Abstract
Cities with the greatest need for hyper-local air-quality and weather forecasts often have the shortest data records and the tightest compute budgets. The usual responses, searching for a tiny architecture or transferring from data-rich cities, assume foundation models are too heavy for such deployments, an assumption adopted far more often than tested. Across 29 cities and two domains (hourly PM2.5 and 2 m temperature), under pre-specified protocols with measured energy and a cost-adjusted decision rule, we compare all three. On PM2.5 the zero-shot model and the tuned specialist are statistically indistinguishable (MASE 0.662 vs 0.692 under deployable causal covariates; formally equivalent at a 0.05-MASE margin only when the specialist is granted perfect-foresight covariates). The specialist's apparent temperature edge (0.533 vs 0.792) shrinks to a small, non-significant difference (0.745) once its covariates are restricted to values knowable at forecast time: a perfect-foresight artifact that can produce spurious cross-domain conclusions in any benchmark using future covariates. Transfer learning never significantly beats zero-shot at any budget. Under frequent retraining the untrained model uses roughly ten times less measured energy, though a once-trained specialist eventually amortizes. Where models are refreshed often, downloading one now matches training one.