The Case for an American Open-Weight AI Strategy: Compete With China on Infrastructure, Not Bans

The US AI policy debate collapsed into a single question last week: what do you do when Chinese open-weight models start matching your closed frontier systems?

Dean Ball, a former Trump policy advisor who joined OpenAI earlier this month to head its Strategic Futures team, offered one answer in an X post: direct federal agencies to issue “soft law” creating regulatory fear around Chinese models. “You just need to direct every agency to issue soft law that creates FUD,” Ball wrote. “A Federal Reserve Advisory Bulletin found that there may be backdoors in Chinese AI models. It needn’t be that well justified.”

The backlash was immediate. Former Trump AI advisor David Sacks called it regulatory capture for OpenAI that would “backfire by imposing approval gates on U.S. labs while Beijing ignores them.” Undersecretary of Defense for Research and Engineering Emil Michael was less diplomatic: “Every industry/ecosystem has its supreme village idiot. @deanwball is that for AI.” Michael’s substantive point: the Pentagon already restricts certain Chinese labs through legislation, not bureaucratic pressure campaigns.

The Third Path

Andrew Glenn’s Building Our Future essay, published July 20, rejects Ball’s approach but takes the underlying concern seriously. If Chinese open-weight models continue narrowing the gap with closed frontier systems, the US cannot compete by simply frightening enterprises away from superior products. It needs to produce competitive open-weight models of its own.

Glenn proposes a “third path” between two failing strategies. Path one: closed frontier APIs from OpenAI and Anthropic, which meter every token and lock customers into vendor relationships. Path two: unrestricted open releases with no safety framework. The third path: American open-weight model laboratories operating under risk-mitigating frameworks that combine broad weight access with commercial licensing and staged release.

The proposal has teeth because the numbers support urgency. The UK AI Security Institute confirmed this month that open-weight models now trail frontier closed models by 4 to 7 months on cyber capabilities, down from 6 to 10 months through most of 2025. Moonshot AI’s Kimi K3 (2.8 trillion parameters, with weights scheduled for release July 27) claims to outperform Anthropic’s Opus 4.8. Alibaba’s Qwen3.8 preview claims to trail only Anthropic’s Fable 5, though it arrived with no model card, no activated-parameter count, and no benchmark data.

Where the US Stands

The essay acknowledges the US has pieces in place. NSF and NVIDIA committed $152 million to the Allen Institute for AI for open multimodal research infrastructure. OpenAI released gpt-oss models under Apache 2.0. NVIDIA organized a Nemotron Coalition with Mistral, Perplexity, Reflection AI, and Thinking Machines. Thinking Machines released Inkling, its first general-purpose open-weight model, last week.

But Glenn argues these are projects, not an ecosystem. No accountable owner. No predictable release schedule. No clear path to the frontier. The Trump administration’s own AI Action Plan stated “we need to ensure America has leading open models founded on American values” but assigned nobody the responsibility of making it happen.

The economic logic works against open-weight release for commercial labs. A closed API lets you meter tokens, retain customers, monitor usage, and feed data back into training. Release the weights and competing hosts drive down inference prices, customers self-serve, and the lab loses control. From the lab’s perspective, closed is rational. From the country’s perspective, the costs accumulate elsewhere: startups locked into vendor dependency, researchers cut off from the systems they study, government agencies unable to deploy in secure environments, and the global developer ecosystem organizing around Chinese model families.

The Agent Portability Question

For agent builders, the policy question maps directly to infrastructure choice. Agents that call closed APIs are operationally bound to those vendors. An OpenClaw deployment running Claude or GPT cannot switch models without rebuilding tool integrations, prompt engineering, and evaluation pipelines. An agent built on open-weight infrastructure can move between models as capability improves.

Glenn’s proposal goes further: government procurement architecture should require contractors to make agent state, tool definitions, evaluations, memory stores, fine-tuning artifacts, and audit records portable across models. This is the policy version of the modular agent stack argument. Lock-in is not just a business risk for individual companies. It becomes a strategic vulnerability when the locked-in vendor is also the one lobbying for regulations that protect its market position.

The question is whether any of this translates into policy. Ball’s “soft law” approach was the easy path: no legislation required, just regulatory signaling. Glenn’s open-weight ecosystem requires sustained investment, institutional coordination, and a government willing to fund competitors to its own lobbying partners. The AI Action Plan punted on that. Whether anything changes depends on whether policymakers treat the Chinese model surge as a trade problem (ban the imports) or an industrial policy problem (build better domestic alternatives). History suggests they will try banning first.