Note that, for any $x$ and any $i$, $f_i(x) \le f(x)$. Hence, if $x, y$ are elements of the domain and $\lambda \in [0, 1]$, then for any index $i$,
$$f_i(\lambda x + (1 - \lambda)y) \le \lambda f_i(x) + (1 - \lambda)f_i(y) \le \lambda f(x) + (1 - \lambda)f(y).$$
Consider the set of values of $f_i(\lambda x + (1 - \lambda)y)$, as $i$ ranges over the index set. We have just shown that this set bounded above by $\lambda f(x) + (1 - \lambda)f(y)$. However, by definition of $f$, its least upper bound is $f(\lambda x + (1 - \lambda)y)$. Therefore,
$$f(\lambda x + (1 - \lambda)y) \le \lambda f(x) + (1 - \lambda)f(y),$$
proving convexity.