That turquoise stuff and they skybox definitely aren’t real polygons. The trees on the left are kind of transparent. This is just midjourney or similar.
However I reckon existing transformer architectures would manage to output 3D models just fine given sufficient training data and compute power. Generating grammatically-correct files is what GitHub copilot already does with other languages (I tried, it does not seem to support .obj though). It would definitely be interesting for procedural game generation, in fact I would be surprised to learn that nobody is working on something like this.