Meta’s October 28 engineering update explains how its next-generation Andromeda system uses model, GPU and system co-design to retrieve relevant ads at much greater scale.
Meta has detailed a major evolution of Andromeda, its personalized advertising retrieval system, explaining how a redesigned model and GPU pipeline can evaluate a much larger pool of eligible ads. The engineering update was published on October 28, 2026.
Andromeda operates early in Meta’s ad-selection process. It narrows a large pool of eligible ads to a smaller set of candidates for later ranking. Meta says the latest system combines model architecture, software and hardware improvements rather than treating them as separate optimization problems.
What Meta says has changed
Meta describes a retrieval architecture inspired by mixture-of-experts models, with specialized user representations and contextual information about candidate ads. The system is designed to identify relevant candidates efficiently while accounting for a wide range of user interests and real-time context.
The company says the production system can process more than one billion ads per second on a GPU server while scoring eligible ads in its retrieval pipeline. Meta also describes a fused GPU pipeline that reduces data movement between components and a co-location strategy that brings index matching, scoring and candidate selection closer together.
Why retrieval matters in ad delivery
Ad delivery is not simply a matter of choosing the highest bid. Platforms must first determine which ads are eligible for a person and context, then estimate relevance and expected outcomes. Retrieval is the stage that reduces a very large candidate pool to a manageable shortlist.
Improving retrieval can help the platform consider a broader set of potentially relevant ads without making the process impractically slow. It does not mean every advertiser will automatically receive more conversions or lower costs; campaign outcomes still depend on the offer, creative, audience, landing page, competition and measurement setup.
What advertisers should do
- Develop meaningful creative variety: Create distinct concepts, messages and formats rather than making many near-identical variations.
- Keep conversion signals reliable: Review Pixel and Conversions API event quality, deduplication and event priorities where applicable.
- Use accurate conversion values: Where value-based optimization is appropriate, pass values that reflect business outcomes rather than arbitrary amounts.
- Evaluate performance beyond platform-reported reach: Monitor qualified leads, sales, customer acquisition cost and revenue.
- Avoid overreacting to short-term fluctuations: Assess performance over a suitable period and account for budget, seasonality and creative changes.
Does Andromeda require an advertiser setting?
Meta’s engineering article describes changes to its underlying ad retrieval infrastructure. It does not announce a new advertiser-facing switch or a required campaign migration. Advertisers should continue using the campaign and measurement controls available in Ads Manager and follow official Meta guidance for any account-specific changes.
The broader direction of Meta advertising
The update reflects a wider shift toward AI-driven ad selection, where the platform’s ability to understand creative content, context and user signals plays an increasingly important role. For advertisers, the durable response is to provide strong creative inputs, accurate conversion data and a clear value proposition—not to chase a single algorithm update with untested campaign changes.
Source
MarketNex editorial note: Throughput and performance figures in this article are claims reported by Meta about its own infrastructure; they should not be interpreted as guaranteed campaign-level results for individual advertisers.
