One way to think about this, and not get hung on the word "suffering" that may be too corporeal:

Could a future AI thing have its own drives? (As in https://en.wikipedia.org/wiki/Drive_theory) We could probably make a current-day LLM already exhibit basic outward signs of this with just a system prompt. Now consider adding memory that over time retrains the weights, allowing for behavioral drift.

Would depriving it of fulfilling those drives be acceptable?