I think that unless (until?) OpenAI releases information about the model itself and the inference engine it runs on, everything is just speculation. Clearly, there's impressive ML and systems engineering at play with GPT-3.5-turbo given how capable, fast, and scalable to their customer base it is.