Latency is of course the biggest concern. Recent videos I saw do not indicate they have a good solution to latency. I'd absolutely love if they used some sort of 'asynchronous reprojection' like you see in VR.

Essentially, send a depth buffer along with video data, and let the client reproject frames before it gets the next keyframe from the server. It can help solve input lag for movement and turning. Not so much for things like shooting.

Wouldn't that require a decent graphics card, largely negating the point?

Not too much for reprojection. That's a very specialized task that can be optimized for in hardware, without having to support all the typical operations of GPUs. That's not to say it's easy. It requires a large amount of engine integration and the client has to now know what amount of reprojection to apply when you turn, which may vary per game or context.

Hugely hugely impacts the complexity.

People used to (still?) say people wouldn't enjoy VR because they'd feel motion sick also.

There are games that do make me feel nauseous occasionally, but that hasn't stopped me from putting over 1,200 hours into the games that don't. Even if latency is an issue on Stadia (contrary to what people who've tried it so far have said), there's gonna be plenty of games that don't need the lowest of latencies to enjoy.