The explosion of video generation and consumption, coupled with an inadequate rise in network bandwidth has led to network delays and decreased Quality of Experience, limiting the opportunities to tap into the full potential of video data. These deficiencies in network resources with a shift to cloud computing models have resulted in the need to revisit the overall mechanism for video transfer and storage of videos between edge devices and the cloud. We propose a novel multi-scale real-time super-resolution convolutional neural network to achieve the composite task of optimizing the entire cost of video transfer with minimal loss of quality that can be used for any application involving the transfer of video data. To achieve this, we develop a cost-optimized video transfer system that optimizes the metrics of video transfer to give the best quality video output, given the user budget. The model makes use of Convolution blocks for extracting features and output creation with multiple sub-pixel convolutions in a novel structure. For upscaling to full High Definition video at 30 fps, the model successfully retained the frame rate while the system achieved savings in transfer time and bandwidth usage. This model has been trained on surveillance videos (VIRAT dataset), but consistent results were obtained during testing even on feature films and sports videos which demonstrates its content invariance. The evaluation of our approach averaged over 376 videos, yielded meager quality losses of 8%, measured by a novel non-referential quality metric, also proposed in this paper. Additionally, average network bandwidth savings of 80% and average video transfer time reduction of 52% were achieved.