Apparently Qt is not significantly better: a Telegram client consumes about 400 MB, and a Qt-based music player Strawberry, about 100 MB.
I can imagine that GPU-based compositing can consume a lot of RAM for the textures used in rendering, but frankly an entire 4K screen at 32bpp is less than 32 MB, while both Telegram and Strawberry take up a small portion of it.
Or, from another angle: can a modern GUI be made to consume little RAM while remaining performant and using a GPU for rendering on larger / high-DPI surfaces?