Who it is for
Data platform and analytics engineering teams, and vendors embedding forecasting into streaming or warehouse products.
Why now
Two independent moves landed in this window: a 330M zero-shot model added native multivariate support and single-pass horizon decoding with release across GitHub, Hugging Face and a planned warehouse channel; and a data-streaming vendor exposed four time-series foundation models natively as SQL functions switchable by a parameter, so users no longer provision GPU serving or move data out of the stream. A third release — a self-supervised CGM foundation model reducing labeled-data dependence — shows the same pattern of label-free adaptation in a regulated vertical. A year ago each series or team needed its own model; the binding constraint is now platform availability and whether one shared model generalizes to a given organization's series.
What would kill it
Independent benchmarks from an evaluator not affiliated with either vendor show zero-shot time-series foundation models lose to per-series statistical models on representative production series, which would remove the demand for SQL-native foundation forecasting.