from someone who runs AI inference pipelines for video production -- the cost per inference is what actually matters to me, not raw speed. right now i'm paying ~$0.003 per image generation and ~7 cents per 10-second animation clip. a full video costs under $2 in compute.

if dedicated ASICs can drop that by 10x while keeping latency reasonable, that changes the economics of the whole content creation space. you could afford to generate way more variations and iterate more, which is where the real quality gains come from. the bottleneck isn't speed, it's cost per creative iteration.