It's because CSS transforms that don't affect layout happen on the GPU, so they may as well be polygons. In fact ThreeJS, the most prominent JS library for 3D rendering in WebGL, can work in DOM rendering mode instead of WebGL :)

I've been playing around with some very extreme forms of server side rendering, and have come to discover that all modern browsers can actually decode and display 1080p+ jpeg images streamed over websocket fast enough to be indistinguishable from a x264 video playing at the same frame rate (60fps in my testing). I have tried on ARM/mobile and it's the same success story there. Latency almost feels lower on my S10 than my Threadripper 2950X.

There's lots of things a browser can do that would surprise a lot of developers.