For us as users, online stuff feels separate from the real world’s resources, but that’s not exactly how it works. The powerful servers running models like GPT-4 need a ton of energy and, just as importantly, a lot of cooling.
tens of millions of dollars well spent–you never know
— Sam Altman (@sama) April 16, 2025
So, how much juice do these chats really use? As noted in this report, data crunched by The Washington Post gives a pretty stark idea, especially about water use. Getting GPT-4 to generate a short, 100-word email can use about half a liter of water. That’s because the data centers processing our prompts get hot, and water cooling systems are needed to keep them from overheating. Scale that up across millions of users asking countless things daily, and the water and energy use skyrockets. We’re not even touching the huge amount of energy needed just to train these AI models in the first place.What makes this even more interesting is how it bumps up against what other tech companies have done. A few years back, Google added a “Pretty Please” feature to its Assistant. It was designed to encourage people, especially kids, to use “please” and “thank you” with the voice AI. The idea was to help kids learn good manners, hoping it would carry over into how they talk to people. Google was basically rewarding politeness. Now, OpenAI is showing us that the very politeness Google wanted to encourage comes with a real-world resource cost.It definitely gives you something to think about. Encouraging respect towards AI seems like a good social goal. But the costs OpenAI mentioned bring up questions about efficiency and environmental impact.
So, in my mind the question is: is the processing power spent on understanding our “pleases” worth the reinforcement of good habits, or should we be more direct with our digital helpers to save resources? It definitely adds a new wrinkle to the whole conversation about AI sustainability and how we interact with these tools.
