 AI model

Assuming the public values the convenience of AI, it is likely that people will choose requests over single almonds. Even with the bags, given that this is a fixed cost only incurred once, the public might still choose to forego 1900 almond bags in the extreme case if it means their preferred AI model works.

Other food examples that are often brought up include dried pasta. Spaghetti (dried), also not including any of the water to cook, only the pasta production, is estimated to require 139 – 185 Litres for a single serving. That is 278 – 370 water bottles.

Numerous other examples can be given, fundamentally it will come down to this:

Determine the value of AI, compare against the value of the other thing, think about what one would rather have more of.

Realistically, putting this comparison into practice on a case-by-case basis would be impractical. If people were actually really concerned about water use the cleanest policy solution would be a water-tax.

Similar to the arguments for a carbon-tax, this sort of idea is the thing that virtually every economist agrees is good policy, but simultaneously becomes politically impractical.

If one really wanted to solve this issue through the opportunity cost lens, simply increase the price of the resource so that its usage needs to be better managed, and let the economy figure it out. This is a charge on resource use, a tax.

While this may appear flippant, if individuals are viewing this through the opportunity cost lens one of three things is generally true:

They are for a water-tax so long as it is equally applied on a consistent basis (Cost in $ per litre as an exact ratio).

They don’t actually care about water usage in this way.

They either don’t know basic economics, or reject market systems entirely, which places their concerns outside of scope.

Threshold #3, AI is Evil

For some, everything above does not matter. This is commonly seen when arguments are made about the above, but without data or citations to justify the position.

An example would be a water scarcity claim “AI is draining our rivers”.

Or a magnitude claim “AI consumes as much water as entire cities”.

The first is case specific, nuanced, and is covered above. Also, it is generally sensational and misleading. The second claim is either false or is using two very extreme edge cases. It might be comparing all data centers globally (including non-AI) with a very small city.

Sometimes claims are made where the above is irrelevant.

An example would be an environmental claim “the environmental cost of AI is unacceptable”.

This is generally assumed from a risk averse climate change position. From this position, anything unnecessary that has a negative climate impact is hastening the destruction of the human species.

Interestingly, from this same perspective, AI is actually being used significantly in the technology component of Impact = Population x Affluence x Technology.

Doing large, resource heavy computations is necessary in rapidly advancing technology. In a way, from the apocalyptic climate perspective, humanity is in a sort of race:

Climate change accelerating technology is required to invent the solution to climate change. This creates a unique problem with two broad options:

1. Accelerate climate change to try and find a solution.

2. Give up on the solution and accept the inevitable end, gain a bit of extra time though.

This creates a strange situation where if one doesn’t believe in climate change, they can assume a pro-AI position, and if one views climate change as an extinction level threat, they can also assume a pro-AI position.

Getting back to other reasons people give, some people also make ethical arguments that “AI is destroying art and culture”.

When water is brought up in that context, it is irrelevant to the argument that is being made.

There is nothing wrong with an individual having the opinion that AI is evil. People have opinions. 

In the interest of going beyond opinion and being productive; if the conversation is about water then the scope has been defined.

Vezina is the CEO of Prepared Canada Corp. and is the author of Continuity 101. He can be reached at info@prepared.ca.