had the opportunity of playing around with b10 inference. you can tweak the min and max replicas, and concurrency targets, and h100's cost .133 dollars per minute, which adds up to ~191.5 per day per replica. if you spin up 20 of these it's 3800 a day. is that expensive or not. i'm not sure. but that is vc money being put to good use for sure, especially given the use case.