Meta has published an engineering post describing Intelligent Delivery, an ML-driven system it uses to allocate compute across its ads serving stack. The company frames the problem as ads ranking models growing toward LLM scale faster than serving hardware can keep up, so each unit of compute has to go where it does the most good. [1]
Meta says the system has contributed roughly 10% incremental ads ranking performance on major surfaces, on top of other model and system advances, from 2017 through the first half of 2026. The figure is Meta's own and is not independently verified; the post does not specify the metric behind it. [1]
At its core is what Meta calls the Resource Core Model, deep learning models with 6 billion sparse and 20 million dense parameters that run once before candidate retrieval and are fine-tuned into specialised predictors of outcome, compute use and latency. Meta says this approach improved identification of the highest-impact ad requests by about 20% in offline benchmarks versus its previous production model. [1]
Allocation is split into a hierarchy of decision engines: shifting traffic between data center regions on a cycle of minutes, matching workloads to hardware across CPUs from Intel, AMD and Arm and accelerators from NVIDIA, AMD and Meta's own MTIA, and deciding per-host how much compute each request gets under a tail-latency target. [1]
Meta also says it increasingly uses AI agents to tune the system. According to the post, agents contributed to about 20% of Intelligent Delivery optimizations in the first half of 2026 and about 80% of engineering effort on its most automated mechanism, with engineers reviewing and approving every change. Meta says this shortens the optimization loop from months to weeks. [1]
Looking ahead, Meta says it wants to fuse request understanding and serving decisions into a single real-time model, and to rework its compute levers for GPUs as large ranking models such as its Generative Ads Model (GEM) increasingly run on them. [1]