China's Open-Weights Strategy

1. Commoditize the model layer to break US capital efficiency

US labs need to justify $100M training runs to VCs who expect proprietary returns. If Chinese labs release frontier-grade open weights for free, the pricing power of US APIs collapses. We're already seeing it: Chinese models now handle nearly 60% of enterprise API traffic on OpenRouter. When the same capability costs $5M instead of $100M, private markets repriced the whole sector downward. R&D budgets compress. US labs slow down. The VCs don't have a thesis anymore.

China's labs don't need to be profitable. DeepSeek, Moonshot, Zhipu, MiniMax: they're all state-adjacent or state-backed entities operating inside an industrial policy framework. They can absorb losses that Sequoia and Andreessen Horowitz will not. This is asymmetric warfare through pricing.

2. The nationalization asymmetry

This is Haseeb's sharpest point, and it's the one most people miss. If private investment pancakes in the US, what's the fallback? The US can't nationalize its AI labs. The constitutional, political, and institutional barriers are enormous. The Cairo Review documented the Anthropic confrontation in early 2026: even that limited government intervention generated massive backlash. There is no clean mechanism for Washington to say "OpenAI is now a national asset."

China can do it with a phone call. The labs are already entangled with the state. If market signals fail, Beijing just continues funding directly. The structural difference is: China has a floor under its AI investment that the US does not. Open weights are how you trigger the condition where that floor matters.

If China forces US labs to cut R&D budgets because private markets get spooked when token prices compress, and then China nationalizes their frontier labs straight away and we don't, why wouldn't that mean they can develop a lead? Haseeb Qureshi, @hosseeb

3. Standards capture via the Global South

Same playbook as Belt and Road, applied to software instead of ports. Xi's WAIC speech positioned China as the champion of "AI for all," pledging 5,000 training slots, cooperation centers, and technology transfer to BRICS, ASEAN, Africa, and Latin America. 29 countries signed onto the China-led World AI Cooperation Organization. Microsoft's own data shows 60% of Global South AI traffic runs on Chinese open-source models.

Once your government builds its AI infrastructure on Qwen, you don't switch. The switching cost is retraining your engineers, reconfiguring your systems, and rebuilding institutional competence. China is giving away the razor to sell blades, except the blades are dependency, data flows, and standards alignment. The Ugandan government's AI system runs on Alibaba's Qwen-3. That's not charity. That's infrastructure capture.

4. The hardware workaround

China can't get the best chips. US export controls on advanced NVIDIA GPUs are real and binding. But open-weight models that are efficient to run (smaller, optimized, quantized) work around this constraint. If you can't win on compute scale, win on compute efficiency and distribution density.

The USCC's March 2026 "Two Loops" report laid this out: China's open AI strategy and its manufacturing dominance are mutually reinforcing. Open models get deployed across China's industrial base: factories, logistics, robotics, generating real-world data that feeds back into model improvement. The deployment gap in embodied AI (humanoids, autonomous driving) may actually favor China because they're already integrating AI into physical systems at scale. The two loops are: open models enable broad deployment, broad deployment generates data, data improves models, improved models get released open, repeat.

5. Distillation as a free R&D pipeline

OpenAI accused DeepSeek of distilling US models in February 2026. But the flow goes both ways. When China releases open weights, every developer globally fine-tunes, extends, and stress-tests them. Bug reports, fine-tuning recipes, and architectural improvements come back as free R&D. It's crowdsourced model improvement at planetary scale, and the state captures the upside.

This is the real 4D chess element. Open weights are not just a diplomatic tool or a pricing weapon. They're a research multiplier. The global developer community becomes an extension of China's AI labs, and none of those developers are on China's payroll.

6. Soft power and narrative control

Xi framed Chinese open-source AI as a "global public good" and warned against "new historical injustices" from unequal access. This is diplomatic cover for commercial imperialism. The Global South hears "we'll help you build AI capacity" and doesn't hear "and you'll be running on our stack forever, with our values embedded in the weights."

The narrative is potent because it contains a real grievance. US AI policy has been exclusionary: export controls, API restrictions, security clearances. China is offering inclusion. The framing isn't dishonest; it's just strategic. You frame your self-interest as public goods provision, and you let the soft power do the standards-setting work that hard power can't.

7. The ASI optionality play

Nic Carter dismissed ASI as "a religious belief, not a markets belief." That's a category error. You don't need ASI to be real for the strategy to work. You need the US to believe it might be real and to allocate accordingly. If there's even a 10% chance that sustained compute advantage produces a decisive capability gap, and China has compressed US investment while sustaining its own via state backing, they've bought a lottery ticket with someone else's money.

The optionality is asymmetric in another way. If ASI doesn't materialize, China still owns the global open-source AI ecosystem, the standards, the developer mindshare, and the Global South dependency network. The downside is bounded. The upside is not. That's a rational bet for a state actor with a 50-year time horizon.

What could go wrong

Reciprocal commoditization. Nic Carter's strongest point: if China open-weights, the US can use those weights too. There's no inherent asymmetry in distillation. If GLM-5.2 is open, US companies can build on it. The attack works both ways.

The US nationalizes anyway. If the strategic threat becomes obvious enough, political resistance to nationalization might collapse. The US has done it before (steel, railroads, banks in wartime). The question is whether the threat is recognized fast enough.

Open weights cut both ways. China's own frontier capabilities become transparent. Others can distill from Chinese models. The openness that creates the R&D multiplier also creates leakage.

The Global South pushes back. Dependency on Chinese AI could trigger the same backlash that Belt and Road faced when countries couldn't repay loans. Sovereignty concerns cut both directions.