Retail recommender systems reach their data ceiling far sooner than operators assume. Meta’s analysis of DLRM-style click-through-rate models found that quality scales as a “power law plus constant” in data size, parameter count, and compute. The arithmetic is unforgiving: each tenfold increase in training data buys a smaller accuracy increment than the last, and the constant represents an error floor that no volume of data can breach. Parameter scaling hit its limit first: for the architecture under study, the researchers concluded it was “out of steam.” Fashion retailers generating millions of browse-click-purchase events daily are already operating on the flat part of that curve, whether they have measured it or not.

When the ceiling is architectural, the fix is architectural. Meta demonstrated this directly with HSTU, a sequential transducer that treats user actions as tokens in a generative framework. The model broke through the parameter plateau where DLRM had stalled. Deployed at 1.5 trillion parameters, HSTU delivered a 12.4% improvement in online A/B tests. On public benchmarks, the NDCG improvement reached 65.8%. No plausible increase in training data volume could have matched those gains under the old architecture.

Fashion recommendation compounds the problem because the signal decays faster than in categories with stable purchase cycles. A customer who bought running shoes six months ago will likely need running shoes again; a customer who bought a floral midi dress may never want another. Seasonal rotation, trend velocity, and visual novelty all shorten the half-life of interaction data in apparel. Climber, deployed in music streaming where consumption signals decay at comparable rates, addresses this directly: its multi-scale sequence extraction processes different time horizons at different resolutions. The 12.19% overall production lift it achieved is the first documented instance of controlled model scaling driving continuous online metric growth on its deployment platform. When data decays this fast, accumulating more of it cannot substitute for architecture that adapts to the decay.

The retailer with the deepest proprietary preference dataset in fashion already tested what happens when architecture fails to keep pace with the data it ingests.

Stitch Fix built an entire business around the premise that proprietary style-preference data creates compounding returns. The company accumulated over a decade of explicit feedback: fit ratings, style profiles, rejection reasons, stylist annotations. By any data-moat theory, this dataset should have widened the competitive gap over time. Instead, the company reported seven consecutive quarters of year-over-year revenue decline averaging 18%, with active clients falling 15% in fiscal 2024. Multiple factors drove that decline, but one stood out by its absence: the data advantage did not compound.

The strongest case against this thesis is that fashion preference is personal in ways that resist architectural shortcuts. A cold-start recommender can surface bestsellers; only a mature dataset can predict that a specific customer wants wide-leg trousers in olive rather than navy. This is true in the narrowest sense — collaborative filtering still outperforms LLMs in data-rich settings. But recent work on dynamic representation learning shows that new users and items can be represented without fine-tuning, through a single forward pass over existing embeddings, outperforming comparable methods by 29.5 to 47.5 percent in cold-start scenarios. If the cold-start gap closes in weeks rather than years, the data advantage is a head start, not a moat.

The investment frontier for retail recommendation has shifted. Meta’s Foundation-Expert paradigm replaces monolithic recommendation models with lightweight, surface-specific experts built on a shared foundation, reducing compute by centralizing general representations rather than retraining from scratch for each deployment surface. A separate line of scaling research achieved over five times the training efficiency and 21 times the inference efficiency through architectural redesign alone, alongside 4% to 8% gains in production consumption and engagement metrics. These are gains that no plausible increase in data collection can match. If this pattern holds, the retailers who maintain recommendation advantages will be those who adopt architectural advances fastest. The board presentation that leads with data volume is answering a question the field moved past years ago.