This is a dangerous chart to take at face value. There are two important considerations that don’t seem to be addressed in this graphic 1. Offer data for these GPUs do not reflect the true transaction by costs. Almost all customers / tenants receive some discount on the sticker price for any one of a number of reasons. 2. There is a danger in aggregating offers across hyperscalers, neoclouds, etc because H100, B200, etc comparisons are not apples-to-apples. eg the hyperscalers are able to charge much more in theory because of their software layers, kubernetes, etc. We @OrnnExchange have access to a better data set that shows how on-demand chip pricing has changed vs time, reflecting market supply/demand dynamics. We will be posting supplementary research soon.