Meta leverages idle AI inference to boost ads recommendations without extra capacity

Oct 13, 2026 · 3 min read

Prepared with AI assistance from source material. An individual human editorial review is not recorded for this archive article. The named archive byline has not been independently verified by TechStaged. Editorial policy and corrections.

TechStaged-owned

Summary

Meta built a system that enhances ads recommendations by leveraging idle inference capacity in its fleet.
Two elasticity levers, Rank More Ads (RMA) and Serve Better Models (SBM), expand workload using opportunistic capacity without harming baseline performance.
The system aggregates pockets of idle capacity into a shared opportunistic capacity pool.

Meta describes a system that enhances ads recommendations by leveraging idle inference capacity in its fleet. The aim is to expand the work of ads ranking and model evaluation when spare capacity is available, without impacting the baseline performance that serves ads with strict latency and reliability targets.

TWO ELASTICITY LEVERS TO USE OPPORTUNISTIC CAPACITY

To safely deploy opportunistic capacity, Meta uses two complementary mechanisms: Rank More Ads (RMA) and Serve Better Models (SBM). RMA increases the number of candidates evaluated in each ranking stage when capacity is available, while SBM adjusts the share of requests routed to a more powerful model. TechStaged has also covered Meta unveils IcePack liquid-cooled rack and IceCube 102.4T switches at OCP Summit 2026 to power AI across scale-up, scale-out, and scale-across networks.

RMA expands candidate evaluation by leveraging CPU and accelerator surplus
SBM routes a portion of requests to more capable models using accelerator capacity while keeping upstream CPU demand stable

FROM FRAGMENTED RESERVATIONS TO A SHARED POOL

Idle inference capacity is inevitable in large fleets. Meta aggregates pockets of idle capacity into a shared opportunistic capacity pool and uses it to improve ads recommendations, aiming to improve relevance for users and outcomes for advertisers without new capacity orders.

SAFEGUARDS TO PROTECT BASELINE PERFORMANCE

The system is designed to prevent loss of opportunistic capacity from harming baseline serving. It employs isolation and feedback control to ensure baseline latency and reliability targets are not violated as capacity fluctuates.

HOW RMA AND SBM OPERATE IN PRACTICE

RMA uses a closed-loop controller to adjust the volume of candidate ads, guided by observed tail latency. SBM uses a central controller and a client-side router that segments requests by predicted engagement, with ongoing calibration to maintain targeted usage of the more powerful model.

RMA relies on a PID-like control loop at the server and client enforcement
SBM uses a request-value signal and explore/exploit updates to keep the more powerful model usage aligned with value

SAFE HANDLING OF CAPACITY SWINGS

When opportunistic capacity contracts, both mechanisms reduce their use of the more powerful paths and may re-route traffic to baseline models. The system also employs value-aware shedding at both server and client levels to absorb rapid capacity changes without overloading baseline serving.

WHAT THIS MEANS FOR ADVERTISERS AND USERS

The approach is intended to deliver more relevant ads and better advertiser outcomes by utilizing otherwise idle resources, without requiring additional capacity orders. Meta notes that the changes are made with safeguards to preserve baseline performance and reliability at scale.

SOURCES

Engineering at Meta: Turning Opportunistic AI Inference Capacity Into Better Ads Recommendations Published Oct 12, 2026 · Primary source