Say you have a simple 5000 triangle horse with a modest 16 frames of animation. With the current average of 10 bytes per triangle, each frame of animation will eat roughly 50KB of bandwidth, making this 5000 triangle object use the same amount of bandwidth and storage as an 80,000 triangle object and use nearly 1MB of bandwidth to download all frames and LODs. Once the object is decoded, its memory footprint per frame would be around 250KB of vertex buffer data, or 4MB for the entire 16 frames, making it not practical to hold all frames in memory, so even after an object is downloaded, flipping frames would cause much disk access and frame transitions would be slow.