Did DeepSeek’s leak hand the US an AI roadmap?

(Edited and refined by Grace Chong, with the assistance of AI translation.)

A transcript of a closed-door meeting between investors and DeepSeek founder Liang Wenfeng has been leaked, exposing remarks the latter made on the gap between China and the US and the company’s reliance on Nvidia chips. The revelations sparked intense debate online, pushing the hashtag “Liang Wenfeng meeting notes leaked, America gets a free advantage” to the top of Weibo’s trending topics.

After raising about US$7 billion in its first funding round in June, DeepSeek launched a second round of fundraising in July ahead of a potential A-share initial public offering (IPO). However, Bloomberg reported on 25 July, citing people familiar with the matter, that DeepSeek had suspended the fundraising. One reason, according to the report, was Liang’s frustration over the leak of discussions with investors during the first round.

From viral praise to censorship

On 22 July, Chinese technology self-media outlet Elsewhere (别处发生) published a WeChat article titled “A Transcript of Liang Wenfeng’s Four-Hour Meeting with Investors”, featuring a collection of Liang’s memorable remarks from an online discussion held with investors in May.

Many of the quotations attributed to “Saint Liang” (梁圣) — a nickname coined by Chinese internet users — struck an aspirational tone. Among them were: “We generally don’t work overtime”; “Restraint is a core company vision”; “We never started this company thinking about how much money we’d make”; and “We were simply a group of ordinary people who accomplished something extraordinary.”

The article spread rapidly across the Chinese internet. Within a day, it had been shared more than 100,000 times on WeChat alone, becoming a viral sensation.

Feng Ji, founder of Game Science and producer of Black Myth: Wukong, reposted the article on Weibo in the early hours of the morning, saying it was “deeply stirring to read in the middle of the night”. He described DeepSeek as “a Chinese story of greatness through inclusiveness, and strength through freedom from desire”.

Meanwhile, Chinese financial media outlet ChinaVenture (投中) struck a more sceptical tone, commenting: “Liang Wenfeng spoke for four hours, and now everyone’s memorising the quotes. Many of the excerpts being circulated don’t even explain who asked the questions, what the context was, or why Liang gave those answers.”

The full transcript emerges

After the article by Elsewhere went viral, Chinese media published follow-up reports on 23 July with further details.

The National Business Daily, citing an investor in DeepSeek, reported that the leaked transcript was authentic. The meeting lasted three hours and 44 minutes, with Liang spending the first two and a half hours outlining the company’s vision before taking questions for about an hour. During the Q&A, he discussed the computing power gap between China and the US, the stability of DeepSeek’s team and China’s domestic AI chip ecosystem.

At the same time, the full transcript began circulating on Chinese and overseas social media.

According to the transcript, attendees were warned before the Q&A began: “Liang Wenfeng has shared a great deal of sensitive information. Please don’t disclose any figures or details, including the number of AI chips we have. And please don’t make screen recordings or share them outside this meeting.”

The transcript, which Chinese media have verified as authentic, also showed Liang discussing DeepSeek’s computing capacity and dependence on Nvidia chips. “We can buy some ‘non-compliant chips’,” he said. He also compared Huawei’s chips with Nvidia’s, claiming that “four Huawei GPUs equal one Nvidia GPU, while still trailing by two years”.

Not long after the transcript spread online, discussion of Liang’s remarks was rapidly curtailed. Several widely shared WeChat articles, including “A Transcript of Liang Wenfeng’s Four-Hour Meeting with Investors”, were deleted and replaced with notices stating: “This content cannot be viewed because it violates relevant regulations.” Reports by several Chinese state media outlets and major web portals were also taken down.

Fears the leak could aid US scrutiny

Chinese technology commentator Lanxi (阑夕) wrote on X on 26 July that Liang’s remarks had effectively been scrubbed from the Chinese internet.

Lanxi argued that, taken out of context, Liang’s comments on computing power could be interpreted as an admission that US export controls were working. “Chinese AI model developers are genuinely suffering from chip shortages and the growing pains of transitioning to a new hardware ecosystem, and it has had a real impact on commercial returns,” he wrote.

On the same day, the hashtag “Liang Wenfeng meeting notes leaked, America gets a free advantage” briefly appeared on Weibo’s trending list.

Zhao Shengye, a prominent online commentator who describes himself as a “rogue strategist”, argued that the level of operational detail in the transcript effectively handed US authorities a ready-made investigative checklist.

“This was very likely more than an accidental leak,” he wrote. “It looks as though someone deliberately released the information, bringing the matter close to espionage and the disclosure of commercial secrets. We need to reassess the seriousness of this incident from both legal and national security perspectives.”

He added that what appeared to be routine operational details were precisely the chain of custody sought by the US Bureau of Industry and Security, which enforces export controls. The recent US allegations against Chinese AI company Moonshot AI’s Kimi K3 model, he said, were “the warning bell that has already been sounded”.

AI rivalry moves beyond business

Released on 16 July, Kimi K3 drew widespread attention for delivering performance comparable to that of the West’s leading closed-source models. But on 22 July, Michael Kratsios, Director of the White House Office of Science and Technology Policy, publicly accused Moonshot AI of distilling Anthropic’s Fable model during K3’s development.

Kratsios said that large-scale, covert industrial distillation aimed at stealing proprietary US technology and undermining American research is “unacceptable”.

He also alleged that Moonshot AI had purchased servers equipped with Nvidia GB300 chips and was operating them in Thailand, “likely to train its AI models”. The GB300 belongs to Nvidia’s Blackwell series, whose sale to Chinese companies is prohibited by the US.

US Treasury Secretary Scott Bessent later wrote on X: “We support open-source AI and the innovation it unlocks. But open source is not open season on American IP. When PRC firms conduct covert, industrial-scale distillation attacks that cross the line into IP theft, sanctions and Entity List designations will be on the table.”

As Washington intensifies pressure on China’s AI industry, AI is expected to feature prominently when Chinese President Xi Jinping visits the US on 24 September, a trip announced by US President Donald Trump on 23 July.

The same day, Hu Xijin, former editor-in-chief of the state-owned Global Times, urged Liang and Moonshot AI founder Yang Zhilin to avoid travelling to the US or any of the Five Eyes countries.

“For the time being, don’t go to the US, and don’t travel to any of the Five Eyes countries either,” he wrote on Weibo. “I’m worried the US could target them with trumped-up accusations of intellectual property theft and threats to US national security.”

As AI becomes a new front in China-US strategic competition, China’s AI leaders — long accustomed to keeping a low profile — are unlikely to remain above the fray. Their every word and action will now be re-examined through a political lens that reaches far beyond business.

This article was first published in Lianhe Zaobao as “外泄梁文锋“金句”给美国递刀子？”.

Related: Cheap, fast and everywhere: Why China leads AI adoption | Winning the AI race without an OpenAI