Intelligent Delivery: Optimizing Compute for Large-Scale Ads Ranking
Meta implements an ML-driven system to dynamically allocate compute resources for ads ranking.
1System Scaling Challenges
Ads ranking models are expanding toward LLM scale, outpacing available hardware growth. This creates a critical efficiency gap where compute resources fail to match the complexity of modern ranking architectures. Meta developed Intelligent Delivery to manage this disparity, ensuring that every unit of compute is utilized effectively across their heterogeneous serving stack.
2Intelligent Delivery Overview
Intelligent Delivery is an ML-driven optimization layer designed to maximize system-wide ranking performance. Since 2017, this framework has delivered approximately 10% incremental performance improvements. It functions by continuously directing resources toward the specific requests, models, and hardware fleets that provide the highest marginal utility, ensuring optimal efficiency under strict latency budgets.
3Resource Core Model
The Resource Core Model uses deep learning with 6B sparse and 20M dense parameters to understand request behavior. By predicting marginal utility before candidate retrieval, it replaces outdated, manual rule-based segmentation. This shared intelligence layer captures non-linear interactions among thousands of signals, achieving a 20% improvement in identifying high-impact ad requests.
4Hierarchical Decision Engines
Meta decomposed the complex resource distribution problem into a hierarchy of decision engines. These engines manage traffic across global regions, assign workloads to specific hardware like CPUs or GPUs, and control host-level resource allocation. This multi-tier approach ensures that scarce hardware is matched precisely to the workloads that generate the most value.
5Hardware Heterogeneity Management
The infrastructure spans diverse hardware, including Intel, AMD, Arm CPUs, and NVIDIA accelerators alongside Meta’s own MTIA chips. Because capacity is not interchangeable, the system benchmarks efficiency for every workload-SKU pair. This allows the framework to solve assignment problems dynamically, directing compute to the most efficient hardware for each specific ranking stage.
6Host-Level Optimization
Within individual hosts, the system manages concurrent requests by balancing latency budgets against expected outcomes. Using real-time predictions, the framework performs constrained optimization to maximize impact. As host load increases, the system automatically trims resource-intensive, low-impact tasks, enforcing control from the kernel-level scheduler up to the application-level thread pools for stability.
7Agentic AI Integration
To manage the vast search space of configurations, Meta employs agentic AI that monitors system landscapes. These agents analyze source code, design documents, and experiment logs to propose optimizations. By automating hypothesis generation and testing, agents have significantly shortened the optimization loop, allowing engineers to focus on high-level strategy rather than manual configuration.
8Agentic Performance Metrics
In the first half of 2026, AI agents contributed to approximately 20% of all Intelligent Delivery optimizations. For the most automated mechanisms, this contribution reached 80% of the engineering effort. This shift reduces the time required for optimization cycles from months to weeks, while maintaining human oversight for every final change.
9Global vs Local Optimality
A primary challenge is that local optimizations often fail to translate into global system efficiency. Because ads delivery involves complex dependencies across data centers and ranking stages, resource shifts must be coordinated. Meta uses closed-loop hardware testing environments to validate changes, ensuring that local adjustments contribute to the global optimum target.
10Future Technical Frontiers
Meta is currently working toward end-to-end real-time optimization by fusing request understanding and serving into a single model. Additionally, they are reimagining levers for GPU-based serving, as large models like the Generative Ads Model require different throughput and memory management strategies compared to traditional CPU-based ranking architectures and legacy serving stacks.
11Strategic System Impact
Intelligent Delivery represents a shift toward self-optimizing infrastructure that adapts to evolving user behavior and model complexity. By leveraging agentic AI and deep learning, Meta maximizes the utility of finite compute resources. This approach ensures that advanced ranking models remain performant and efficient, ultimately serving better ads to global users.