The NEM-sensitive fusion protein, NSF, together with SNAPs (soluble NSF attachment proteins) and the SNAREs (SNAP receptors), is thought to be generally used for the fusion of transport vesicles to their target membranes. NSF is a homotrimer whose polypeptide subunits are made up of three distinct domains: an amino-terminal domain (N) and two homologous ATP-binding domains (D1 and D2). Mutants of NSF were produced in which either the order or composition of the three domains were altered. These mutants could not support intra-Golgi transport, but they indicated that the D2 domain was required for trimerization of the NSF subunits. Mutations of the first ATP-binding site that affected either the binding (K266A) or hydrolysis (E329Q) of ATP completely eliminated NSF activity. The hydrolysis mutant was an effective, reversible inhibitor of Golgi transport with an IC50 of 125 ng/50 microliters assay. Mutants in the second ATP-binding site (binding, K549A; hydrolysis, D604Q) had either 14 or 42% the specific activity of the wild-type protein, respectively. Using coexpression of an inactive mutant with wild-type subunits, it was possible to produce a recombinant form of trimeric NSF that contained a mixture of subunits. The mixed NSF trimers were inactive, even when only one mutant subunit was present, suggesting that NSF action requires each of the three subunits in a concerted mechanism. These studies demonstrate that the ability of the D1 domain to hydrolyze ATP is required for NSF activity and, therefore is required for membrane fusion. The D2 domain is required for trimerization, but its ability to hydrolyze ATP is not absolutely required for NSF function.