Modeling interventions that include treatment times is an important task in many domains such as healthcare, finance, and others. Scaling up models for these interventional distributions is challenging, since popular architectures, e.g. transformers that predict the next event in a sequence, do not naturally support alterting times of specific treatments.We develop competing event models, an autoregressive generative approach in which estimating interventions in both when and what treatments are applied is made simple. The key element in our solution draws from the competing risks literature and models the timing of each type of event, e.g. a treatment or an observation, given that its the next to occur.This design allows straightforward treatment timing interventions via next-token prediction, admits a simple likelihood-based objective, and yields valid effect estimates under standard assumptions. We evaluate on a simulated benchmark for effect estimation of sequential treatments.