Level of detail and imposters work for reasons you may not fully understand.

SimonDev has a video on his channel talking about optimization, he goes over various topics and talks about different techniques and how it works.

Optimization through Levels of Detail (LOD): This technique reduces the complexity of objects as they move farther from the camera, using less detailed versions (or LODs). When objects are far away, the viewer cannot perceive fine details, so switching to simpler models reduces computational load. This frees up the GPU to handle more tasks without compromising visual quality.

Imposters and Billboards: These are extremely low-detail versions of objects that appear as 2D textures or "pictures" instead of 3D models. From far enough away, this technique is nearly indistinguishable from the real object and dramatically cuts down on GPU processing. This is used, for example, with large forests of trees in the distance.

Why It Works So Well: Reducing vertices and complexity of 3D models decreases the workload on the GPU, which has limited resources for processing vertices. Modern GPUs balance the workload between vertex and pixel shading, making it more efficient than older fixed-function pipeline architectures. However, even with powerful GPUs, unnecessary small triangles can drastically reduce performance because of the limitations in the primitive assembly and rasterization stages.

Primitive Assembly Bottleneck: The primitive assembly stage is where vertices are turned into triangles. This stage can become a bottleneck when you have too many small triangles because GPUs process them in batches (usually 2x2 pixel quads). Small triangles can waste GPU resources, as the GPU has to perform computations even for empty parts of these quads.

Nanite Technology in Unreal Engine: Nanite is Unreal Engine's automatic LOD system, which solves the problem of micro triangles by dynamically adjusting the level of detail based on camera distance, automating what previously required manual optimization. It avoids the GPU’s primitive assembly bottleneck by using a software rasterizer for tiny triangles, which is more efficient than traditional hardware rasterization.

The deeper explanation behind these optimizations reveals that while GPUs have tremendous power, efficiently managing resources (like handling vertices and triangles) is key to preventing performance bottlenecks, especially when rendering complex scenes in real-time games.