Kimi K3 is so powerful that it has made OpenAI’s strategic leadership start discussing how the U.S. should fortify its defenses

According to Beating monitoring, OpenAI’s strategy lead Dean W. Ball said Kimi K3 is a very strong model. Its performance in Agent programming is already close to the best publicly available models in the first quarter of 2026. Such capability cannot be explained simply by distillation.

But when he says this, he isn’t just trying to praise Kimi K3.

What Ball really wants to discuss is why China is still willing to open up models at this level, and how this will affect the U.S. AI industry.

In his view, the pressure brought by China’s open-source models isn’t just that there is another cheap competitor. As long as open models are strong enough, developers won’t need to keep paying high prices for closed-source models. Model vendors’ profits will be squeezed, and investors will become more cautious as well. The momentum for U.S. companies to invest tens of billions of dollars in training the next generation of models may weaken accordingly.

Open weights can help technology spread faster, but they may also make it increasingly difficult for training frontier models to make money. In the end, model research and development can only rely on subsidies from other businesses, or on government funding—turning AI into a public infrastructure similar to power grids and roads.

Ball speculates that part of the reason China is willing to do this is that it hasn’t fully appreciated the risks of advanced AI. Another part is that after the U.S. restricts the export of advanced chips, China lacks compute capacity to provide inference services to users worldwide. Since it’s hard to monopolize users through APIs, open weights becomes a way to expand influence instead.

He also predicts that the U.S. government will eventually find a way to prevent Chinese models from entering domestic enterprises. The U.S. doesn’t need to directly ban open-source models; it only needs to continuously raise concerns about backdoors, data security, and compliance risks, and regulated industries such as banks will proactively avoid them.

Ball even believes these warnings don’t need particularly solid evidence. As long as enough uncertainty is created, it can make companies reluctant to adopt the models—while also not forcing all developers to switch to overseas service providers that are harder to regulate.