Meta Andromeda retrieval exceeds one billion ads scored per second in production
Meta Engineering published the next generation of Andromeda, Meta’s personalized ads retrieval engine, built through co-design of the model, serving system, and NVIDIA GPU hardware. In production, Meta reports the system scores every eligible ad at near ranking-level quality at retrieval scale and surfaces top candidates at over one billion ads per second, a throughput level the company frames as a hundredfold lift versus its 2024 Andromeda system within the same latency budget. Meta also attributes a 2.4 percent conversion increase on Facebook since 2025 to matching ads across a larger candidate pool at retrieval. The post positions Andromeda as the first stage that narrows vast ad inventories before a heavier ranking pass. Operators should treat the throughput and conversion figures as vendor-reported until independent measurement details appear.
Meta Andromeda retrieval exceeds one billion ads scored per second in production
Meta published the next generation of Andromeda, its ads retrieval engine, after model-system-GPU co-design with NVIDIA. In production it runs at near ranking-level quality at retrieval scale and surfaces top candidates at over one billion ads per second, described as a 100x throughput lift versus the 2024 Andromeda system within the same latency budget.
Key takeaway
Retrieval-stage ads scoring at near ranking quality and billion-per-second throughput is now claimed in production, but evidence in the packet is Meta-reported only.
What happened
Meta AI published an engineering post on October 28, 2026 describing the next generation of Andromeda, its personalized ads retrieval engine, after model-system-GPU co-design with NVIDIA.
Meta states the production system scores eligible ads with near ranking-level model performance at retrieval scale, surfaces top candidates at over one billion ads per second, cites a 100x throughput lift versus the 2024 Andromeda system within the same latency budget, and reports a 2.4% increase in conversions on Facebook since 2025.
Evidence
Production Andromeda surfaces top ad candidates at over one billion ads per second.
Meta AI · attributed
Meta reports the system surfaces top candidates at over one billion ads per second, a 100x lift compared to the 2024 Andromeda system
The system scores every eligible ad with near ranking-level model performance at retrieval scale.
Meta AI · attributed
In production, the system scores every eligible ad with near ranking-level model performance at retrieval scale
Meta Andromeda delivered a 2.4% increase in conversions on Facebook since 2025.
Meta AI · attributed
Since 2025, Meta Andromeda has delivered a 2.4% increase in conversions on Facebook.
The Information reported Meta plans to raise up to $95 billion for AI through equity sales.
The Information AI · attributed
Meta Plans to Raise Up To $95 Billion for AI Through Equity Sales - The Information
Why it matters
Large-scale recommendation and ads teams may reconsider splitting retrieval and ranking quality if co-designed hardware and models can keep latency flat while expanding scored candidate pools.
Limits and uncertainties
The packet provides Meta self-reported metrics without independent verification or detailed methodology for the 100x throughput benchmark or 2.4% conversion lift.
Available text is an excerpt; latency budget specifics and sustained generalization beyond Facebook are not established in the packet.
Practical implications
Teams architecting two-stage recommenders should compare retrieval compute budgets against ranking-grade models when planning GPU co-design with vendors.
Treat advertised throughput and conversion gains as hypotheses until replication data or third-party benchmarks appear.
What to watch
Whether Meta publishes measurement methodology for the 100x throughput claim and the Facebook conversion lift.
Independent benchmarks or customer-visible ad performance trends after broader Andromeda rollout.
