Single-cell foundation models obey data scaling laws on the optimized loss

Building ever-larger cell atlases is the field’s main bet for improving single-cell foundation models (scFMs), and it pays off only if these models keep improving as the data grows. A recent large-scale evaluation found that they do not — downstream performance stops improving after the first 1–10% of the corpus — implying the remaining data is not worth collecting. We trace this conclusion to what was measured rather than to scFMs themselves. Downstream transfer scores a frozen model on a task it was not trained for, through a metric already near its ceiling with little room left to register improvement; scaling laws are instead defined on the pretraining loss the model actually optimizes. Measured there, across ∼180 controlled runs spanning three orders of magnitude in data (10^4 to 10^7 cells), scFM pretraining obeys a saturating data scaling law: the loss falls steadily toward a positive floor. The law is predictive — fit on every data scale but the largest, it forecasts the loss at the held-out largest scale to within 2.5%, close enough to predict the most expensive run from cheaper ones without paying to train it, while a PCA control stays flat. Whether this law reaches a frozen model’s downstream performance is governed most decisively by how well the pretraining objective is aligned with the task — holding everything else fixed, swapping one aligned likelihood for a misaligned one flips scaling from a power law to flat. We claim saturation to a positive floor, not unbounded LLM-style scaling: data scaling in scFMs is real but conditional, and a downstream proxy can hide it entirely.