Bloomberg: China's AI-for-All Offensive Breaks US Containment｜Open Weights Are Air You Can't Tariff
TL;DR
Bloomberg analysis: China's "AI for all" offensive defies US containment playbook — open-weight models make tariff tools useless.
Bloomberg's July 22 long analysis — the rapidly rising global competitiveness of Chinese AI models is ringing alarm bells in Washington, presenting new challenges to the standard protectionist playbook.
US posture is clear — US Trade Representative Greer Tuesday: "We're watching very closely how China is promoting its AI development." Treasury Secretary Bessent separately: the US can sanction any foreign model found stealing US IP, and hinted at pressure on companies using Chinese AI — "You can't use counterfeit goods."
Bloomberg's sharp read — if the Trump administration decides to try to restrict China's open-weight models, the US will face challenges unseen in prior protectionist actions.
Three comparisons build the argument — Chinese EVs can be tariffed (100% tariff landed in May), Huawei 5G can be banned (Rip and Replace cost cumulative $5B), but these measures are much harder to use against software that can be downloaded, modified, and run locally the moment it releases.
The root is distribution cost is zero — a 700B parameter model, download once, replicate infinitely, modify weights = new model, decentralized inference services (Together, Fireworks, Modal) can spin up nodes in Singapore, Dubai, São Paulo, Frankfurt.
US remaining tools — pressure US firms that use Chinese AI (Microsoft evaluating Kimi K3 gets caught), delist Chinese weights from GitHub/HuggingFace (DeepSeek-R1 tried, community mirrors resist), new legislation to expand export controls to "open-source models" (legally hard to define).
China's "AI for all" policy — Kimi K3, Qwen 3.8-Max, Qwen-Image-3.0, Zhipu GLM-5.2, DeepSeek-V4 canary — five frontier models all open-sourced in the past two months. Weights free, fine-tuning open — treating AI as infrastructure, not (as the US does) as product.
Bloomberg quotes a former USTR official — "This is the first time we've faced a 'software Huawei' — and this time there are no base stations to dismantle."
via Bloomberg
US posture is clear — US Trade Representative Greer Tuesday: "We're watching very closely how China is promoting its AI development." Treasury Secretary Bessent separately: the US can sanction any foreign model found stealing US IP, and hinted at pressure on companies using Chinese AI — "You can't use counterfeit goods."
Bloomberg's sharp read — if the Trump administration decides to try to restrict China's open-weight models, the US will face challenges unseen in prior protectionist actions.
Three comparisons build the argument — Chinese EVs can be tariffed (100% tariff landed in May), Huawei 5G can be banned (Rip and Replace cost cumulative $5B), but these measures are much harder to use against software that can be downloaded, modified, and run locally the moment it releases.
The root is distribution cost is zero — a 700B parameter model, download once, replicate infinitely, modify weights = new model, decentralized inference services (Together, Fireworks, Modal) can spin up nodes in Singapore, Dubai, São Paulo, Frankfurt.
US remaining tools — pressure US firms that use Chinese AI (Microsoft evaluating Kimi K3 gets caught), delist Chinese weights from GitHub/HuggingFace (DeepSeek-R1 tried, community mirrors resist), new legislation to expand export controls to "open-source models" (legally hard to define).
China's "AI for all" policy — Kimi K3, Qwen 3.8-Max, Qwen-Image-3.0, Zhipu GLM-5.2, DeepSeek-V4 canary — five frontier models all open-sourced in the past two months. Weights free, fine-tuning open — treating AI as infrastructure, not (as the US does) as product.
Bloomberg quotes a former USTR official — "This is the first time we've faced a 'software Huawei' — and this time there are no base stations to dismantle."
via Bloomberg
