Meta's New Ad Engine Scores a Billion Ads a Second

Meta has published the details of Andromeda, the machine-learning system that now decides which advertisements to show across Facebook and Instagram, and the engineering numbers are startling. In a post on 28 October, the company said the system can score every eligible ad and surface the strongest candidates at more than a billion ads per second on a single server. That is roughly a hundred times the throughput of the setup it ran in 2024.

The interesting part is how Meta got there. Ad retrieval, the job of narrowing billions of possible ads down to a handful worth ranking closely, has long been a brute-force problem. Andromeda borrows an idea from large language models instead, adding what are called mixture-of-experts layers so that only the relevant slice of the model fires for any given query. Meta says that lets it approach the quality of its heavier ranking models while using about 300 times fewer calculations per ad.

At that pace the bottleneck stops being arithmetic. Meta's engineers found that simply shuttling data between the chip's processing cores was generating terabytes per second of memory traffic, so they redesigned the pipeline to keep the matching, scoring and selection steps on one GPU, cutting network traffic by around a hundredfold. Working with Nvidia, they pushed one key step past ten billion ads per second on a single GH200 chip.

Does any of this matter outside Meta's data centres? Financially, yes. The company says Andromeda has lifted conversions on Facebook by 2.4 percent since 2025, a figure it reports without a stated baseline, so read it with some caution. The deeper point is that the architecture behind chatbots is now quietly reshaping the advertising machine that pays for much of the consumer internet. The model that drafts your emails and the model that picks your ads increasingly share a blueprint.