In-line holography has recently made the transition from silver-halide based recording media, with laser reconstruction, to recording with large-area pixel detectors and computer-based reconstruction. This form of holographic imaging is an established technique for the study of fine particulates, such as cloud or fuel droplets, marine plankton and alluvial sediments, and enables a true 3D object field to be recorded at high resolution over a considerable depth. The move to digital holography promises rapid, if not instantaneous, feedback as it avoids the need for the time-consuming chemical development of plates or film film and a dedicated replay system, but with the growing use of video-rate holographic recording, and the desire to reconstruct fully every frame, the computational challenge becomes considerable. To replay a digital hologram a 2D FFT must be calculated for every depth slice desired in the replayed image volume. A typical hologram of ~100 micrometre particles over a depth of a few hundred millimetres will require (10^3) 2D FFT operations to be performed on a hologram of typically a few million pixels. In this paper we discuss the technical challenges in converting our existing reconstruction code to make efficient use of NVIDIA CUDA-based GPU cards and show how near real-time video slice reconstruction can be obtained with holograms as large as 4096 by 4096 pixels. Our performance to date for a number of different NVIDIA GPU running under both Linux and Microsoft Windows is presented. We consider the implications for grid and cloud computing, and the extent to which GPU can replace these approaches, when the important step of locating focussed objects within a reconstructed volume is included.