I mean yes, current large models are essentially compressing incredible amounts of content into something manageable by a single Accelerator/GPU, and making it available for retrieval through inference.