Before Facebook or Instagram decides which ad to show, it has to eliminate an enormous number of candidates. That early selection step is handled by Andromeda. On October 28, 2026, Meta published details of a rebuilt version that, it says, evaluates more than one billion eligible ads per second on a single GPU server — around 100 times the throughput of the 2024 system.

Meta's creative generation tools have contributed to the growth in eligible ads. That makes initial selection harder: the system needs to narrow a huge pool within milliseconds without discarding promising candidates before the more expensive ranking stage gets to examine them.

Behind the performance claim

The new pipeline puts targeting-rule checks, candidate scoring and selection on the GPU instead of moving large lists between separate servers. Meta also adopted a mixture-of-experts approach, using specialist components to interpret different user and context signals. Its engineering team reports roughly 300 times fewer inference-time calculations for the retrieval model and about a hundredfold reduction in network traffic against the previous design.

Meta attributes a 2.4% increase in Facebook conversions since 2025 to Andromeda. It is a company-wide reported effect, not a published forecast for any individual account. The engineering article does not provide the kind of campaign-level comparison a buyer would need to turn that number into an ROI expectation.

Does a buyer need to act?

Andromeda is part of Meta's delivery infrastructure, not a campaign setting that media buyers can enable. The announcement provides no reason on its own to rebuild an account or abandon a working targeting approach. Buyers can keep testing creative quality and conversion measurement against their own results, while treating Meta's throughput figures as an explanation of how the platform is changing under the hood.

The more revealing trend is the pressure on the retrieval stage: as creating ad variants becomes cheaper, choosing the right subset quickly becomes a larger part of the advertising problem.

Before Facebook or Instagram decides which ad to show, it has to eliminate an enormous number of candidates. That early selection step is handled by Andromeda. On October 28, 2026, Meta published details of a rebuilt version that, it says, evaluates more than one billion eligible ads per second on a single GPU server — around 100 times the throughput of the 2024 system.

Meta's creative generation tools have contributed to the growth in eligible ads. That makes initial selection harder: the system needs to narrow a huge pool within milliseconds without discarding promising candidates before the more expensive ranking stage gets to examine them.

Behind the performance claim

The new pipeline puts targeting-rule checks, candidate scoring and selection on the GPU instead of moving large lists between separate servers. Meta also adopted a mixture-of-experts approach, using specialist components to interpret different user and context signals. Its engineering team reports roughly 300 times fewer inference-time calculations for the retrieval model and about a hundredfold reduction in network traffic against the previous design.

Meta attributes a 2.4% increase in Facebook conversions since 2025 to Andromeda. It is a company-wide reported effect, not a published forecast for any individual account. The engineering article does not provide the kind of campaign-level comparison a buyer would need to turn that number into an ROI expectation.

Does a buyer need to act?

Andromeda is part of Meta's delivery infrastructure, not a campaign setting that media buyers can enable. The announcement provides no reason on its own to rebuild an account or abandon a working targeting approach. Buyers can keep testing creative quality and conversion measurement against their own results, while treating Meta's throughput figures as an explanation of how the platform is changing under the hood.

The more revealing trend is the pressure on the retrieval stage: as creating ad variants becomes cheaper, choosing the right subset quickly becomes a larger part of the advertising problem.

Before Facebook or Instagram decides which ad to show, it has to eliminate an enormous number of candidates. That early selection step is handled by Andromeda. On October 28, 2026, Meta published details of a rebuilt version that, it says, evaluates more than one billion eligible ads per second on a single GPU server — around 100 times the throughput of the 2024 system.

Meta's creative generation tools have contributed to the growth in eligible ads. That makes initial selection harder: the system needs to narrow a huge pool within milliseconds without discarding promising candidates before the more expensive ranking stage gets to examine them.

Behind the performance claim

The new pipeline puts targeting-rule checks, candidate scoring and selection on the GPU instead of moving large lists between separate servers. Meta also adopted a mixture-of-experts approach, using specialist components to interpret different user and context signals. Its engineering team reports roughly 300 times fewer inference-time calculations for the retrieval model and about a hundredfold reduction in network traffic against the previous design.

Meta attributes a 2.4% increase in Facebook conversions since 2025 to Andromeda. It is a company-wide reported effect, not a published forecast for any individual account. The engineering article does not provide the kind of campaign-level comparison a buyer would need to turn that number into an ROI expectation.

Does a buyer need to act?

Andromeda is part of Meta's delivery infrastructure, not a campaign setting that media buyers can enable. The announcement provides no reason on its own to rebuild an account or abandon a working targeting approach. Buyers can keep testing creative quality and conversion measurement against their own results, while treating Meta's throughput figures as an explanation of how the platform is changing under the hood.

The more revealing trend is the pressure on the retrieval stage: as creating ad variants becomes cheaper, choosing the right subset quickly becomes a larger part of the advertising problem.

Before Facebook or Instagram decides which ad to show, it has to eliminate an enormous number of candidates. That early selection step is handled by Andromeda. On October 28, 2026, Meta published details of a rebuilt version that, it says, evaluates more than one billion eligible ads per second on a single GPU server — around 100 times the throughput of the 2024 system.

Meta's creative generation tools have contributed to the growth in eligible ads. That makes initial selection harder: the system needs to narrow a huge pool within milliseconds without discarding promising candidates before the more expensive ranking stage gets to examine them.

Behind the performance claim

The new pipeline puts targeting-rule checks, candidate scoring and selection on the GPU instead of moving large lists between separate servers. Meta also adopted a mixture-of-experts approach, using specialist components to interpret different user and context signals. Its engineering team reports roughly 300 times fewer inference-time calculations for the retrieval model and about a hundredfold reduction in network traffic against the previous design.

Meta attributes a 2.4% increase in Facebook conversions since 2025 to Andromeda. It is a company-wide reported effect, not a published forecast for any individual account. The engineering article does not provide the kind of campaign-level comparison a buyer would need to turn that number into an ROI expectation.

Does a buyer need to act?

Andromeda is part of Meta's delivery infrastructure, not a campaign setting that media buyers can enable. The announcement provides no reason on its own to rebuild an account or abandon a working targeting approach. Buyers can keep testing creative quality and conversion measurement against their own results, while treating Meta's throughput figures as an explanation of how the platform is changing under the hood.

The more revealing trend is the pressure on the retrieval stage: as creating ad variants becomes cheaper, choosing the right subset quickly becomes a larger part of the advertising problem.