Chinese AI and chips are ahead of where export controls assumed
Two threads landed together today that tell a single story. The TOP500 list at ISC'26 revealed a new number-one supercomputer built with Chinese interconnects and Chinese chips, the first to cross 2 exaflops. Commenters called it 'a wake-up call' without hedging. Separately, GLM-5.2 is being benchmarked against Claude in security tasks, and one commenter is already predicting Commerce will force OpenRouter and HuggingFace to delist Chinese open models within months.
The pattern: the US export control strategy assumed a gap that is narrowing faster than policy can track. A separate thread on knowledge distillation of black-box LLMs, published in 2024, resurfaced today, with commenters noting that distillation techniques let Chinese labs extract capability from US models without direct access.
The counterpoint from the threads: some argue banning models from HuggingFace 'wouldn't make any sense' since the weights are already distributed. But that's exactly what makes this a harder problem than chips, you can embargo a fab, you can't embargo a torrent.
So what?
If you're building on the assumption that US frontier models have a durable capability lead, that assumption needs revisiting now. Founders choosing a model provider or building AI products should start evaluating Chinese open-weight models seriously, because your competitors almost certainly will.