The success of modern machine learning has been fueled by increasingly large models, with parameter counts scaling from millions to billions and now even trillions. While these models demonstrate extraordinary capability, their size creates a serious obstacle to practical deployment. Training costs are already immense, but the larger challenge lies in inference. Serving models at scale requires memory capacity, compute throughput, and latency reduction that exceed what many organizations can afford. Without strategies to compress models and optimize inference, even the most powerful systems remain impractical for widespread use.