Meta's engineering team has published a detailed look at Intelligent Delivery, an internal ML system that decides how much serving compute each ad request receives. Meta says the system has added roughly 10% incremental ads ranking performance on top of model and system advances between 2017 and the first half of 2026. It is infrastructure, not an Ads Manager feature, but it shapes which ads get the deepest ranking treatment.
What Intelligent Delivery does
Meta cannot run its heaviest ranking models on every request, so Intelligent Delivery predicts which ad requests are likely to matter most and gives them more compute. The post describes a hierarchy of decisions: which region serves a request, which hardware it runs on (Intel, AMD and Arm CPUs; NVIDIA, AMD and Meta's own MTIA accelerators), and how much host capacity it gets.
Traffic between regions is rebalanced on a cycle of minutes, and a feedback controller manages capacity down to thread pools.
The models and the numbers
At the core is what Meta calls the Resource Core Model, with 6 billion sparse and 20 million dense parameters. It runs once before candidate retrieval and is shared across the funnel, with specialised models fine-tuned from it.
Meta reports that its request-understanding layer is about 20% better than the previous production model at identifying the highest-impact ad requests in offline benchmarks. AI agents contributed to about 20% of Intelligent Delivery optimisations in the first half of 2026, and Meta says engineers still review and approve each change before it ships.
- ~10% incremental ads ranking performance, 2017 to H1 2026 (Meta's figure)
- ~20% offline gain in spotting high-impact requests
- Agents behind ~20% of optimisations in H1 2026
Why it matters to media buyers
None of this is a setting you can touch, and Meta gives no methodology for its figures. The practical point is that delivery quality is not uniform: requests the system judges as high impact get more ranking work, which helps explain why broad, signal-rich campaigns tend to stabilise faster than tightly constrained ones.
Meta also lists unfinished work, including serving large ranking models such as GEM on GPUs, so expect more delivery changes that show up as performance shifts rather than announced features.
What to do with this
Treat sudden CPM or CPA swings without a change on your side as a possible platform-side shift, and compare against a stable control campaign before rebuilding structure. Keep conversion signals clean through Pixel and Conversions API so the system has good data to judge your requests by. If a shift coincides with an account being restricted or throttled, Power Ads' 24/7 operator can help you check whether the issue is the account or the delivery.
What it means for you
- Intelligent Delivery is internal; no Ads Manager toggle changes
- Strong, consistent conversion signals help your requests rank as high impact
- Benchmark unexplained swings against a control campaign before restructuring
- Expect ongoing, unannounced delivery changes as Meta tunes compute
