It is difficult to calculate the exact environmental costs of GenAI. The size of the model, the training approach used and the capabilities of the tool influence how much energy and water the model uses. Likewise, there are very different energy needs for training a model and for using it. That said, the energy costs of training and running the tools are estimated to be considerable.
One recent peer-reviewed study estimates that by 2027, the annual electricity usage of AI servers could equal the annual electricity usage of a country like the Netherlands or Sweden. This amounts to 0.5 percent of the world’s current electricity use.
As for water consumption - used for electricity generation and cooling of the data servers - according to one study , ChatGPT could use up to 500 ml of water per 5-50 prompts.
‘The environmental costs of generative AI’, Microsoft Designer.