Sure, generally speaking. Is that true for static, fixed parameter count LLMs like GPT4?

I think you're hand waving a lot just to claim that OpenAI are (somehow) reducing accuracy of their models during high load. And I'm not sure why.