Meta has rebuilt the first gate in its ad delivery pipeline — the retrieval stage that decides which ads get a shot at auction. The company says the new Andromeda system scored every eligible ad at more than one billion ads per second on a single GPU server and delivered a 2.4% increase in Facebook conversions since 2025.
The gate before the auction
Andromeda is the first stage of Meta’s ads recommendation pipeline. In milliseconds it shortlists a large pool of eligible ads into a small candidate set; a more expensive ranking stage then scores the survivors. If an ad is not shortlisted, it never reaches the auction. That makes retrieval a hidden but decisive step for performance marketers.
Meta wrote that the system now “scores every eligible ad with near ranking-level model performance at retrieval scale.” In practice, that means it is checking all eligible candidates rather than an approximation, which changes which ads get a chance in principle.
What changed under the hood
The rebuild moves beyond the version Meta first described on December 2, 2024. The model now borrows a mixture-of-experts design from large language models, using specialised user towers, real-time context in the candidate ad tower and a routing network that activates only the most relevant experts for each request.
At the system level, Meta fused the scoring kernel so intermediate results stay in fast memory, and moved targeting checks, scoring and top-K selection onto one GPU server. That reduces network traffic by about 100x, Meta says.
- The push was driven by generative AI creative tools, which helped add hundreds of millions of new ads and grew the candidate pool.
- Meta names three design constraints: model quality, throughput at billions of ads per second, and hardware headroom.
- Nvidia helped optimise GPU kernels, with component benchmarks of 2x to 5x in early 2026.
The numbers need a closer look
The headline 2.4% conversion lift is Meta’s own figure. It comes without a baseline, end date, test design or Instagram result. A separate investor update on October 21 said Andromeda improvements had lifted Q3 Facebook conversions by 1.4% — a different number that Meta has not reconciled with the new claim.
The 100x comparison also switches from “per GPU card” to “per GPU server”, so treat it as an order-of-magnitude benchmark rather than a clean like-for-like. Meta’s average price per ad also rose 14% worldwide in Q3 2026 while impressions grew 10%, so cheaper ad serving does not automatically mean lower advertiser costs.
Advertisers get no new Ads Manager setting or report. Andromeda’s impact shows up only indirectly in delivery. That matters because platform-reported lifts do not always survive incrementality checks; Meta offers no incrementality data in the engineering post.
What to do about it
- Watch your own conversion and cost-per-action trends before crediting the stack.
- Run incrementality tests if you can, so you validate the lift against your own data.
- Do not expect a new lever to switch on; this is automatic delivery infrastructure.
The rebuild is a real infrastructure shift, but for Meta advertisers the first move should be measurement, not reallocation.
Source: PPC Land
