The first wave of successful GPU accelerations has been crowded with highly-parallel methods that adapted well to the hardware. But the easy-pickings are now running out. The truly challenging applications require "going back to the algorithmic drawing board." To develop new versions of the most effective fast algorithms, such that our science can most benefit, an ideal environment is created by the open software model, where efforts can be shared. We will describe one area of application --electrostatics of biomolecules in solution-- where we see at work the triad of extreme computing: fast algorithms, open software, and heterogeneous computing.