Memory requirement will remain a big issue until GPU memory becomes scalable through multiple G/C or a virtual gpu memory system (gpu ram - cpu ram - hdd) is built. I actually experienced many failures rendering (blender cycles) with 20million+ triangles on 980 ti (6GB system) due to the lack of memory.
Tesla… I would say please don’t overestimate Tesla. The fastest GPU is actually driven by Geforce cards. Teslas are targetted for enterprise market. It has more expensive capacitors and ECC ram (error correction), but it is in general slower than the best geforce card (e.g., 1080). The reason why supercomputers use Tesla is they need to put so many cards (100+) there. In such a situation error rate is very important to both system engineers and customers. But for indie developers, 4-way geforce 1080 (I bet the performance is better than 8 best teslas) is much more adequate for use. Tesla also requires server O/S, server M/B, etc. which means many headaches…