This proposal led by the University of Oxford, with support from the Alan Turing Institute (ATI), Bristol, Edinburgh, KCL, QMUL, Sheffield, Southampton and UCL is for a national GPU system that will support multidisciplinary science with a focus on machine learning and molecular dynamics. The architecture is based on ``fat'' GPU compute nodes, with 8 of NVIDIA's new Pascal GPUs, each with a) 16GB 720GB/s HBM2 memory, b) an 80GB/s NVlink interconnect to other GPUs, c) 6GB/s bandwidth to main system memory, d) 6GB/s bandwidth to the Infiniband external network. Each server also has two 20-core Xeons, 512 GB DDR4 memory and 8TB SSD. The motivation for selecting this architecture is the huge growth in research in machine learning and associated areas of data science within the UK, particularly within the universities which are members of the Alan Turing Institute, or SES. The same architecture is also ideally suited for molecular dynamics, medical imaging and a number of other application areas. The system will be run as a national facility, similar to Archer in being free to all academic users with computing time available to all through a lightweight Resource Allocation Panel, with a top-level steering committee determining the policy on resource allocation between the different application areas (Machine Learning, Molecular Dynamics, Other).