You’re confusing LLMs with other AI models, as LLMs are magnitudes more energy demanding than other AI. It’s easy to see why if you’ve ever looked at self hosting AI, you need a cluster of top line business GPUs to run modern LLMs while an image generator can be run on most consumer 2000, 3000, 4000 Nvidia GPUs at home. Generating images is about as costly as playing a modern video game, and only when it’s generating.