If what people go for are AAA games with hyperdetailed graphics and massive playing spaces, the tendency is for games to grow in size (all those highly detailed textures and masses of data for terrain and object placement really add up) and the only alternatives for trying to deliver some of that using less data, such as No Man's Sky and their heavy use of generation, end up with results that quickly feel repetitive after some playing and an inferior experience on the adventure side than a carefully crafted gamespace with carefully crafted chracters and encounters.

There are plenty of smaller games from indies which focus mostly on engaging game mechanics and hence are much smaller datawise, but if all you're going for is something like GTA or Fallout, don't be surprised when the tens of thousands of highly detailed objects and characters, days worth of voice data and hundreds of square kilometers of gameplay area translate into more than 100GB.

Mind you, the industry uses tons of generation in game making (nobody is going around making, say, the various maps in a chainmail texture by hand) but it's all vetted and costumized by actual people and the best results end up properly fitted to the models and stored as mainly static stuff in the game data files so big and varied gameplay ares will add up to lots of data even if a lot of it was done with the help of generative tools.

So far and from what I've seen, unsupervised AI can't really deliver good results in a lot of that, so whilst it will probably be a massive leap foward in the area of generative tools for game making, you will still end up with massive game data files containing the output of the AI generation, carefully curated and even customised by actual humans.