Each bidder endpoint allows finite QPS. Every low-value call displaces a high-CPM one. At processed-auction-request, modules score bid probability per bidder -- publisher, geo, device, fill history -- and prune unlikely bidders from imp[].ext.prebid.bidder to pack the QPS budget with highest-probability, highest-CPM.
FIELD NOTES
- Low-fill traffic displaces high-CPM slots in the shared QPS budget
- ML shaping cuts ~20% of calls with near-zero revenue loss
- No native QPS cap -- most hosts limit at the load balancer