Does that water usage take into account all the water that went into training the models, including the electricity it took run the data centers the GPUs are housed in?

Training is part of why there's two orders of magnitude range in the estimates. However, I believe yes, they do include energy production impact.