Meta turns idle inference capacity into better ads recommendations

Meta says it has built a system that uses idle inference capacity across its fleet to improve ads recommendations without new capacity orders.

The system relies on two elasticity levers, Rank More Ads (RMA), which evaluates more candidate ads when capacity is available, and Serve Better Models (SBM), which routes a share of requests to a more powerful model.

Meta says the approach protects baseline latency and reliability when opportunistic capacity is reclaimed.