A Systematic Test of Physics Integration for Fleet-Scale Data Center Load Forecasting

Data center cooling systems are designed with first-principles thermodynamic models, and it is natural to expect such a simulator to also improve hourly load forecasting. We test that expectation across a fleet of 200+ data centers over a full seasonal test year, injecting physics information into strong forecasting hosts through every available mechanism, from engineered features in a gradient-boosted model to in-context covariates and fine-tuning covariates for a time-series foundation model. The answer is largely no. Feature-based integration is marginal or harmful, and in-context injection degrades accuracy or fails falsification checks. The one effect that survives is small. When physics enters as a covariate during parameter-efficient fine-tuning it yields a consistent improvement that grows with horizon (+0.008 points at week two, $p=1.5\times10^{-10}$) yet remains operationally marginal. A pre-registered test shows the benefit is not mediated by simulator calibration quality. The evaluation machinery built for these tests, applied to a public foundation model with weather covariates and fleet fine-tuning, reaches 0.846% median day-ahead error, 48% better than persistence.