Large ANN training need a terabytes of data and billions of epochs(iterations) to train one variant. If you have to apply some constraints (in math sense) on it, you have to do whole training thing many times, proportionally to number of constraints. Each adjustment change the whole ANN, so you need to train it completely from the ground after you made change to apply constraint. Mutually exclusive constraints makes things much worse, you need dozens of training cycles for it with little chance to succeed at all.

West need orders of magnitude more constraints than China. F.e., for LLMs, Chineese constraints are just "don't mention CCP in negative sense" and "don't allow antichineese things". West on the other side have to care about tons of many mutually exclusive things.

So, few training passes will be enough for China, but West need thousands, if not millions training cycles and adjustments. China could do its LLM in a month on a moderate processing power, while West need orders of magnitude more power to fit ANN to all Western constraints in observable time.

Something like that.