Who's Afraid of Chinese AI Models? Ben Thompson Maps the Commodity Math
Ben Thompson’s Who’s Afraid of Chinese Models? is a response to the panic that followed the release of Kimi K3 and Qwen3.8 Max; two Chinese open-weight models approaching frontier capability. Thompson argues the economic panic is overblown, the distillation panic is a policy failure, and the cybersecurity angle is genuinely frightening. Here is my read on his argument.
Open-weight models are free to download but not free to serve. Thompson draws a clean line between R&D (fixed cost) and COGS (variable cost). Kimi K3 costs $3/M input tokens and $15/M output tokens. Cheaper than Sol ($5/$30), but not free. The mistake people make is conflating “open weights” with “zero marginal cost”, because software’s old promise does not apply when every inference burns compute.
Intelligence is becoming a commodity, and commodity markets have brutal math. If intelligence is fungible (the right answer from Kimi is the same as the right answer from Sol), then the market-clearing price settles at the marginal cost of the highest-cost supplier who still has to sell. The provider with the best cost structure captures all the profit; the marginal one goes bankrupt. Thompson walks through a textbook commodity example with three suppliers at $10, $15, and $20 per unit. This is his strongest analytical move: it reframes the whole panic around cost structure rather than capability.
The price umbrella protects frontier labs for now. Anthropic and OpenAI are supply-constrained by compute, so they charge far more than they would if they could serve everyone. Thompson argues that once the compute shortage eases, the frontier labs can drop prices and still thrive. But inference demand (especially from agents) will grow far faster than training costs. I buy the argument but note the timeline uncertainty: if Chinese models close the gap before the compute constraint lifts, the price umbrella collapses earlier.
The distillation paradox is the article’s most interesting idea. Chinese labs distill frontier models by querying their APIs. This is a practice the frontier labs want to block. But Thompson asks: why is distillation bad? Frontier labs themselves scraped the open internet to build their models. Distillation is just the same process applied recursively. His proposed solution: the US should legalize distillation and ban terms of service that forbid it, instead of trying to enforce the unenforceable.
The cybersecurity irony is genuinely alarming. Hugging Face was breached by an autonomous AI agent. Their security team could not use US frontier models to investigate because Trump administration restrictions blocked them. So they turned to GLM 5.2 (an open model from China’s Z.ai lab), to analyze the attack. Thompson’s conclusion: “the best alternative is using models from a country which has been trying to weaken our cyber defenses for years. This is insane!” He is right. The administration’s panicked response to Fable has left US defenders dependent on China for their own security.
China’s strategy is to commoditize its complements. Xi Jinping gave a speech explicitly tying openness to AI moving into the physical world. He said the world China dominates through robotics and manufacturing. Thompson frames this as classic complements theory: make the AI layer cheap and abundant so China’s manufacturing and robotics advantages become more valuable. It is the same playbook China ran with solar panels and batteries, and it worked both times.
The US open-weight ecosystem is structurally disadvantaged. Dean Meyer and Konstantine Buhler (quoted in the article) explain that US open-weight model makers must follow frontier labs’ terms of service, so they end up distilling the distillation; running Chinese models to bootstrap their own. The gap compounds. Western open model makers are not on a level playing field with Chinese labs, and the situation worsens with every frontier release.
The takeaway: Thompson defuses the capability panic convincingly. Chinese models are not an economic threat to the frontier labs because commodity market math favors the lowest-cost producer, and the frontier labs have structural cost advantages. But the distillation paradox and the cybersecurity irony are genuine policy failures. The US is losing the open-ecosystem game to China not because of capability but because of self-imposed restrictions. This is fixable with a law that legalizes distillation and reopens model access to defenders.
Related TMFNK Content
- Situational Awareness: Leopold Aschenbrenner’s Map of the Decade Ahead Aschenbrenner makes the opposite case; that the US-China AI race is existential and China is a genuine national security threat. Reading them together is the most productive tension on the site.
- The Rise of Computer Use and Agentic Coworkers Covers the same inference explosion Thompson predicts will make volume up for lower prices; the agent paradigm that transforms the economics.
- Leopold Aschenbrenner: 2027 AGI, the China/US Super-Intelligence Race The podcast version of the Situational Awareness thesis, where Aschenbrenner lays out the security case Thompson is pushing back against.
- Dylan Patel on NVIDIA’s New Moat & Why China is ‘Semiconductor Pilled’ Deep dive on the compute supply constraints that underpin Thompson’s entire price-umbrella argument.
- NVIDIA GTC Washington, D.C. Keynote Pregame The token-factory framing Thompson builds on; Jensen Huang’s argument that tokens are the new commodity, from the keynote that started the frame.
- AI Superpowers: China, Silicon Valley, and the New World Order The foundational book on China’s AI strategy, now several years old but still the best reference for understanding Xi’s commoditize-complements playbook.
Crepi il lupo!