Financial frictions are a prominent feature of modern macroeconomic analysis. Using a novel loan-level dataset, we evaluate whether leading theories for these frictions are broadly consistent with observed interest rates. We find that they are not. There is a lot of dispersion in rates paid by firms, even on observationally equivalent loans, and risk---the dominant theory for dispersion---can account for less than 15\% of this. Alternative theories for dispersion, based on banks and firm-bank relationships, also have quantitatively little explanatory power. Instead, the data dictates that the major source of variation is persistent idiosyncratic firm characteristics that are not closely connected to economic size or performance. We provide a search-based theory of firm financing that can rationalize these facts.