Don't use cloud stuff. Just run your LLMs locally, if you have to. But you have an old GPU? Just take a smaller model. "But the quality of smaller/local models is worse?" Well, since you are using LLMs, I assumed that you didn't care about quality. (There are some proper use cases for LLMs, especially regarding accessibility. I am totally for using LLMs in that area.) But in general: You want bad text, bad pictures, bad code. Stop clinging to old concepts of "better or worse". You want quick, not good.