I never thought that game streaming was impossible, because the data involved is miniscule, and all calculations are done remotely. It's no different in many respects to me using PCAnywhere in the 90s over ISDN to support clients in Israel and Japan with remote desktop.
I also don't think you understand what I mean by async. Once you have a big event happening with a ton of pieces moving, all that has to be synced out between client and server, otherwise you have no real interaction with those objects possible without breaking the physics model.
Let's get a decent basic hypothetical : you are playing a game where you control a player who can untie a ribbon and release hundreds of various size balls from the ceiling in a net. These balls vary in size and weight. Half of the floor is rubber, half linoleum. The graphics are all rendered by your console. Okay, pull the string, and everything tumbles down and starts bouncing. This is an event being calculated by the server, server tells the client where things are, client draws them. Now walk up and start bouncing the balls against each other. You are now changing the variables dramatically with every action, and that input needs to get back to the server, and there is no longer a 'set piece' in place, but rather a constant variable series of actions that all have immediate consequences. If there is lag, then the physics model won't appear realistic, or worse, it will fail to render and cause client side issues/glitching.
Does that make sense? Because a big set-piece action triggered by the user, but that is not interactive beyond the start/end of it, is useless. We can already do that locally. The bigger stuff, huge synchronous cloud to client and back physics events of a continually changing cycle : THAT is both what would be incredibly cool AND is ridiculously impractical.
It bears repeating that dedicated hardware physx cards start having problems when they're moved from a bus of 2000MB/sec to one that is 500MB/sec. Moving that same physx engine to one that exists on 1MB/sec (8 megabit) is astonishingly unlikely to be impressive or effective.