America’s Open-Model Paradox

Sequoia's latest memo makes an uncomfortable case: the open layer of AI, the free weights everyone builds on, is quietly becoming Chinese infrastructure. According to ATOM's Report, Qwen's share of new open-model fine-tunes and adaptations jumped from 1% in January 2024 to 69% by February 2026. That's not a niche trend. It means the majority of American AI startups have Chinese weights buried somewhere in their stack, whether they advertise it or not.

The more interesting wrinkle is upstream. Western frontier labs are starting to lean on Chinese open models as teachers. Thinking Machines pre-trained its Inkling model independently but used synthetic data from Moonshot's Kimi K2.5 to bootstrap supervised fine-tuning. Nothing illegal there — but doing the same thing with GPT or Claude outputs would violate OpenAI's or Anthropic's terms of service. So a bizarre asymmetry has formed: American labs have a lawful path to learn from Chinese models, but no lawful path to learn from each other.

That asymmetry matters because post-training, not pre-training, is where raw capability turns into something usable — coding, reasoning, tool use, agents. A strong teacher model turns that expensive discovery process into a cheap imitation problem. Distillation doesn't explain all of China's open-model lead; its labs have real compute, real researchers, real hardware-software integration. But distillation is exactly the kind of shortcut that compresses the gap between a decent base model and a near-frontier one, and every new American breakthrough hands Chinese labs another teacher to copy from.

Sequoia also flags a quieter risk: openness isn't the same as permanence. Reuters has reported that Chinese regulators are discussing restrictions on overseas access to advanced models, including unreleased ones. If Beijing decides to stop sharing tomorrow, nothing breaks immediately — Western products just stop improving. And there's a security wrinkle too: open weights don't reveal what data trained them, what got filtered, or whether a dormant backdoor is sitting inside, since research shows implanted behavior can survive fine-tuning and adversarial training. Fine for a chatbot. Not fine for defense or critical infrastructure.

The proposed fix is basically a legal domestic distillation market: let American frontier labs sell structured, metered training rights to vetted US and allied companies, while keeping the riskiest bio and cyber capabilities locked down. Pair that with tougher enforcement against unauthorized foreign extraction. The pitch is straightforward — either the US builds a legal pipeline for turning its own frontier work into cheaper downstream models, or it keeps outsourcing that job to Beijing by accident.