See my other comment, this is likely only for the prefill GPU pool, a normal 1M token session would work fine. I think Cloudflare should probably link an explainer on how model serving works at the top of the post because a lot of people are not familiar at that level of detail.