Who's Afraid of Chinese Models? The Return of Marginal Costs in AI
Who's Afraid of Chinese Models?
I realized that the era of zero marginal costs in software is ending as AI inference costs resurface. While open weights models like Kimi K3 reduce R&D expenses, the real cost lies in serving tokens. Intelligence is becoming a commodity where profitability depends on superior cost structures rather than higher prices, fundamentally shifting the industry dynamics.
"What is fungible is what is constructed from tokens, which is to say intelligence."
HN discussion
- LLMs operate under manufacturing economics rather than software economics, where marginal costs of inference determine profitability and US labs currently lead in token efficiency over Chinese models.
- Chinese AI development relies on a centralized data strategy where the government aggregates web crawls and API usage traces from American models to train domestic companies, bypassing the need for independent data acquisition.
- Banning distillation clauses is legally feasible as governments can declare specific contract terms unenforceable, similar to how they regulate other commercial provisions.
- Open weights models offer superior auditability compared to closed systems, though they may still conceal undetectable backdoors.
- Model identity hallucinations, such as a model claiming to be Claude, are common artifacts of training data rather than evidence of covert distillation or security breaches.