The expected shortfall is defined by \begin{align*} ES_{\alpha} = \frac{1}{1-\alpha}\int_{\alpha}^1 VaR_{p}(L) dp, \end{align*} where $L$ is the loss function. For the case with 500 scenarios, the $\alpha=99\,\%$ percentile VaR is approximately the $5^{\rm th}$ worst loss scenario. The expected shortfall can then be approximated by the average of the 5 worst losses, times $-1$ (we take $ES_{\alpha}$ to be positive). That is, \begin{align*} ES_{\alpha}(L) = -\frac{1}{5}\sum_{i=1}^5 L(i), \end{align*} where $L(i)$ is the $i^{\rm th}$ worst loss scenario. Assuming that the loss $L$ can be decomposed into the losses $L_1$ and $L_2$ from two sub-portfolios. That is $$L=L_1+L_2.$$ Then $$L(i) = L_1(i)+L_2(i).$$ However, it is easy to see that, though $L(1), \ldots, L(5)$ are the 5 worst loss scenarios of $L$, $L_j(1), \ldots, L_j(5)$, for $j=1, 2$, are not necessarily the 5 worst loss scenarios for $L_j$. In other words, for $j=1, 2$, \begin{align*} ES_{\alpha}(L_j) \ge -\frac{1}{5}\sum_{i=1}^5 L_j(i). \end{align*} Then \begin{align*} ES_{\alpha}(L) &= -\frac{1}{5}\sum_{i=1}^5 L(i)\\ &=-\frac{1}{5}\sum_{i=1}^5 L_1(i)-\frac{1}{5}\sum_{i=1}^5 L_2(i)\\ &\le ES_{\alpha}(L_1) + ES_{\alpha}(L_2). \end{align*}