I think they'll just be used to force VDI on consumers.

Which is a shame, because VDI could be amazing for consumers, but it won't be.

Mass surveillance, is of questionable value beyond just creating a database of footage. Police DON'T aim to "solve all crime and be good guys". They barely aim for investigating all crimes, and most crime they'd just need a time and date for to look at the footage only IF and WHEN a crime is reported. They're not going to spend police resources to have cops sift through footage the AI has guessed there's crime in every day. That would just cost time/money. At best it might be used for facial recognition at protests - but facial recognition systems predate AI, so it's superfluous.

war games, takes appropriate training data which is increadibly difficult to simulate and get the right answers from. There are already various war-gaming departments who have been trying to do this. The YouTube channel People Make Games has an episode on this - war just has so many variables that beyond the basics of troops numbers and geography it's kind of difficult. But even if this becomes a use, it's niche, it wouldn't require a large amount of compute, each country only has one military... And spending on research projects (eg. DARPA) isn't new. It's not enough to be a "big win" for 3 AI giants, it's more a step back to where they came from.

predictive policing (and population control) hang on, wasn't this your first example? Again, mis-matches and police time checking the AI are going to cost too much, and those programs would cost big money in lawsuits and time. It'd be the first program cut by most departments.

I see nothing realistic here.

More likely uses are jobs that burn through a lot of copy writers: spam/marketing, realestate ads, search engines and researching, little buddies and toys, publishing (sadly), video games, weather bureaus, text transcription and translation, scams and telemarketing. Things there's lots of data on, and close to zero risk with.