Turning Opportunistic AI Inference Capacity Into Better Ads Recommendations
Meta built a system that repurposes idle inference capacity to improve ads ranking quality.
Meta Engineering describes a production system that aggregates locally idle GPU/accelerator capacity across its fleet into a shared opportunistic pool used to run heavier ads ranking models and evaluate more candidates. Two elasticity levers—Rank More Ads and Serve Better Models—expand or contract workload dynamically as capacity is reclaimed, with isolation and graceful degradation protecting baseline latency. The engineering is notable for its scale but represents incremental infrastructure optimization rather than a broader industry shift.