Chinese AI lab DeepSeek is in talks to raise at least $300 million in its first outside funding round, a deal that would value the company at $10 billion or more, Reuters reported Friday, citing a story from The Information. It's a sharp reversal for a startup that spent years turning down offers from China's biggest VCs and tech firms.
Why now
Up to this point, DeepSeek has lived on money from its parent, the Chinese hedge fund High-Flyer Capital Management. That arrangement worked when R1 could reportedly be trained for around $6 million on export-restricted Nvidia H800s. It works less well now that the company needs to keep pace with labs spending at a completely different scale.
The valuation is the interesting number. $10 billion sounds impressive until you look sideways: OpenAI closed a $40 billion round in March at roughly $300 billion, and Anthropic is reportedly fielding offers near $800 billion. DeepSeek's pricing reflects both what a Chinese AI company can realistically fetch right now and the caution US capital will bring to anything with Beijing in the background.
The talent keeps walking out
Funding is one problem. Keeping the people who built the models is another. Luo Fuli, a core researcher who co-developed DeepSeek-V2, surfaced at Xiaomi last November leading a new team on multimodal spatial intelligence, per a Pandaily report. Xiaomi founder Lei Jun had reportedly been courting her with a "tens of millions of RMB" package for months, the kind of number usually reserved for retention at Alibaba or ByteDance.
More recently, Guo Daya, first author of the DeepSeek-R1 paper that put the company on the map, turned up at ByteDance's Seed team this month. He'd quietly left DeepSeek in March. Baidu was the first rumored destination and publicly denied it. One industry flash pegged his ByteDance move to April 15.
DeepSeek has always run lean, with a team reportedly in the low hundreds. Losing two names from that roster hits harder than the same departures would at a company with 40,000 engineers.
The infrastructure is straining
On the night of March 29, the chatbot went dark for roughly seven hours, with fixes extending into Monday morning. SCMP reporting put the affected user base at more than 355 million, based on February figures. DeepSeek hasn't publicly explained what broke. Chinese analysts blamed surging demand hitting limited compute, which is both the convenient explanation and probably at least partly true.
The company is working on the bottom of the stack. Bloomberg flagged job listings last week for two data center roles in Inner Mongolia, the first time DeepSeek has publicly acknowledged running its own facility there. US officials claim the site hosts smuggled Nvidia Blackwell GPUs. Nvidia calls the claim "farfetched," and DeepSeek denies it.
Founder Liang Wenfeng has resisted outside capital for years on the theory that it would dilute the research culture. V4 is reportedly close to release. Whether DeepSeek closes this round before that launch, and whether the zeros on the valuation climb or stay put, will say something about how the market currently prices a Chinese frontier lab.
