The duration of large wildfires (> 400 ha) in the western United States has shown a significant uptrend over the course of 1992–2020 (with increasing rates of 0.76 days yr^-1 and 0.55 days yr^-1 in the summertime and fall, respectively). However, the exact mechanism driving such a fire duration increase is unknown, posing challenges for predictions of fire variability and risk assessments. To address this problem, we employ multiple linear regression models to investigate the factors driving such observed trends and variability of large fire duration. Our analysis reveals that the maximum daily wind speed during large fire events U_dmax,fire, which also possesses an uptrend during 1992–2020, is the key predictor explaining the fire duration variability in both seasons. Surprisingly, while U_dmax,fire has a historically rising trend, there is no significant trend in the background wind or mean wind conditions. Our statistical analysis suggests that prolonged fires could generate enhanced boundary-layer wind fluctuations and thus U_dmax,fire, which might then possibly lead to further longer fire duration. We hypothesize that the robust correlation between the historical uptrends in both fire duration and U_dmax,fire are due to not only (1) the elevated buoyancy-driven wind fluctuations by fires but also (2) a potential dynamical feedback by the elevated U_dmax,fire and wind fluctuations that make fires even harder to contain. These findings highlight the growing challenges in fire control efforts as wildfires become more intense and persistent, driven in part by changing wind dynamics induced by the larger fires.