COBURNICUSThis film started with a gap that surprised me.
The computers that train a powerful AI model can fill a building and run for weeks or months. When they finish, the result is a set of files you can carry on a portable drive. Thousands of expensive chips go in. Something you can hold in your hand comes out.
So why does everyone keep building bigger data centers? And can AI keep getting better without eating more electricity, water and raw materials every year?
Millions of hours, one drive
Meta’s Llama 3.1 405B is a good example because Meta published the numbers. According to its model card, the model was trained on more than 16,000 NVIDIA H100 processors for about 30.84 million GPU-hours. A GPU-hour is one graphics processor running for one hour.
The finished model has about 405 billion parameters, the tiny internal dials that training tunes. Stored at 16-bit precision, all of them take roughly 810 gigabytes. That fits on a one-terabyte portable SSD.
Training is how you find the weights. The weights are what you keep. ### The part that never stops
Training happens once. Inference, the work of running the model every time someone asks it something, happens billions of times.
One question is small. In August 2025 Google reported that the median text prompt to its Gemini app used about 0.24 watt-hours of electricity and about a quarter of a milliliter of water. A 2026 study in Joule by researchers at Microsoft landed in the same range for very large models, about 0.31 watt-hours for a typical query.
The same study found that a long reasoning question, where the model works step by step before answering, used about 3.9 watt-hours. That is more than ten times as much. Video, images and agents that run for hours use more still.
Nobody asks one question. That is why the building spree keeps going.
The cloud has weight
The International Energy Agency estimates that data centers worldwide used about 485 terawatt-hours of electricity in 2025 and expects about 950 by 2030. That covers all data centers. The IEA says AI-focused use roughly triples over that stretch.
Water is the other cost. Lawrence Berkeley National Lab estimated U.S. data centers used about 17 billion gallons of water directly in 2023. The water used to generate their electricity was around 211 billion gallons.
In southwest Memphis, xAI built a supercomputer called Colossus and brought in natural gas turbines on trailers to power it. The Southern Environmental Law Center said aerial photos in April 2025 showed 35 turbines at the site, next to Boxtown, a mostly Black neighborhood that already lives with heavy industrial pollution. The county permitted 15 turbines in July 2025. At a second site in Southaven, Mississippi, the NAACP sued in April 2026 over 27 turbines it says ran without permits. xAI argued they were mobile equipment. On October 30, 2026, a federal judge rejected that defense, and the case continues.
In Newton County, Georgia, Beverly and Jeff Morris live about a thousand feet from a Meta data center. They say their water pressure dropped and sediment started coming from their taps after construction began. They have spent about $5,000 trying to fix it and can’t afford a roughly $25,000 new well. Meta says its own study found it “unlikely” the data center affected local groundwater. Nobody has proven it either way. The facility uses about half a million gallons a day, roughly ten percent of the county’s use.
Your electric bill
PJM runs the grid for much of the eastern United States and holds an annual auction for future power capacity. The price went from about $28.92 per megawatt-day for 2024/25 to about $269.92 for 2025/26. It has sat at or near the cap since.
PJM’s independent market monitor found that without data center demand, the 2025/26 auction would have cost about $9.3 billion less. Across four auctions, it attributes about $29.4 billion of $63.6 billion in costs to data centers. That money is spread across everyone’s bills in the region.
The other side
The film also looks at the counterweights.
Big tech companies are signing nuclear deals. In September 2024 Microsoft signed a 20-year deal with Constellation to restart a Three Mile Island reactor, now the Crane Clean Energy Center, targeted for 2027 and still awaiting a federal regulatory decision. Google, Amazon and Meta have their own nuclear deals. Most of those new reactors are years away. In the meantime a lot of new demand is being met with natural gas.
And the water story has two sides. One analyst, Andy Masley, estimated that data centers in the Phoenix area used about 905 million gallons of water in 2025, against about 29 billion for golf courses. It is one person’s estimate. Arizona Governor Katie Hobbs still proposed in January 2026 that data centers pay the same penny a gallon families do. The legislature has not passed it.
Computing also keeps getting more efficient. A 2025 study of the world’s top supercomputers by Prieto and colleagues found that the computing you get per unit of energy doubled about every 2.3 years from 2008 to 2023. The historical Koomey’s Law pace was about every 1.6 years, so the gains are still coming, more slowly.
The coal question
In 1865 William Stanley Jevons noticed that better steam engines didn’t reduce Britain’s coal use. They made steam power cheap enough to spread everywhere, and coal use went up. He called the idea that efficiency means using less fuel “wholly a confusion of ideas.”
That is the question at the center of the film. If each AI task takes half the electricity but the world runs four times as many tasks, total use doubles. The electricity per task can fall while the total keeps climbing.
What didn’t make the cut
Two numbers from the research stayed out of the film. Berkeley Lab’s 2025 update projects U.S. data centers could use about 11.8 percent of the country’s electricity by 2030, within a range of 9.5 to 15.3 percent. In 2023 the figure was 4.4 percent.
And the IEA’s October 2026 report found that more than a third of identified AI data center capacity sits in areas of high or extremely high water stress. It also estimates data center electricity produced about 150 million tonnes of CO2, rising to around 350 million by 2035 in its central case.
Where I land, and what I don’t know
One thing about this still amazes me. Somebody spends a fortune training a model, and the result can be handed to independent developers, researchers and small businesses who could never have paid for that training. The people I’d bet on in AI are the ones getting the most useful work out of the least computing.
What I can’t answer is when the two lines cross. Efficiency per task is falling. The number of tasks is rising, and the tasks keep getting bigger. Will efficiency outrun demand, and if it does, will the towns hosting the servers get a say in what it costs them along the way?
Sources
- Meta, Llama 3.1 405B model card (2024): https://huggingface.co/meta-llama/Llama-3.1-405B
- Google, Measuring the environmental impact of AI inference (Aug 2025): https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference
- Oviedo et al., “Energy use of AI inference, efficiency pathways, and test-time scaling,” Joule (2026): https://arxiv.org/abs/2509.20241
- IEA, Key Questions on Energy and AI (April 2026) and Powering Artificial Intelligence (October 2026): https://www.iea.org/
- Lawrence Berkeley National Lab, U.S. Data Center Energy Usage Report (Dec 2024) and 2025 Update (June 2026): https://eta-publications.lbl.gov/publications/united-states-data-center-energy-2025
- Southern Environmental Law Center on xAI in Memphis: https://www.selc.org/news/inside-memphis-fight-against-xai/
- The New York Times / Press Democrat on Meta’s Georgia data center (July 2025): https://www.pressdemocrat.com/2025/07/15/meta-built-a-new-data-center-in-georgia-their-neighbors-water-ran-dry/
- PJM capacity auction reports: https://www.pjm.com/markets-and-operations/rpm
- Monitoring Analytics (PJM independent market monitor): https://www.monitoringanalytics.com/
- Prieto et al., “Evolution of computing energy efficiency: Koomey’s law revisited,” Cluster Computing (2025): https://doi.org/10.1007/s10586-024-04767-y
- Arizona’s Family, “Data centers aren’t the water villains you think they are” (Feb 2026): https://www.youtube.com/watch?v=_K0HDa4c1nY
- W. S. Jevons, The Coal Question (1865)
Subscribe to get the next film and the research behind it. The full film and the rest of the documentaries are on the YouTube channel: https://www.youtube.com/@coburnicusmind?sub_confirmation=1