Tried uploading simple models with a cube instanced at multiple places both as fbx and gltf to 3D Tiles but once I’ve done that the resulting mesh consist of hundreds of polygons.
From this user story Swiss Geospatial Portal: Switzerland’s 3D Map Viewer for Federal Infrastructure Data – Cesium it certainly looks possible.
The information needed to use Instanced Tiling is dependent on the data source. For example, if you use CityGML or KML+COLLADA (which is used in the Swiss example), then the pipelines can identify the instancing and take advantage of it.
In the 3D Model Tiling Pipeline (FBX, glTF etc), at this time the pipelines do not take advantage of instancing as the algorithms to tile and create LODs are different. Instead of instancing, the meshes are merged and batched to provide optimized rendering. The example you gave of cubes being instanced is not usually the input for 3D Tiling, so the pipelines are optimized more for massive photogrammetry meshes and 3D models.
Are you able to share a sample of your data for us to look at?