Prefill vs. decode: Why the same model feels fast in one product and slow in another

There's a question that comes up a lot when teams start measuring AI performance. "Why does our model feel slow? We're using the same model as everyone else." The answer almost always lives in the sam