The cost of visualizing computational fluid dynamics and other flow field data sets is increasing rapidly due to ever-increasing grid sizes that constantly strain platform memory capacity and bandwidth. To address this problem of 'big data', techniques have been developed in two areas: out-of-core visualization, which exploits the fact that most flow visualizations require a very sparse traversal of the data set, and remote visualization, in which images are rendered by large-scale computing systems and transmitted via network to desktop systems. A new method, which combines out-of-core and remote techniques, offers a potentially significant improvement in both scalability and cost. By incorporating new techniques for spatial partitioning, data prediction, and explicit memory management, this new method enables desktop computing applications to selectively read the contents of massive data sets from remote servers connected by local or wide area networks. Initial testing has shown that local memory usage is nearly independent of data ste size, overcoming the key limitation of prior out- of-core methods. By performing the visualization computations and graphics rendering on the local/desktop platform, the new method also provides a significant improvement in price-performance ratio compared to current remote visualization methods.