Published: May 22, 2026
For someone who chooses not to use generative AI, I certainly think about its implications quite a bit. The recent phenomenon of regular people rebuking AI data centers gives me quite a bit of hope. I do have a word of caution, however, and of course it has to do with semantics.
Let's back up first and talk about how this stuff even works. Generative AI models rely on GPUs, or "graphics processing units", for the bulk of their operation. At some point over the past 25 years, we figured out that the specialized chips on the video cards people would install in their gaming PCs were really good at specific mathematical operations and parallel processing, so their uses have expanded. One of the most famous examples of a benevolent use has been the long-running Folding@home project, utilizing excess capacity[1] to simulate biological processes for medical research. For a while, GPUs were also the primary mechanism behind most blockchain implementations due to their speed at performing cryptographic hash functions. Manufacturers like Nvidia met the demand in these markets by gradually decoupling the chip itself from its original use - which was real-time graphics and image processing - and building new products from the same architecture tailored to these computational uses.
Essentially... AI data centers are buildings stuffed to the brim with gaming PC video cards, except these things have been force-fed a diet of nails, broken glass, steroids, and protein supplements. These GPUs are only useful for doing a lot of math very quickly. To be fair, they're quite well-suited to this task.
To accomplish the task, however, the type of GPUs found in AI data centers suck down an astounding amount of energy. The average LED light bulb you'll screw into a lamp on your nightstand dissipates perhaps ten watts of energy while it's on. The combined compute capacity of the GPUs in an AI data center are literally measured by how many megawatts (or gigawatts) of energy they consume. A single megawatt - or one million watts - is roughly equivalent to 100,000 of those light bulbs turned on all at once, and a gigawatt is 1,000 megawatts.
I find it peculiar that in the late 2010s, the prevailing narrative in discourse about electric cars was that the power grid didn't have the capacity to charge all these cars. If you owned an electric vehicle and charged at home overnight, a good middle-of-the-road home charging station could draw maybe 10 kilowatts from the grid while active. A 1.5-gigawatt AI data center would continuously draw enough energy to charge 150 million electric cars at once. According to US Department of Transportation data[2], as of 2023, there weren't even 100 million passenger automobiles actively registered in the United States. On top of that, an EV stops drawing energy from the grid once its battery pack is fully charged. These AI data centers don't have any plans to shut off.
It looks like we already had the required capacity to completely electrify American transportation, and that old talking point was a complete lie. However, if you're wondering why your energy prices are still going up even though we clearly didn't have an overall energy constraint, you have to remember that a 1.5-gigawatt AI data center is concentrating the energy use required to charge every electric car in America (if every car on the road today was electric) in a single location. How many of these things do they want to build, again?
We need to take a moment to talk about the water use argument against AI data centers. While all these GPUs do require massive amounts of water for cooling, this use doesn't contaminate the water in any meaningful way. This week, a video showing AOC holding up a jar of contaminated water from an AI data center at a Congressional hearing has been circulating on the internet. It's important to note that this is runoff from the construction of the facility and not its operation. Construction-related water contamination is a problem that is not unique to AI data centers, but that doesn't mean it's not worth solving.
When water is used for cooling, it's passed through some sort of heat exchanger and then discharged or circulated. Heating up water doesn't contaminate it, but such a temperature difference may make nearby waterways less hospitable to aquatic life. To be clear, this problem is also not unique to AI data centers. Many nuclear power stations also use water for cooling[3] - that's also why they tend to be located near lakes, rivers, or oceans. Over time, several of these power stations have built retention ponds or other infrastructure to ensure this water has a chance to cool down itself before finding its way back to the waterway it came from.
Despite the shaky foundation of the water use argument against AI data centers, the energy requirements are terrifying. I truly believe it's good to oppose this wasteful use of resources.
When opposing or making arguments against these AI data centers, it's critically important that you specify "AI data centers" and to tell you why, we need to dig into some light historical details.
The generic term "data center" has been around for many years. It's used to describe any facility housing lots of computers to provide services on the internet. Since before the era of generative AI, data centers have been critical infrastructure on the internet. Just about every web site you've ever visited in your life has been hosted on a server in a data center[4].
What's the difference? Traditional data centers hold - for all intents and purposes - regular-ass general purpose computers. Sure, server hardware tends to be a great deal more powerful than your typical desktop or laptop PC, but they (mostly) run the exact same architecture with an Intel, AMD, or Arm-based CPU and the same supporting peripherals. While a GPU is only really good at graphics and math, a traditional CPU can perform pretty much any task you throw at it... that's why we use them for almost everything else.
When organizing any sort of opposition to AI data centers, we must qualify the "AI" part of it. It would not be good to entangle traditional data centers in the rhetoric around generative AI. These traditional data centers and internet exchange points have been fundamental to the evolution and continued existence of the parts of the internet that have brought us together, and they will be critical to reclaiming the artisan and organic nature of the internet that thrived in eras past. It's not a stretch to think those fueling this hype cycle are now counting on this ambiguity to cause lots of collateral damage when the party is ultimately over for them.
I know this may seem like useless "GNU/Linux" style pedantry[5], but there's this old saying about babies and bath water....
read: gaming PCs when not being used to play video games ↩︎
Here's a link to the query I used ↩︎
Water cooling is what those stereotypical curvy-shaped towers are for. ↩︎
As of this writing, even this very blog is hosted from a (traditional) data center in St. Louis, Missouri. ↩︎
Now I think I need a shower. ↩︎