How Meta uses AI to decide where every bit of computing power goes for better ads

How it works

An AI system inside Meta's ad engine that routes extra computing power to the ad requests that need it most.

1. 🔍A big AI model studies each ad request first
2. 📊Smaller models predict result, speed, and compute needed
3. ⚡Extra computing power goes to the best ad requests
4. 🔄Agent tools keep adjusting as things change

What you could do with it

• Meta could deploy its Intelligent Delivery system to automatically allocate extra computing power to the ad requests that will benefit most from it ¹.
• An advertiser using Meta's platforms could get better, more relevant ads shown to them without Meta needing to spend more on servers ¹.

Words to know

Compute: The processing power a computer uses to run calculations and programs · Marginal utility: How much extra benefit you get from one additional unit of something

What the problem looks like

Ads ranking models — the systems that pick which ads to show you when you open an app — are getting bigger and more powerful faster than the chips running them can keep up ¹. When Meta adds more computing capacity to its ad engine, that extra power does not help every ad request equally . Some requests, some types of ads, and some pieces of hardware get much more out of it than others .

The question is: how do you decide where each unit of compute should go so the entire system works as well as possible ¹?

Meta answered with a system called Intelligent Delivery ¹. Since 2017 it has added roughly 10 percent to ads ranking performance beyond what improvements in models and hardware alone would have produced . That means better ads for people who use Meta's apps and services .

Why this is hard

There are three reasons this is tricky ¹. First, the same extra bit of computing power can be worth very different amounts depending on what is happening at that moment — what product you are on, what you seem interested in, which ads are being considered . Second, Meta's ad system stretches across many stages, hundreds of different models, several generations of chips and data centers around the world; moving resources between all those parts is not straightforward . Third, the best answer keeps changing as new models arrive, hardware gets replaced and user behavior shifts .

How the system works

Instead of having engineers write rules to split ad requests into buckets and treat each bucket the same way, Meta reframed the problem as two machine learning questions ¹: which requests are most likely to benefit from extra compute, and how will different resource settings change their results and speed ?

To answer the first question they built the Resource Core Model, a deep learning model with 6 billion sparse parameters and 20 million dense parameters that runs once before the system even starts looking at candidate ads ¹. It learned patterns that human-made categories could not capture .

They then fine-tuned that core model into smaller specialized versions that predict outcome, compute use and delay for each request ¹. Together these form a shared intelligence layer that understands each ad request and its context .

In offline tests, this approach improved the ability to spot the highest-impact ad requests by about 20 percent compared with the previous production model ¹.

The system also increasingly uses agent-like AI tools that continuously watch for changes in how the system behaves and adjust resource allocation without human intervention ¹.

How good is it

According to Meta, the Resource Core Model improved the ability to spot the highest-impact ad requests by about 20% in offline benchmarks compared with their previous production model ¹. Since 2017, Intelligent Delivery has contributed roughly 10% additional ads ranking performance improvement on top of what model and hardware gains alone produced .

Corrections and updates

• Corrected on 12 November 2026. An earlier version specified particular apps (Instagram, Facebook) and claimed no need to buy more servers; the source does not list those apps by name nor state that no servers are needed. It only says better ads for people who use their apps and services. An earlier version claimed an ML engineer could study and adapt the approach for other systems, but the source does not discuss generalizing the method beyond Meta's own infrastructure or advise third-party adaptation. Removing the unsupported speculation is appropriate.

Key facts

• Intelligent Delivery has added ~10% incremental ranking performance since 2017 at Meta
• The Resource Core Model has 6B sparse parameters and 20M dense parameters
• About 20% improvement in identifying highest-impact ad requests vs. previous model
• System uses ML and increasingly agentic AI to allocate compute dynamically
• Optimization spans multi-stage ranking, hundreds of models, multiple chip generations

Sources

1. Intelligent Delivery: Making Every Unit of Compute Serve Better Ads — engineering.fb.com