There’s been an interesting debate on X these past few days about Chinese open-source models, especially Kimi: When a highly capable, low-cost model makes its weights public, allowing anyone to deploy and use it, is it accelerating AI progress, or is it a form of “dumping” that could undermine the motivation of U.S. AI companies to innovate? This perspective was first raised by @deanwball, who recently joined OpenAI. He argues that training cutting-edge models requires investments of tens or even hundreds of billions of dollars. But once the model weights are made public, other cloud providers and developers can use them to offer services, making it hard for the original company that bore the training costs to recoup its investment. The potential outcome? In the short term, users get cheaper access to AI. But in the long term, U.S. labs might see their revenues, valuations, and funding capabilities decline, reducing their willingness to invest in training the next generation of models. Dean likens this scenario to “dumping”—capturing market share by offering products below cost, ultimately weakening competitors’ ability to continue investing. a16z partner @martin_casado strongly disagrees, arguing that the idea that “open-sourcing weights will slow down AI” doesn’t make sense. Open-sourcing typically leads to more innovators, more competition, and more applications, not less progress. Dragonfly managing partner @hosseeb, however, believes that Dean isn’t simply opposing open source but is raising a more specific question: When model training becomes extremely capital-intensive, but deployment can be easily replicated by anyone, who will fund the next generation of cutting-edge models? Does the short-term accessibility brought by open source risk undermining the commercial foundation for advanced model development in the future? Is this just OpenAI’s anxiety about China catching up in the AI race, or is it an underappreciated, legitimate concern? What do you think? Share your thoughts