I know that GPULM currently only uses the proxy meshes, which generally results in low quality/artifacted GI. I wouldn’t be surprised if CPULM alone supports full-quality Nanite in the near-term, as the amount of VRAM needed to hold every Nanite mesh in a complex scene, would be quite large.
Generally speaking, Lumen and baked lighting don’t play well together, including Baked GI and lumen reflections. What the documentation suggests for scalablility to lower-level hardware is to use a combination of DFAO and SSR/SSGI, which generally lets you keep art direction the same for both. Lumen also supports baked AO on static meshes, which brings down the cost as well.
The biggest problem you may run into is dynamic reflections, but SSR + cubemaps + well-planned planar reflections can still work great if you have the time to configure and optimize. DFAO handles specular occlusion as well, so it can help minimize light leaking.