Background: Optimal timing and criteria for implementation of androgen deprivation therapy (ADT) in men with relapsed M0 prostate cancer (PCa) remains undefined. Early ADT induces the non-metastatic castration resistant PCa (nmCRPC) clinical state. FDA approved apalutamide (SPARTAN) and enzalutamide (PROSPER) for nmCRPC based on prolongation of metastasis free survival (MFS) which is now considered a valid endpoint for drug approval. Because overall survival (OS) in PCa is usually long and long-term ADT is associated with irreversible adverse events and high costs, we sought to evaluate OS and other outcomes of men with relapsed PCa and ADT deferred until metastasis. Methods: Retrospective review of 2,636 men who had radical prostatectomy (RP) between 1981-2017 and developed biochemically recurrent PCa from a single-institution. Patients who received ADT prior to metastasis were excluded. Kaplan-Meier survival estimates of MFS and OS were defined from RP to event or censor. Multivariable Cox proportional hazards regression was used to identify prognostic factors. Results: 1,686 men treated with deferred ADT until metastasis or censored metastasis-free were eligible. Medians: follow up 10 years (IQR5-16), age 60 years, PSADT 33 months, Gleason < 7 (24%), Gleason 7 (55%), Gleason > 7 (21%). 688 (41%) received salvage radiotherapy. Median MFS and OS were 21 and 22 years, respectively (Table). In multivariable models, age (HR 1.06, 95% CI 1.04-1.1), Gleason < 8 vs ≥8 (HR 0.4; 0.3-0.5) , RP stage (organ confined vs not 0.6; 0.5-0.8), PSADT (0.995, 0.993-0.997) and salvage RT (0.88; 0.81, 0.96) were associated with OS. Conclusions: Deferred ADT in relapsing M0 patients is associated with long OS measured from time of local treatment, comparable to OS with salvage ADT in contemporary experience. Drug approval trials in nmCRPC should focus on patients at high risk for metastasis and death prior to ADT, and determine standardized criteria for initiation of ADT. Prolongation of MFS in nmCRPC requires further validation and may not necessarily reflect a net OS benefit.