Many studies have demonstrated attenuated verbal working memory (WM) under articulatory suppression. However, performance is not completely abolished, suggesting a less efficient, non-articulatory mechanism for the maintenance of verbal information. The neural causes for the reduced efficiency of such a putative complementary maintenance system have not yet been addressed. The present study was conducted to fill this gap. Subjects performed a Sternberg task (a) under articulatory maintenance at low, high, and supracapacity set sizes and (b) under non-articulatory maintenance at low and high set sizes. With functional magnetic resonance imaging, set-size related increases in activity were compared between subvocal articulatory rehearsal and non-articulatory maintenance. First, the results replicate previous findings showing different networks underlying these two maintenance strategies. Second, activation of all key nodes of the articulatory maintenance network increased with the amount of memorized information, showing no plateau at high set sizes. In contrast, for non-articulatory maintenance, there was evidence for a plateau at high set sizes in all relevant areas of the network. Third, for articulatory maintenance, the non-articulatory maintenance network was additionally recruited at supracapacity set sizes, presumably to assist processing in this highly demanding condition. This is the first demonstration of differential neural bottlenecks for articulatory and non-articulatory maintenance. This study adds to our understanding of the performance differences between these two strategies supporting verbal WM.