Processing 500 million tokens on a custom model with proprietary hardware can be 97% cheaper than using a large commercial model like Claude