Well, when people say 50% overhead (the lower bound), they mean that the cost to the grant is `1.5*direct_cost`, not that half the grant goes to indirect costs.
Anyway, research is the primary motivation for purchasing journal subscriptions. It does not make financial sense for a non-research institution to subscribe to many journals. So regardless of which budget the subscriptions are formally taken from, they come from, it should be considered a cost of research. Making the system more efficient frees up money somewhere that reduces the cost of (externally- and internally-funded) research.
Anyway, research is the primary motivation for purchasing journal subscriptions. It does not make financial sense for a non-research institution to subscribe to many journals. So regardless of which budget the subscriptions are formally taken from, they come from, it should be considered a cost of research. Making the system more efficient frees up money somewhere that reduces the cost of (externally- and internally-funded) research.