Not for the first time, a Chinese chatbot has set off a firestorm in Washington.
Moonshot AI, a Beijing-based artificial intelligence firm, released its latest model, called Kimi K3, on Friday, claiming that it had “consistently outperform[ed]” every other model on the market except the latest ones from U.S. rivals OpenAI and Anthropic. Chinese tech giant Alibaba followed two days later with similar claims about the prowess of its soon-to-be-released model, Qwen 3.8.
Not for the first time, a Chinese chatbot has set off a firestorm in Washington.
Moonshot AI, a Beijing-based artificial intelligence firm, released its latest model, called Kimi K3, on Friday, claiming that it had “consistently outperform[ed]” every other model on the market except the latest ones from U.S. rivals OpenAI and Anthropic. Chinese tech giant Alibaba followed two days later with similar claims about the prowess of its soon-to-be-released model, Qwen 3.8.
Much like Chinese company DeepSeek did last year, the new models have reignited questions about whether the United States still leads China in the AI race, how close Beijing is to closing the gap, and what Washington can do to maintain the U.S. lead.
The two new models have made a big splash—particularly Kimi, which got so popular that Moonshot had to suspend new subscriptions within 48 hours of launching, citing an overload of the hardware used to power its models (referred to in industry parlance as “compute”). But while having to stop sign-ups because demand is so high may be a great marketing win, analysts say it also betrays the very real compute constraints that Chinese AI developers face.
U.S. companies such as Nvidia dominate the production of the advanced semiconductor chips used to train and power AI models, and the likes of OpenAI, Anthropic, Google, and Microsoft have spent tens of billions of dollars on compute—all while the U.S. government has cut off China’s access to those chips through export controls.
Kimi’s new launch is “a serious development” and can realistically be described as a “second DeepSeek moment,” said Ryan Fedasiuk, a fellow at the American Enterprise Institute in Washington, D.C., who previously served as a China and technology advisor at the U.S. State Department. But it “also illustrated some of the serious structural constraints facing China’s AI industry,” he added.
What’s causing concern within the U.S. government and industry are the ways in which China has overcome those constraints. OpenAI and Anthropic have accused Chinese AI companies of using advanced U.S. models to train their own, effectively piggybacking off the power of the American models while using a fraction of the computing power through a process known as distillation. While distillation is a relatively common practice within the AI industry (companies often use it internally to create more basic, affordable versions of their models), U.S. tech giants and officials allege that Chinese companies have used the technique for “illicit purposes” to steal proprietary technology.
“It’s essentially helping make up for their compute deficit,” said Matt Sheehan, a senior fellow at the Carnegie Endowment for International Peace who studies China’s AI ecosystem. “If they should be 40 percent behind because of compute constraints, distillation makes up half of that,” he added. “This does not mean that Chinese AI companies are not extremely good and extremely innovative—I think if we took away the compute constraints, we would expect the gap to close by a large degree.”
Multiple Trump administration officials on Wednesday accused Moonshot of “stealing” U.S. technology to train its breakthrough model, with White House Office of Science and Technology Policy director Michael Kratsios citing “information” that the Chinese firm distilled Anthropic’s latest Claude Fable model and illegally accessed Nvidia processing units. “[L]arge-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable,” he said.
“Lying, cheating, and stealing is not innovation,” added U.S. Undersecretary of State for Economic Affairs Jacob Helberg. “It’s corrosive to productive economic activity and to the healthy economic development of nations.”
Treasury Secretary Scott Bessent went further, threatening in a post on X to add Chinese AI companies to a U.S. government list that restricts their access to U.S. technologies on national security grounds.
Such statements are fueling speculation that the Trump administration is laying the groundwork for a comprehensive ban on Chinese AI models that could restrict their use by U.S. companies and others who use U.S. technology.
But doing so would be anything but straightforward. It would also face opposition both domestically and internationally because of another key pillar of China’s AI strategy.
Kimi and Qwen’s recent announcements came with a commitment to publicly release all the numerical parameters that make their models tick—known as model weights. That would allow anyone to download the bones of the models for free and build their own applications or features on top of them.
That so-called open-weight approach, also employed by DeepSeek, stands in contrast to the more closed black-box systems of Anthropic and OpenAI’s most advanced models, which can only be used by subscribing to the models themselves and paying to continually use them.
“If it is closed weight, that means that functionally the only way you can access the model is through the company providing it,” Sheehan said. “I’m working entirely through their system … and they can charge me whatever they want,” he added, whereas with open-weight models “you can effectively download the entire model yourself, and at that point it’s totally detached, the company no longer has any control over it and you can use it however you want.”
That also makes Chinese models far cheaper and easier to use for U.S. start-ups—dozens of which petitioned the Trump administration this week not to move forward with a ban. “Am I going to pay these very high subscription fees to send all my requests back to Claude and have it give me the best answers, or would I rather just download a Chinese model that’s mostly as good and is totally fine for my use case, and then the only cost is the cost of running the model?” said Sheehan, explaining the calculus of those smaller companies.
That logic extends to other countries as well, particularly middle powers in the global south that are choosing whether to put their limited resources toward U.S. or Chinese technology. The Trump administration’s recent blocking of Anthropic’s latest models from being used by foreign nationals will have further deepened the global distrust of Washington and has given Chinese President Xi Jinping yet another avenue to present Beijing as the more responsible and reliable global player.
Xi did exactly that at a speech at China’s World Artificial Intelligence Conference in Shanghai on Friday, where he repeatedly advocated for “openness” in global AI development. “We should seize this rare, historic opportunity to encourage open source, openness, collaboration and sharing,” he said, adding that “China is ready to be more open, take more practical actions, and assume a more visionary perspective.” There are already signs that that pitch has found a receptive overseas audience, with 29 countries joining China’s World Artificial Intelligence Cooperation Organization launched ahead of the Shanghai conference.
How long that openness will last within China’s authoritarian system and its tight restrictions on its own internet remains an open question, though. Indeed, even in that same speech, Xi cited a need to “forestall loss of control” over the technology.
Sheehan said China’s embrace of the open-weight approach came more out of necessity from lower compute access. “I don’t think they developed their open-weight strategy because of geopolitics, but it fits very well with it,” he said. “The only way to start getting buy-in from the rest of the world is to offer these models up for free.”
Despite the early hype and promise, the true test for Chinese models will be whether they can keep up their performance levels long term before the compute shortages stymie their wide adoption, Fedasiuk said. “The nature of the competition has shifted,” he said. “It’s not about building a model that scores the best on benchmarks. It’s about serving models to people and getting them into people’s hands, and this is really an area where China does struggle and will continue to struggle.”
But the Chinese companies’ open-weight strategy also means wide enough adoption could negate the compute shortage itself, said Mohammed Soliman, a director at the strategic advisory firm McLarty Associates who focuses on global AI infrastructure. “Because it’s open source, you’re able to push this to data centers, to cloud providers, to users, to run their own versions of the models” and thus proliferate the models through those users’ compute, he said.
It’s a flood-the-zone tactic that China has used effectively for other technologies. “This strategy is not coming out of nowhere; it’s also coming from Chinese industrial strategy around solar, around electric vehicles, around space—you release the model because you want the biggest amount of users to be using it,” Soliman added. “So yeah, they might be constrained today, but they’re not really constrained in the future.”
