CRD — variance of a group mean minus the grand mean

In a completely randomized design suppose $\bar y_{i\cdot}$ is computed from $n_i$ observations ($n_i$ not all equal with $\sum_{i=1}^h n_i=N$), then what will be $\operatorname{Var}(\bar y_{i\cdot}-\bar y_{\cdot\cdot})$?