5% is probably optimistic.
The big problem is the <5% of non-nonsense applications are the ones that were being actively used before the 'deep learning' boom, mainly machine vision and data filtering.
Large Language Models, where the vast majority of investment is currently going, have basically no utility beyond toys. Image generation is in Photoshop Era (or if you're older, the digital drum-machine era) of artistic panic: the bit where everyone thinks it will take their jobs by having anyone push a button and receive art, and before the bit where everyone figures out that you still ned to be an artist to actually get anything of actual value out of it and it just becomes another digital tool that some artists will use productively and others will eschew for various reasons to various effect (e.g. the auteur directors who will still demand celluloid film vs. the many directors who cannot afford the massive costs of celluloid film). But even then, image generation is a pretty niche use, and will be relegated to mundane non-consumer-facing applications like creating tiling non-repeating rock textures from a small number of seed textures in game engines, or mapping user-generated-avatar expressions to face models without having to make sure the individual composited face elements can tolerate all possible poses.