Trump officials clash over Chinese AI model as debate mirrors open-source software wars

The conflict over China's Kimi K3 has divided the Trump administration into two camps. Michael Kratsios, Trump's tech adviser, and Treasury Secretary Scott Bessent have accused China of intellectual-property theft and threatened retaliation. Simultaneously, Dean Ball of OpenAI is urging the White House to restrict US companies from using Kimi K3. On the opposite side, venture capitalist Bill Gurley has argued the market should decide, saying "we need to let the free market work." The article reframes this debate by drawing a parallel to how the software industry has evolved. Large language models, the author argues, are ultimately just software — extraordinarily expensive software worth billions of dollars, but software nonetheless. Silicon Valley's history shows that the most successful innovation occurs when software is open and free; companies make money not by charging for the software itself but by building valuable implementations around it. The frontier labs — OpenAI, Anthropic, and others — were never going to maintain a permanent monopoly. It was inevitable that competitors would catch up, and even use the frontier models themselves to generate training data for rival models. This is not theft, the author contends, any more than it was theft when AI labs used the author's book to train their own models. Furthermore, Chinese imitation is unlikely to destroy American frontier labs. By definition, frontier models can maintain their lead even as competitors multiply, especially if those frontier models improve recursively and leap further ahead in capability. The size, rate of improvement, and power of these models could make their creators among the most valuable companies in history. The real value, however, lies not just in the model weights but in everything built on top of them. OpenClaw and other agent harnesses showed that the true magic comes from what people can do with AI, not the model alone. When OpenAI and Anthropic built their own harnesses, demand for tokens exploded — even though in principle any model could plug in. The frontier labs won because they optimized their harnesses for their own models. Going forward, frontier labs will compete by making their products — Claude Cowork, Codex, and others — easier and better to use. They will build guardrails and infrastructure that lock in customers not through secrecy but through superior user experience. Regarding regulation, the author notes that software cannot be effectively stopped at borders, and it is difficult to legally prevent companies from using whatever software they choose. Export controls, while potentially effective at reducing Chinese inference capacity (where most economic gains occur), will not prevent Chinese firms from distilling US models. Restricting access to US software also does not address the cybersecurity and bioterrorism risks from powerful Chinese AI — those risks are prevented by hardening US defenses. The article concludes that if frontier labs cannot build a sustainable business around the most powerful software ever created, they do not deserve to survive. However, the author expects them to do well. Even if they falter, investment in massive models may shrink, but innovation will not end; the greatest minds will shift to other breakthroughs. The final message is straightforward: "It's just software."

The clash between Trump officials and free-market advocates over Kimi K3 repeats a pattern from Silicon Valley's own history with open-source software. The article argues that frontier AI labs operate like software companies, where openness and competition have historically driven innovation rather than killed it. The body's central claim is that locking down AI models was never realistic — competitors would inevitably catch up, just as they do in software. The frontier labs' real competitive moat comes not from keeping weights secret but from building better products on top of those models. OpenAI and Anthropic's experience with harnesses demonstrates this: when they built their own tools on top of their models, token demand exploded, but any model could theoretically plug in. The difference was that the frontier labs tuned their harnesses to their own models, creating a better user experience. This mirrors how software companies like Microsoft or Apple built value not by hoarding open-source code but by creating indispensable applications on top of it. The article's framing also suggests that export controls, while potentially effective at limiting inference capacity in China, cannot realistically prevent the spread of model weights themselves — a software-like asset that travels at the speed of the internet. The real strategic question, then, is not whether to stop Chinese models from existing, but how to harden US defenses against the cybersecurity risks they pose.

No comments yet. Be the first to share your thoughts!