Abstract: The unstructured grids in power systems represent complex network problems which in turn often rely on sparse numerical solvers in their algebra kernels. This problem class’ structure, or lack thereof, leads to potential performance bottlenecks on the upcoming GPU based supercomputer generation. This talk shows our implementation of a power flow application (ExaPF) that leverages the compactness and flexibility of the programming language Julia to run fully on the GPU without any transfer from host to device during the main computational loop. This includes Newton-Raphson, automatically differentiated derivatives, and a preconditioned (block Jacobi) iterative linear solver (BiCGSTAB). We use this application to extrapolate its runtime behavior to optimal power flow.