I'm not familiar with all the varied uses of GPUs but it seems like image generation could feasibly be distributed: large upfront download of models, then small inputs of text and settings, and small output of resulting images.