Yes they can render it in-engine, but they can use bespoke character animations which don't have any blending artifacts, and they can put the camera on a rail and tune the assets and particle effects until they are certain every single frame can be rendered in under 16ms. They can hide billboards and other rendering tricks and be certain the camera is never going to hit them at an angle which gives them away.
Real-world game play scenarios are much more unpredictable, and the results are likely to fall well shy of this mark.
I mean... that lighting was really, really good, and I think this is the first triangle-based demo where the surfaces really don't look oddly angular almost anywhere (maybe with exception of the stalactites). But it's a far cry from the hand-tuned look of something prerendered.