“Achieving model speeds of 10,000 tokens per second would be highly beneficial for tasks like chain-of-thought reasoning, parallel rollouts, and extensive code generation with verification.”