Your lakehouse is fresh. Why are your forecasts still days old?
Most predictive ML architectures contain the same contradiction: a continuously updated, zero-copy lakehouse feeds a pipeline that freezes the data, trains one model per tenant and use case, and retrains it on a schedule. Three thousand tenants and four use cases already mean twelve thousand model artifacts—each with its own lifecycle and its own shadow copy of data you spent a decade centralizing.
Tabular foundation models offer a radically different architecture. They can predict from labeled rows supplied as context at request time: no per-tenant training run, no model artifact, no frozen extract. One shared, stateless model can replace the **N × M × R** multiplication of tenants, use cases, and retrains.
What happens when you delete the training pipeline? Tenant onboarding becomes a query scope. The feature store loses its reason to exist. The federated Iceberg lakehouse becomes the serving layer—and model freshness collapses into data freshness. The architectural fit is unusually exact: tabular foundation models need live, governed, point-in-time addressable data; the modern lakehouse was built to provide precisely that.
But the work does not disappear. It moves into context selection, governance, point-in-time reproducibility, and inference latency.
This article traces that architectural shift and tests it through three experiments using TabPFN, Dremio, and Iceberg: comparing shared and per-tenant model accuracy, preserving isolation across physically separate tables, and reproducing past predictions from lakehouse snapshots.
