DeepSeek halts $1.5B raise after leaked call exposes US compute gap

DeepSeek has verbally suspended a planned second funding round — targeting at least ¥10B ($1.5B) at a roughly ¥480B ($71B) pre-money valuation — after a transcript of founder Liang Wenfeng addressing investors leaked and spread virally. The pause landed just weeks after the company closed a record ~$7B first round backed by Tencent, CATL, and a state-linked AI fund. Notably, the trigger wasn’t diligence or terms; it was the leak itself. Links circulating the meeting notes were scrubbed within hours, and it remains unclear whether that was a company decision or state pressure. The transcript’s authenticity has not been independently confirmed.

The substance is what made it radioactive. Liang reportedly framed compute access as the single variable explaining every gap with US labs — talent, capability, and applications all downstream of it — pegging DeepSeek at roughly two years (elsewhere 12–18 months) behind frontier American firms while running on about one-twentieth of their compute. He detailed a push to cut Nvidia dependence via an in-house compiler stack (TileLang), argued Huawei’s 950 super-node can rival Nvidia’s GB200/GB300 on price and performance, and still conceded a ‘fourfold plus two years’ hardware deficit bottlenecked by manufacturing capacity, alongside an AGI roadmap with dates attached.

The significance is as much political as financial. A Chinese national champion openly cataloguing its reliance on Nvidia and its lag behind the US is exactly the kind of candor that draws scrutiny in Beijing, which plausibly explains both the scramble to delete the posts and the abrupt fundraising freeze. Beyond DeepSeek, the leak is a rare unfiltered look at how much China’s AI ambitions still hinge on compute — and how directly US export controls shape the race.