Abstract: Knowledge of the spatial distribution of total carbon is important for understanding the impact of regional land use change on the global carbon cycle. We studied spatial total carbon variability using transect sampling in an Imperata grassland area. Spatial variability was modeled following an isotropic stationary process with spherical and exponential variogram functions. Range and sill were estimated at 100 m and 82.29 ton2 ha−2, respectively. For nugget, sill ratio was estimated at 24%, implying a rather strong spatial dependence. In a subsequent total carbon stock inventory based on the sampling design mentioned above, we applied three types of estimators, namely, “naive average procedure,” “spatial average procedure,” and “spatial optimal procedure.” Estimation of total carbon stock (in ton ha−1) following naive average procedure (which erroneously ignores the spatial dependence) resulted in a considerably too narrow 95% confidence interval of 37.52 to 39.75, whereas the outcomes using spatial average procedure and spatial optimal procedure were 36.54 to 40.73 and 37.14 to 40.78), respectively, using the spherical model, and 36.63 to 40.64 and 37.07 to 40.64, respectively, using the exponential model. Our research indicated that, when total carbon stock estimation is the main goal, random sampling is optimal, whereas wide design sampling (i.e., shortest distance between sampling locations not less than the range) can be preferred in some cases.