In our recent workshop from TU Graz at the Austrian Cultural Forum in Warsaw on AI literacy and prompting, one question came up: What is the environmental cost of a prompt? It ties in with one of the myths in our preprint Seven Myths of AI Use: that AI is environmentally harmless.

Cloud AI: small numbers, big scale

Google reports 0.24 Wh of energy and 0.26 ml of water per Gemini text prompt; OpenAI states 0.34 Wh for an average ChatGPT query. These figures come from the providers themselves and exclude training, images, network transfer and indirect water use. Independent estimates range up to 2.9 Wh per query, and billions of daily prompts add up.

Google reports 0.24 Wh of energy and 0.26 ml of water per Gemini text prompt; OpenAI states 0.34 Wh for an average ChatGPT query. These figures come from the providers themselves and cover only text prompts, excluding training, images, network transfer and indirect water use. An early academic estimate in Joule (de Vries, 2023) put a ChatGPT request at up to 2.9 Wh. Today’s standard queries are likely far below that, but long reasoning answers can again reach several watt-hours. And billions of daily prompts add up.

Local AI: transparent and water-free

On oer.one/ki-lokal we share open guides for running models on your own device with Ollama or Llamafile. A good example is gpt-oss-20b: only about 3.6 of its 20.9 billion parameters are active at a time. Measurements by Hugging Face show around 2 Wh for a 100-token query, about eight times less than a comparable 30B model. A Stanford study found that local models answer 88.7% of chat queries accurately, and that combining local and cloud models can cut energy use by 40–65%. And since a laptop is air-cooled, no cooling water is needed.

Takeaway

What does this mean for a workshop? If twenty participants send 50 prompts each, that makes 1,000 prompts. Based on the providers’ own figures, this amounts to about 0.24–0.34 kWh, roughly the energy needed to boil a kettle three times. With frontier models or long reasoning answers, it could be closer to 4 kWh. For comparison, 20 laptops (50 W each) running for a four-hour workshop use around 4 kWh themselves (our rough estimate). The real driver is not a single prompt but scale, answer length and reasoning.

Local is not automatically greener per query, but it makes energy use visible and builds a valuable habit: a small model for a small task. That is AI literacy in practice.