Because that's how it actually works. It generates the text token by token and you're seeing it live. GPT-3 Lab does the same, you can see when it gets slower/faster.