Large language models are now everywhere. Whether you consider them helpful or an intrusion, it’s up to you, but everything from Google searches to mundane interactions with an e-commerce site now features an AI component.
And, I realized, most of it is just… well, a form of ephemeral e-waste. I just did a Google search, wanted to get the exact syntax for a curl command example to interact with an AI through the API. Google helpfully responded with a page-long AI-generated answer, which included an example. I copy-pasted the example, edited it to my liking, and used it… and never even read the rest of Google’s response before closing the tab. Perhaps Google caches such responses so they’re not totally wasteful. Otherwise, those words headed straight from GPU to the bit bucket, never read by a human.
They may not have been read by a human, but they were still generated on a GPU somewhere, consuming non-trivial quantities of power. How much power? Well, take my setup. A modest 16 GB GPU, consuming about 180 W of power when running at full steam. It is running the open-weights model Gemma, producing a tad fewer than 50 tokens, or about 30 words per second. In other words, each word consumes about 6 joules of electrical energy. That — taking into account power generation and transmission losses — is very roughly the electrical energy extracted from one milligram of coal in a conventional power plant. Not a lot… just a milligram. But it’s a milligram for every word produced, whether or not it is read.
Scale it up to a planetary scale, say, a billion users and 1000 unread words per user per day, and that’s the equivalent of 1,000 metric tons of coal per day… just to produce words that no one ever reads.
Yes, caching responses helps. Using more efficient algorithms helps, as well as more advanced hardware. Nonetheless, this little calculation perhaps illustrates why our AI boom produced an unprecedented demand for electrical power for the new data centers.