Meta's Andromeda retrieval system now processes more than one billion eligible ad candidates per second on a single GPU server, roughly a 100x throughput increase over its 2024 system. That number is easy to misread. So let's start with what it actually measures, then work through what it means for your campaigns and measurement setup.

Throughput Is Not Delivery

The billion-ads-per-second figure is a retrieval throughput metric: how fast Andromeda can score candidates in the first stage of Meta's ad pipeline. It is not impressions served per second. Retrieval is the shortlisting step. Andromeda narrows a massive eligible pool to a smaller candidate set so that heavier ranking models can evaluate each survivor more deeply. Only after ranking and auction does an ad actually get shown.

Retrieval throughput determines which ads get considered. Ranking and auction determine which ads get delivered. Two different problems, two different levers.

Three Layers, Not One Model

Meta's October announcements covered three linked upgrades, and conflating them leads to bad diagnostic decisions.

Andromeda (retrieval): Scores the eligible pool and produces a candidate shortlist. The redesign uses model-system co-design (mixture-of-experts with attention-based gating, fused GPU kernels that reduce data movement) to push throughput up dramatically. Each ad now routes to specialized "expert" sub-models, making candidate selection more precise.

GEM (foundation model): Generative Ads Recommendation Model, a large model trained on broad ads-system data. It doesn't pick ads itself. It transfers learned representations into smaller production models that handle real-time ranking across Facebook, Instagram Reels, and off-Meta conversions. The new version reads twice as long a user history. As Meta VP Matt Steiner described it, GEM is a large model whose learning gets distilled into smaller models for specific uses.

GenRecs (semantic tokenization): This is where LLMs enter the picture. GenRecs uses language models to read an ad's text and images, then describes it by key attributes (brand, color, fit, style) using semantic tokens. Every ad gets described in the same vocabulary, so the system can see what a new ad has in common with existing ones. The payoff: shorter cold-start learning phases. Steiner was clear that it doesn't eliminate the cold-start problem ("You still have to learn," he told AdExchanger), but it gives the system a head start.

What This Actually Changes for Your Campaigns

If Andromeda retrieves candidates differently, the competitive set your ads face in ranking shifts. That can show up as changes in delivery patterns, learning-phase behavior, or CPA movement that aren't caused by anything you changed. PPC Land reported Meta's claim of a 2.4% Facebook conversion lift associated with the newer Andromeda system. That's a platform-level average, not a guarantee for your account.

The practical move: separate retrieval effects from ranking/auction effects when you diagnose performance changes. If CPA spikes on a campaign you haven't touched, check whether delivery volume or audience composition shifted (retrieval change) versus whether conversion rates dropped on the same audience (ranking or signal problem). Different root causes, different fixes.

Signal quality becomes more important, not less. GEM's distillation approach and GenRecs' semantic tokenization both depend on clean, consistent conversion events. If your event instrumentation is sloppy (duplicate events, missing parameters, inconsistent naming), the models will learn the wrong patterns. For B2B teams optimizing toward pipeline outcomes, your conversion API setup, event deduplication, and creative taxonomy need to be tight. Garbage signal in, garbage optimization out, faster than before.

Creative refresh cadence may need adjustment. GenRecs shortens the cold-start phase by giving the system semantic context before engagement data accumulates. New creatives ramp faster, which also means creative fatigue detection and rotation need to keep pace. If your testing cadence was calibrated to a longer learning phase, revisit it.

The Targeting Question

Steiner told AdExchanger that "the importance of targeting is diminishing over time as ranking models get better at selecting who an ad would be relevant for." That's a strong claim from the person who runs the infrastructure. It doesn't mean targeting controls are useless today, but the value of manual audience segmentation erodes with every model upgrade.

For B2B advertisers, the trade-off is real: broad targeting lets Meta's models find high-value prospects you wouldn't have segmented into, but it requires that your conversion signals clearly define what "high value" means. If you're sending Meta a form-fill event when what you actually care about is qualified meetings, the model will optimize for form fills. Broadening targeting without tightening your conversion definition is how you get volume that looks great on the dashboard and produces nothing downstream.

The hypothesis worth testing: if you tighten your conversion event to a deeper funnel stage (say, qualified opportunity creation) and let Meta's broader retrieval do the audience work, does pipeline quality improve even if volume drops? Set a stop-loss at a 25% CPA increase over two weeks. If quality moves and volume recovers, you've found your new baseline. If not, you've learned something about your signal quality before the next model update makes it matter even more.