Your point is badly conceived. Just because China built a model that requires less compute than the US ones doesn’t mean thar compute doesn’t need to be scaled. The compute isn’t built for model training long-term. It’s for inference and real-world use cases.

→ View original post on X — @ninadschick,