Deep learning applications require powerful multi-GPU systems for development and operation, which can be very expensive to rent in the cloud for long-term operations. Which infrastructure offers the best compromise between time-to-solution, cost-to-solution and availability of resources?