Vulkan has a very verbose setup phase, but after that you only make some tiny calls here and there (Draw a frame, update some UBOs), so the CPU load isn't that high because you record most things at start-up.
Multi-threading is possible with SharedArrayBuffers [0] and N-API allows to take the memory address of it. NVK is only a thin layer above Vulkan and JavaScript <-> C++ interopability is pretty fast. You only notice some overhead when making thousands of calls, because the C++ code cannot be inlined. But again, this is not something common in Vulkan.
Also JavaScript doesn't have essential features like Compute shaders for years now. Vulkan has all of it. Even though WebGPU [1] is in development, I don't expect it to give access to the Vulkan RTX pipeline anytime soon.
[0] SharedArrayBuffer: https://v8docs.nodesource.com/node-4.8/d4/db8/classv8_1_1_sh...
[1] WebGPU: https://github.com/gpuweb/gpuweb
