It’s not necessarily that any of the open models are cheaper to train (for similar model size), it’s more that China has deeper pockets. And MoE inference is cheaper than dense models I believe