Divide the trillion training points by the 330 million parameters and you get about three thousand points per parameter [13][2], which is a statement about what is being sold: enough exposure to the shapes a series can take to guess at one the model has never seen, well short of the capacity to memorise any single client's history.
The mechanism that matters to a planning team is small and specific. For covariates whose future values are already known, each token concatenates the current patch with future patches, a lookahead that keeps those signals visible across the forecast horizon while the target series stays masked [8]. In Google's own promotion illustration the univariate line does the only thing it can, projecting the historical weekly pattern forward with no knowledge that a promotion is booked [11]. That gap is the job description of a lot of in-house demand models.
So the argument for keeping one moves, or rather the more interesting version of it moves: it stops being about whether a generic forecaster can ingest a promotion calendar and starts being about data lineage and whose name is on the number when a modelled lift arrives at a quarter of its forecast size. Google says the model significantly outperforms other forecasting models across major benchmarks [12], which is the vendor's sentence and should be read as one.
Worth noticing what got dropped to make this work. Every earlier version generated forecasts one patch at a time, paying for it in latency, accumulating error and compute [7], and the architecture budget went instead into a stack that alternates attention along time and across series over patches of 32 steps [10]. A model advertised as needing no task-specific fine-tuning [5] is also a decision about where revenue comes from: inference, rather than per-customer tuning work.
This could land a few different ways. Zero-shot covariates might prove good enough on the median series, and internal forecasting collapses into data plumbing. Or they are fine on the median and weak in the promotion weeks and product launches where the margin actually sits, in which case bespoke models survive exactly where they were always expensive. Or the model becomes the free baseline that every internal model has to beat in writing, which is a documentation cost rather than a headcount cut. This is probably wrong, but I would take the third. What would break it: quantile bands that come out miscalibrated on a firm's own history, or a fine-tuned baseline that still beats zero-shot by more than the loaded cost of the people who maintain it.
