Meta's Intelligent Delivery adds ~10% ads ranking gains

Meta's post frames a structural problem: ads ranking models are growing toward LLM scale faster than the hardware that runs them can keep up. The same extra unit of compute produces very different returns depending on the surface, user intent, candidate quality and system state, and local fixes do not add up to a global optimum. Intelligent Delivery is Meta's answer: an ML layer that estimates marginal utility per request, paired with a hierarchy of decision engines that shift traffic between regions, assign workloads to the right chip among Intel, AMD, Arm CPUs and NVIDIA, AMD and MTIA accelerators, and decide how much of each host a request gets.

The results are presented as cumulative: about a 20% improvement in identifying the highest-impact ad requests in offline benchmarks relative to the previous production model, and roughly 10% incremental ranking performance overall. Meta also describes a shift in how it finds configurations, moving from manual engineering and a constrained optimization approach to AI agents that read source code, design documents and a decade of experiment logs, then propose changes for engineer approval. That shortened the optimization loop from months to weeks, according to the post.

The forward-looking parts point to two open problems: fusing request understanding and serving into a single real-time model, and adapting the allocation levers to GPUs, where capacity comes in whole fixed blocks rather than divisible CPU time. Both suggest Meta expects the next round of gains to come from operating models more efficiently, not only from making them larger.