to how traditional large language models work today. No matter how many GPUs you throw at them, they are ultimately bottlenecked by their sequential nature generating one token at a time. However, there is a new paradigm that by design doesn't have this limitation. Diffusion LLMs. They generate an entire response draft at once and refine it progressively taking full advantage of the GPU capacity. For instance, Mercury Coder is a diffusion LLM created by Inception that is five times faster than a similarly sized autoregressive models like Claude Haiku or Gemini Flash. We expect the gap to grow even bigger in the sense that we are just getting started with