The arrival of the Chinese open-weight model Kimi K3 has intensified debate over AI ownership, use, and regulation. According to Reed, the model’s “blast radius” keeps widening: some U.S. officials — including Michael Kratsios, former tech adviser in the Trump administration, and Scott Bessent of the Treasury — have urged action and accused Chinese actors of theft. At the same time, Dean Ball of OpenAI has been pushing the White House to limit U.S. companies’ use of Kimi K3, while venture capitalist Bill Gurley has argued for letting the free market operate.
The core argument: "AI is software"
Reed suggests that understanding the dispute is easier if you substitute the word "software" for "AI." Large language models — especially those trained using other labs’ data — are ultimately massive pieces of software built with hundreds of millions or billions of dollars. Historically in Silicon Valley, innovation has often worked best when the underlying software was free or open, and commercial returns were realized through implementations, services, and the capabilities those systems enabled.
Reed argues that frontier labs could never permanently lock down models and avoid competition. It was inevitable others would catch up and even use frontier models to generate valuable training data for competing models. He says that is not theft any more than using his own book to help train models would be (a practice he supports).
Ecosystems, agent harnesses, and competitive edges
Tools like OpenClaw and other agent harnesses showed that the value isn't just in the base model but in the surrounding tooling. When OpenAI and Anthropic built their own harnesses, demand for tokens surged. Any model can be integrated into such an ecosystem, but frontier labs held an advantage because they tuned their harnesses to their homegrown models.
Once frontier labs get consumers accustomed to services — Reed mentions offerings comparable to Claude Cowork or Codex — their work is to keep making those services easier to use, to build guardrails and infrastructure around them, and to make them so good that users are unwilling to switch to an almost-as-good, slightly cheaper alternative.
Limits of regulation and security implications
Reed contends regulators face practical limits: you cannot fully stop software at a border, and it is legally difficult to prevent companies from using any software they can access. Export controls alone will not be effective at stopping Chinese firms from distilling U.S. models, even if measures succeed in limiting China’s inference access — and Reed suggests restricting inference access may be a more important objective given where economic gains and innovation are likely to occur.
Moreover, cutting off access to U.S. software does not eliminate the cybersecurity or bioterrorism risks posed by powerful foreign models like Kimi K3. Mitigating those risks requires hardening one’s own defenses rather than relying solely on access restrictions.
Business models, survival, and market adjustment
Ultimately, Reed argues that if frontier labs cannot develop sustainable business models around the most powerful software ever created, they will not deserve to survive. He is optimistic that the frontier labs will manage. Were they to falter, investment in massive models would decline, but innovation would not stop: leading talent would move to other breakthroughs and the market would adapt.
Recent developments
- The Associated Press reported that Kimi K3 suspended new subscriptions temporarily after a surge in demand overwhelmed the model’s capacity.
- The Information reports that Microsoft is working to add Kimi K3 to its Azure cloud services.
Reed’s concluding point: despite the heated debate, the matter can be viewed pragmatically — it is, at base, software. The question is how market dynamics and ecosystems determine winners, and what role regulation can realistically and usefully play.
