Whoops, looks like ChatGPT got too thorsty and drank all the water in California
Watch the short here: https://youtube.com/shorts/0dUTFR-eCAA
Table of Contents
Additional context
This is a very complex topic, and I didn’t have the time to research nor write about all of it. If you’re here, it means you’re probably looking for extra context, so here are the main things I want to clarify:
- The 12.3 billion litres for power generation is a rough estimate. A lot of these tech companies produce some of their own power, and almost all of them technically run on “100% renewable energy” (but that’s mostly accomplished through Power Purchase Agreements (PPAs) and other things that make it work on paper, but it’s not literally true). It also uses some averages from Jin et al. (2019), some of which have broad data points (namely hydropower).
- All power generation plants and all data centers are different. They all have different means of cooling, which means different rates of water consumption. But my point for this video isn’t to get into the nuance, it’s to point out that in the big picture, this is a problem and will continue to be for the foreseeable future unless we push back.
- Funny enough, I’m not actually 100% on where I stand with AI. I don’t like it under capitalism because it’s hyper-exploitative of artists and the environment, but I’m not totally sure if I believe it’s “plagiarism.” What I do know is that most AI models (that get used) were trained on data without the artists’ consent, and I am NOT a fan of that.
- Minor thing, I accidentally referenced the same EPRI article as both 1 and 8 in the video. That’s all.
Transcript
- California is on fire again, and this time, part of the blame may be on AI and ChatGPT.
- You may have heard some talk about how AI uses so much water, but how does it do that? Why is the computer so thirsty? I thought water was bad for computers?
- The answer is that water is consumed at multiple steps in the process of plagiarizing the hard work of artists (Wang, 2024)!
- But even before Chat GPT steals my juicy tumblr Supernatural fanfic, water is consumed in manufacturing all the GPUs and other hardware that AI servers need to function.
- Then, once the servers are up and running, most large datacenters have water cooling towers on the roof that consume lots of water. These work by basically transferring the heat energy from the servers into the water and letting it evaporate off (Li et al., 2023; Mytton, 2021).
- Inside the datacenter, the computers turn electricity into plagiarism; and generating that electricity takes water too! Most electricity is generated by steam generators. Fossil fuels are burned to heat water and generate steam. The steam rises and spins turbines to generate power, but then the steam needs to be cooled down. So a lot of facilities use water to cool the water.
- A lot of plants use “once-through” cooling systems, where they pull cool water from a river or lake, but then they release that water back after it’s absorbed some of the heat. These systems have high water “withdraw,” but they actually have very low water “consumption.”
- The steam that’s in the loop is considered “consumed,” but that’s not much because it’s in a closed loop, so it’s recirculated. Most of the consumption comes from plants that use evaporative cooling– that’s like the cooling towers on the data centers (Jin et al., 2019; Rathi, 2018; Union of Concerned Scientists, 2010).
- But that’s just steam generators. Big picture, how much water and electricity does AI consume, and how might that have affected the California wildfires?
- Well, per-request, AI uses about TEN TIMES as much electricity as a regular Google search, and that’s only for LLMs like ChatGPT. AIs that generate images, video, and music take WAY more than that (EPRI, 2024).
- For every watt of electricity used, heat is generated that you need water to cool. That’s why energy usage and water usage are so tightly correlated.
- And both of them are growing pretty quickly at every major tech company, with Google leading the way, consistently growing their usage of water and electricity about 15-20% year-over-year (Apple, 2024; Google, 2024; Meta, 2024; Microsoft, 2024)
- It’s not hard to draw a line between the rise in AI and these growing numbers, but Google themselves even attribute a lot of this growth to expansion of AI services (Google, 2024).
- Bringing it back to California, we can use math and numbers to estimate that data centers in California consumed about 12.3 BILLION liters of water from energy production alone, so that’s not counting server cooling.
- And this comes during a time when water scarcity is increasing and much of the western US is in moderate and severe drought conditions.
- With all that dry air and soil, that’s why [loop] California is on fire again…
Resources
Charts
Tech company water consumption per year

Tech company electricity consumption per year

Code
Here’s my messy code that I used to calculate the water usage of California data centers
california_gwh = {
# gwh
"coal": 4981,
"gas": 102774,
"nuclear": 26272,
"hydro": 32886+4988,
"geothermal": 13567,
"solar": 47869,
"wind": 31399
}
water_per_mwh = {
# litres per mwh
"coal": 2220,
"gas": 598,
"nuclear": 2290,
"hydro": 4961,
"geothermal": 1022,
"solar": 330,
"wind": 43
}
def calc_water_consumption(energy_in_twh, state_power, water_rates):
"""
Take in stats (state_power = power generation sources for a given state, water_rates = water consumption rate (litres per mwh) for each type of generation)
and TWh and calculate how much water (litres) is consumed by producing that much energy
"""
ret = {
"twh": energy_in_twh,
"total": 0
}
energy_in_mwh = energy_in_twh * 1000 * 1000
state_power_sum = sum(state_power.values())
state_power_prc = {}
for name, value in state_power.items():
state_power_prc[name] = value / state_power_sum
wc_per_mwh = 0
prc_accounted_for = 0
for name, wc_rate in water_rates.items():
this_prc = state_power_prc.get(name, 0)
prc_accounted_for += this_prc * 100
ret[name] = this_prc * wc_rate * energy_in_mwh
wc_per_mwh += this_prc * wc_rate
missing_types = (set(state_power.keys()) - set(water_rates.keys()))
if len(missing_types) > 0:
print(f"WARNING: {missing_types} are not fully accounted for")
print(f"{prc_accounted_for:.2f}% of state energy accounted for")
ret["total"] = (energy_in_mwh * wc_per_mwh)
return ret
def pretty_wc(wc_from_energy):
total = wc_from_energy.pop('total')
twh = wc_from_energy.pop('twh')
print()
print(f"Calculated values for {twh} TWh")
for name, value in wc_from_energy.items():
print(f"Litres from {name}: \t\t{value / (1000**2):.2f} million")
print(f"---------------\nLitres total: \t\t{total / (1000**2):.2f} million")
pretty_wc(calc_water_consumption(9.31, california_gwh, water_per_mwh))
Image credits
- House on fire: Josh Edelson/AFP/Getty Images
- GPU manufacturing: PC Games N
- Steam Turbine Generator diagram: Unknown
- Water towers: SPX Cooling
- Water tower diagram: US Office of Energy Efficiency & Renewable Energy
- Drought map: University of Nebraska-Lincoln / NOAA / droughtmonitor.unl.edu
References
(3) Apple. (2024). Apple Environmental Progress Report. In Apple. Apple. https://www.apple.com/environment/pdf/Apple_Environmental_Progress_Report_2024.pdf
(6) California Energy Commission. (2023). 2023 Total System Electric Generation. California Energy Commission. https://www.energy.ca.gov/data-reports/energy-almanac/california-electricity-data/2023-total-system-electric-generation
(1, 8) EPRI. (2024). Powering Intelligence: Analyzing Artificial Intelligence and Data Center Energy Consumption. In EPRI. EPRI. https://www.epri.com/research/products/000000003002028905
(2) Google. (2024). Google 2024 Environmental Report. In Google Sustainability. Google Sustainability. https://www.gstatic.com/gumdrop/sustainability/google-2024-environmental-report.pdf
(7) Jin, Y., Behrens, P., Tukker, A., & Scherer, L. (2019). Water use of electricity technologies: A global meta-analysis. Renewable and Sustainable Energy Reviews, 115, 109391. https://doi.org/10.1016/j.rser.2019.109391
Li, P., Yang, J., Islam, M., & Ren, S. (2023). Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models. https://arxiv.org/pdf/2304.03271
(4) Meta. (2024). 2024 Sustainability Report. In atmeta.com. Meta. https://sustainability.atmeta.com/wp-content/uploads/2024/08/Meta-2024-Sustainability-Report.pdf
(5) Microsoft. (2024). 2024 Environmental Sustainability Report. In Microsoft. Microsoft. https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/microsoft/msc/documents/presentations/CSR/2024-Environmental-Sustainability-Report-Data-Fact.pdf
Mytton, D. (2021). Data centre water consumption. Npj Clean Water, 4(1). https://doi.org/10.1038/s41545-021-00101-w
Rathi, A. (2018, August 8). You probably have no idea just how much water is needed to produce electricity. Quartz. https://qz.com/1351279/the-hidden-water-footprint-of-fossil-fuel-and-nuclear-power-plants
Union of Concerned Scientists. (2010, September 30). How it Works: Water for Electricity. Union of Concerned Scientists. https://www.ucsusa.org/resources/how-it-works-water-electricity. Updated 9 Nov, 2017.
Wang, Z. (2024, January 22). How AI Consumes Water: The unspoken environmental footprint. Deepgram. https://deepgram.com/learn/how-ai-consumes-water