New techniques in large language model inference are optimizing the critical trade-off between latency and throughput. These methods either shift a deployment's position along the existing efficiency frontier or expand the boundary entirely to unlock greater overall system performance.