The Pulse
Meta says compute allocation added about 10% to ad ranking on top of model and system advances
Meta described an ad-serving system that directs computing resources across requests, models and hardware. The company says it contributed an additional ~10% improvement in ranking performance, on top of model and system advancements, across major surfaces from 2017 through the first half of 2026.
Meta says its system for directing computing resources across ad requests contributed about a 10% improvement in ads ranking performance across major surfaces from 2017 through the first half of 2026, on top of gains from model and system advances. An engineering post published November 11 describes how the system, called Intelligent Delivery, estimates where extra computing will have the most effect and adjusts its allocation across the serving stack.
The system operates as Meta’s ad ranking models grow toward the scale of large language models. Its allocation decisions cover requests, models, hardware fleets and ranking stages, within the limits of available serving capacity and response time.
Predicting the value of compute
Extra computation can have different effects depending on the product surface, user intent, quality of candidate ads and system state. Meta initially grouped requests by attributes such as surface and context, with engineers setting allocation policies for each group; the company says the approach improved results but did not scale as signals multiplied.
The Resource Core Model has 6 billion sparse parameters and 20 million dense parameters. It runs once before candidate retrieval and is shared across the delivery funnel, with specialized models fine-tuned from it to generate signals about outcomes, compute consumption and latency.
In offline benchmarks on major surfaces, the machine-learning approach improved identification of the highest-impact requests by about 20% relative to the previous production model, according to Meta. The predictions feed mechanisms that determine how to allocate resources.
Moving resources across the stack
Allocation decisions are organized into a hierarchy. Infrastructure shifts traffic between regions in a cycle of minutes, taking account of differences in hardware and available capacity.
Within the hardware fleet, workloads are matched to available machines using measured efficiency for each workload and hardware combination. Retrieval is constrained by memory capacity, ranking by compute and model inference by accelerator throughput, the company says.
Inside a host, the system predicts each request’s expected outcome and its compute and latency costs. It then weighs those estimates against the host’s latency budget; as load rises, resource-intensive work expected to have a low impact is trimmed first.
Other decisions include how many ads to rank at each stage and which candidates enter the pool. Meta tests changes in a closed-loop hardware environment because a change at one point, such as shifting traffic between regions, can affect decisions elsewhere.
Agents and future work
AI agents monitor the system for changes and inefficiencies, examine source code, design documents and experiment logs, and propose changes to test. Engineers review and approve final changes before they ship.
In the first half of 2026, agents contributed to about 20% of Intelligent Delivery optimizations, Meta says. For its most automated mechanism, the company says agents now contribute to about 80% of the engineering effort and have shortened the optimization loop from months to weeks.
Understanding a request and deciding how to serve it remain separate steps. Meta is working toward combining them in a single real-time model, while also studying allocation for large ranking models that increasingly run on GPUs, whose capacity, batching and memory behave differently from CPU time.