[[Preamble - I'm just a regular human and reserve the right to make mistakes and to change my opinion in light of new information. Just because I wrote it here once doesn't mean I'll always believe it or assume I'm right.]]

Since I'm running into a number of threads on Bluesky with a mix of semi-credible to barely googled sources, I thought it best to figure out what is going on with data centers and if are they on the right track.

The IEA produced a report on energy and AI use in 2024. In it they mention that data centers used about 415 TWh of all energy produced. If you read between the lines and did the calculation yourself, you'd also find that they estimate that AI accounted for about 15% of that use or 62.25TWh.

It's important to keep a sense of scale. Global energy generation was 30,857 TWh in 2024, up 4% from 2023. Data centers as a whole account for 1.5% of all consumption. AI by extension accounts then for 0.2% of global energy consumption.

On the whole, manufacturing uses the majority of energy produced globally.

I took the time to read through the sustainability reports of a number of companies, some audited by firms like Deloitte, others not. I wanted to have a comprehensive picture of how is FAANG (or MANGA with the Meta rebrand) handling their use of electricity as well as water?

None of the reports have a wholly consistent reporting standard between them, so it took some spreadsheetin' and math to figure some things out. If I got something wrong, do feel free to cite your source and I'll do my best to make a timely correction with attribution.

Sources

I looked at the largest companies in the space and tried to find the most up-to-date sustainability reports they could provide. Even if the report was published in a certain year, often they follow the fiscal year, so do keep that in mind.

Nvidia - 2025. 2024

Amazon - 2024

Netflix - 2024 2023

Microsoft - 2024 2024 Azure

Meta - 2025

Google - 2025

less helpful, but found

Apple - 2024

JPMorgan - 2024

Oracle - 2025

Broadcom - 2024

One thing I was trying to figure out is how could I find a reasonable metric for expected water use if AI keeps booming?

It made sense to me that it should be some ratio relative to the amount of power used and if I could find a kind of average value across data centers I could extrapolate for the industry.

As I read over the reports, I found the industry already has a term for this, WUE. (Mircosoft's Azure report in the Sources explains it pretty well.) It's a ratio of liters to kilowatthours.

Data centers use water to cool their machines (not directly, they use fancy heat exchangers and whatnot), but there are two major varieties. Closed loop and open loop. Open loop are the ones where water is pulled in from a local source (typically a stream or river) and is passed through the heat exchanger. The water then goes to a cooling tower and a certain amount is turned into steam. Sometimes that same water is also used to wash the towers as minerals build up from evaporation. The water is never 'destroyed', but the steam is inaccessible to local use until it condenses into rain and falls somewhere else on Earth.

It's important then for data centers to have a WUE that is approaching zero. Some data centers in cold regions like Svalbard get very close to that! However, we also need data centers close to where some operations exist. For Fintech like quantitative trading, even a few hundred feet makes all the difference. Some companies, like Nvidia are are moving to more closed loop systems which use a mix of water and coolants, but do not involve local water systems.

Aside from WUE, I also looked at total water usage versus total water discharge to find a water source efficiency ratio. In that case, closer to 1 is better. For every liter you take out of the river or stream, you should be putting just about that same amount back in after getting it cooled back down to ambient waterway temp and making sure it's treated to avoid adding unnecessary chemicals. Some companies like Microsoft, Google, and others are aiming for net positive water creation by 2030.

Lastly, I wanted to look at both the trend of electricity use and the sources of electricity generation. Are these companies producing their own electricity on-site or purchasing carbon offsets or doing direct purchases from solar and wind farms?

I wanted a ratio of renewable energy purchased/used versus total energy used. Turns out, the industry already has a term for that too, PUE.

Fudge Factor

As these are largely technical companies and the reporting methods inconsistent, I used total power and did not separate out data center specific consumption. That will skew results by about 5% in some cases, but this is for an informal post, not a PhD dissertation. This also impacts the WUE as a proper calculation is the total annual data center water consumption divided by the total annual IT equipment energy use.

While Google, Amazon, and Microsoft represent some of the most well known large cloud providers, I included Nvidia as well as it straddles manufacturing with AI and is critical to the growth of AI.

As someone who enjoys fishing every time I visit my hometown and the occasional canoe trip or rope swing, water use is a pretty strong concern of mine. Especially since we've seen the results of strip mining and fracking have had on killing off local wildlife or making some natural sources nonpotable.

WUE trends:

Nvidia has seen steady progress since 2023 on reducing the amount of energy used per kilowatthour. Amazon has had a solid trend as well as has Microsoft and Meta. Google on the other hand does not look as good.

The calculation across all companies was consistent as it look their total water usage in Ml and divided by their total electricity usage in GWh.

Here are the most recent scores I could generate based on available information (remember, lower is better):

Amazon - 2024 - 0.15 *(their WUE is listed, but raw numbers were hard to find)

Meta - 2024 - 0.305

Nvidia - 2024 - 0.499

Microsoft - 2023 - 0.539

Google - 2024 - 1.273

Leaving out the Amazon score due to a lack of supporting data, that would put the industry average near 0.654, or put in lay terms, for every liter that goes into the system, more than half doesn't come back out in the same spot.

Water source efficiency trends:

Nvidia saw a substantial drop in efficiency from 2022 to 2023, afterwards it has been on a recovery trend. Microsoft is slowly slipping in efficiency. Meta has had highly variable efficiency and Google has been incredibly consistent.

Here are the most recent scores I could generate based on available information (remember, the close to 1 the better and many of these companies have a goal to be greater than 1 by 2030)

Nvidia - 2025 - 0.654

Meta - 2024 - 0.445

Microsoft - 2023 - 0.394

Google - 2024 - 0.261

As mentioned in the IEA report, data centers account for 415 TWh of all energy consumed last year. Also, we can draw some kind of relationship between electricity consumed and water consumed as it is the electronic components which heat up with use create cooling demand.

Here are the usage statistics I could find, all adjusted to GWh:

Google - 2024 - 32,727.8 GWh

Microsoft - 2023 - 24,007.86 GWh

Meta - 2024 - 18,423.63 GWh

Nvidia - 2025 - 821.2 GWh

Netflix - 2024 - 158.03 GWh

So of the 415 TWh consumed in 2024, we can ballpark that Google, Microsoft, and Meta accounted for about 18%. Amazon, being the most obfuscated with respect to sustainability data is likely in the upper 20k to 30k GWh range as well.

What about the energy trend?

Here are the year-over-year trends in energy usage and percent renewable energy purchase or generated (negative growth is possible through more efficient systems):

Google - 2021 - 20% growth 98% renewable

- 2022 - 20% growth 97% renewable

- 2023 - 15% growth 97% renewable

- 2024 - 26% growth 98% renewable

Microsoft - 2021 - 25% growth 92% renewable

- 2022 - 26% growth 97% renewable

- 2023 - 28% growth 98% renewable

Meta - 2021 - 31% growth

- 2022 - 22% growth

- 2023 - 33% growth

- 2024 - 20% growth

Netflix - 2020 - 16% growth 100% renewable

- 2021 - 66% growth 100% renewable

- 2022 - .2% growth 100% renewable

- 2023 - -27% growth 100% renewable

- 2024 - 38% growth 100% renewable

Nvidia - 2023 - 16% growth 36% renewable

- 2024 - 23% growth 68% renewable

- 2025 - 34% growth 95% renewable

Many of these companies are working hard on sustainable computing, and it shows. There is more work to be done when it comes to direct renewable power purchasing, shifting cloud work to other regions throughout the day to avoid peak usage and take advantage of overnight electricity rates and cooler temperatures, etc. There is also a meaningful update to be had when FY 2025 closes and we have updated sustainability reports.

Despite being such a large entity in the space, it's disappointing to see Amazon lagging in clear reporting. I'd love to see Kara Hurst bring Amazon up to par with the industry.

None of the WUE's look great, but as long as Amazon isn't cooking the books, they'll be just behind Meta in hitting a 0 score before 2030. I'm genuinely more concerned about the water source efficiency. While companies state that they're 'moving toward' closed loop systems, we'll have to see over the next year or two how fast that will really be.

Is there a correlation between WUE and electricity use?

Using Pearson's r score, only Meta was consistent. That said, it's not a huge dataset, only a few years per company. What I might take away from this is since Meta is not a vendor for their data centers, it's an economic as well as ecological concern that pushes efficiency.

From an electricity perspective though, I found a lot of relief assuming the sustainability reports are to be believed. Most of these companies appear to be prioritizing renewable sources of electricity for purchase. That's a big deal and not what I expected.

What about growth?

In the 'Further Reading' section, I provide a few links that have projections about future demand. Based on historic trend, I'm expecting a year-over-year rate of about 25%. That would mean data centers as a whole go from 415 TWh in 2024 to 518 TWh this year, 648 TWh the next, and so on. For AI specifically, that would mean 62.25 TWh in 2024 to 77.8 TWh for this year, 97.2 TWh the next, and so on. That is if it's not all a bubble and growth is consistent. Big IFs in my opinion.

While I could do an entirely separate writeup on energy production and time to market, I will briefly note that solar year over year growth was 25% in 2022 according to the IEA and that NREL shows that in 2024 over 550GW of solar was added which implies a year of year growth of over 70%. Assuming data centers continue with their sustainability trend of purchasing a majority, if not all of their power, from renewable sources, that will continue to drive the industry.

What about water usage?

There isn't a consistent correlation between WUE and electricity use for major providers, so it's VERY difficult to project usage. That said, if we want to take the average of 0.654 Ml/GWh and apply that to the IEA numbers and my projections, here is what we get for the amounts consumed globally (not withdrawn or discharged):

2024 - 271,710Ml for all data centers, 40,711Ml for AI

2025 - 338,771Ml for all data centers, 50,881Ml for AI

2026 - 423,702Ml for all data centers, 63,529Ml for AI

Annual approximate renewable water resources per region (flow:

Americas (North + South) - 19,700,000,000 Ml/year

Asia (East and Southeast) - 12,250,000,000 Ml/year

Europe - 6,800,000,000 Ml/year

Africa - 3,940,000,000 Ml/year

According to the UN World Water Development Report human withdrawal of water was about 3,984,000,000 Ml globally.

That would mean data centers account for about 0.0068 % of global water use and that AI specifically accounts for 0.0010 %.

So... yes.

My personal takeaway, the handwringing about AI's environmental impact is really just an extension of the existential concern AI provides especially during a time when we have an absolutely shit economy and plenty of unemployment.

If we want to look at the big tent poles when it comes to resource use, for electricity it's manufacturing and for water it's farming (72%).

Can data centers do better when it comes to resource utilization and ESPECIALLY reporting? Absolutely. Is AI going to melt the polar ice caps, drink all our fresh water, and kill us all? No and it's not even close.

** It should be noted that I did not include anything related to decentralized AI as that merits its own research and many models are becoming powerful and compact enough to run on repurposed gaming rigs in the home.

Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models - Pengfei Li, Jianyi Yang, Mohammad A. Islam, Shaolei Ren

This research paper is thorough and has a very interesting set of projections when it comes to AI energy use which feels in-line with the extrapolation I made based on the energy and AI report from the IEA.

IEA Electricity 2024 - Analysis and projection to 2026 - a separate report that projects datacenter usage could grow to 1,000 TWh by 2026.