We used the patch arid land simulator (PALS-FT) -- a simple, mechanistic ecosystem model -- to explore long-term variation in evapotranspiration (ET) as a function of variability in rainfall and plant functional type (FT) at a warm desert site in southern New Mexico. PALS-FT predicts soil evaporation and plant transpiration of a canopy composed of five principal plant FTs: annuals, perennial forbs, C₄ grasses, sub-shrubs, and evergreen shrubs. For each FT, the fractional contribution to transpiration depends upon phenological activity and cover as well as daily leaf stomatal conductance, which is a function of plant water potential, calculated from root-weighted soil water potential in six soil layers. Simulations of wat6er loss from two plant community types (grass- vs. shrub-dominated) were carried out for the Jornada Basin, New Mexico, using 100 years of daily precipitation data (1891-1990). In order to emphasize variability associated with rainfall and fundamental differences in FT composition between communities, the seasonal patterns cover of perennials were held constant from year to year. Because the relative amount of year to year cover of winter and summer annual species is highly variable in this ecosystem, we examined their influence on model predictions of ET by allowing their cover to be variable, fixed, or absent. Over the entire 100-yr period, total annual ET is highly correlated with total annual rainfall in both community types, although T and E alone are less strongly correlated with rainfall, and variation in transpiration is nearly 3 times greater than evaporation and 2 times greater than variation in rainfall (CV of rainfall = 35%). Water use shows a relatively high similarity between the grass- and shrub-dominated communities, with a 100-yr average T/ET of 34% for both communities. However, based on a year-by-year comparison between communities, T/ET was significantly greater in the grass-dominated community, reflecting the fact that over the long term more than half of the rain occurs in the summer and is used slightly more efficiently (T>E) by the C₄-grass community than the shrub community, although we found some rainfall patterns that resulted in much greater T/ET in the shrub community in a given year. Percent of water lost as transpiration (T/ET) suggests that while there is a general trend toward increased T/ET with rainfall in both community types, T/ET is extremely variable over the 100-yr simulation, especially for normal and below normal amounts of rainfall (T/ET values range from 1 to 58% for the grass-dominated site and 6 to 60% for the shrub-dominated site). These predictions suggest that because of the relatively shallow distribution of soil water, there is little opportunity for vertical partitioning of the soil water resource by differential rooting depths of the plant FTs. However, functional types may avoid competition by keying on particular "windows" of moisture availability via differences in phenologies. We found very little differences in average, long-term model predictions of T, E, and ET when annual plant cover was variable, fixed, or absent. The results of our simulations help reconcile some of the disparate conclusions drawn from experimental studies about the relative contribution of transpiration vs. evaporation to total evapotranspiration, primarily by revealing the great year-to-year variability that is possible.