Also, they found that to get the best performance from it, it sucked down in the region of 20gb per hour of data. Which is a lot for countries with capped internet allowances.
Also I suspect this is not something that will sit on any CDNs due to the specialist nature so this is going to transit much further than the ISP's edge.
Equally, because you know so much about the scene, there are many more opportunities to do smart video compression - however, those would be extremely difficult to take advantage of.
Ultimately they could cache textures, models etc. on the client system and only transmit data about the players' locations and inputs. Then the codec would render the game on the client... hey wait.
They could potentially implement something similar to Retroarch's runahead feature, but I suppose it would requires support from the game itself and more infrastructure so probably not feasible...