Meta details Intelligent Delivery system for more efficient ad ranking

A compute allocation problem at LLM scale

Meta says ads ranking models are growing toward LLM scale faster than the hardware available to run them. Because an additional unit of compute can have different effects depending on the ad request, model, chip and system state, the company built Intelligent Delivery to direct resources toward the highest marginal utility.

The system is designed to operate across a serving environment with finite capacity and a tight latency budget. Meta says it has contributed an additional ~10% improvement in ads ranking performance across major surfaces, on top of model and system advancements, from 2017 through the first half of 2026.

The company describes Intelligent Delivery as an adaptive optimization layer that continuously directs resources across requests, models, hardware fleets and ranking stages. Its stated goal is for the delivery system to keep moving toward a global optimum as user behavior, models and the serving stack change.

Request-level predictions replace coarse segments

The first capability, Request Understanding, estimates which ad requests are likely to benefit from extra computation and how different resource settings may affect outcomes, compute consumption and latency. Meta says its earlier approach grouped requests using attributes such as surface, context and ad creative representation, with engineers creating policies for each group.

That rule-based method became difficult to scale as every new signal multiplied the number of rules. Meta says it improved the identification of the highest-impact ad requests by about 20% in offline benchmarks on major surfaces, compared with its previous production model, after reframing the task as two machine learning problems.

At the centre of this layer is the Resource Core Model. Meta describes it as a set of deep learning models with 6B sparse parameters and 20M dense parameters. The models run once before candidate retrieval and are shared across the delivery funnel, using learned embeddings and deep feature interactions to capture request patterns. Specialized models are then fine-tuned to generate signals about outcome, compute consumption and latency impact.

A hierarchy moves resources through the serving stack

Meta says one allocation mechanism cannot control its globally distributed and heterogeneous infrastructure end to end. Intelligent Delivery therefore uses a hierarchy of decision engines covering data center regions, hardware assignments and resource allocation inside individual hosts.

At the regional level, real-time infrastructure can shift traffic between regions in a cycle of minutes. The company says this accounts for differences in traffic, capacity, hardware mix and headroom, with the aim of making the marginal unit of capacity similarly efficient across regions.

Hardware selection is also workload-specific. Meta's fleet includes CPUs from Intel, AMD and Arm, accelerators from NVIDIA and AMD, and its own Meta Training and Inference Accelerator, or MTIA, across several generations. Since retrieval, ranking and model inference are bound by different resources, Meta benchmarks efficiency for each workload and hardware pairing before assigning scarce capacity.

Within a host, the system predicts each request's expected outcome, compute cost and latency cost in real time. It then solves a constrained optimization under a hard tail-latency SLO. As load rises, Meta says resource-intensive work with lower impact is trimmed first.

Why the system matters for Meta's ad infrastructure

Intelligent Delivery addresses a problem that becomes more difficult as ranking systems use larger models, broader candidate sets and richer delivery architectures. Meta says compute cannot scale as quickly as model and system complexity, so improving serving efficiency is as important as building more powerful models.

The framework extends from kernel-level scheduling to application-level thread pools and also covers decisions such as how many ads to rank at each stage and which candidates enter the pool. For people using Meta's apps and services, the company says the purpose is to run advanced ads ranking models efficiently in production and use each unit of compute to serve better ads.