Using novel AI patent data, we document significant production of AI innovation as early as 1990. Then, focusing on publicly traded firms, we show that their AI production is motivated by the mutually reinforcing effect of their innovation capacity and AI exposure. We use corresponding exogenous variation in R&D stock and AI exposure to instrument for AI production. Producing AI creates firm value through a large, permanent decrease in risk (both systematic and idiosyncratic), rather through profitability. Further evidence suggests that AI lowers physical capital intensity and increases bargaining power for producing firms. However, AI production increases future stock returns (roughly 5% per year). Coupled with no evidence of investor learning over three decades, the results suggest that AI innovation is undervalued by investors. We empirically distinguish production of AI innovation from AI adoption, automation, general technology, and other potential confounds.