DeepSeek opens its doors to investors at a $44 billion valuation
Rumors have been flying for weeks: $10 billion, $20 billion, even $30 to $40 billion. The whispers about DeepSeek's pre-funding valuation kept shifting, with talk that Tencent and Alibaba were both trying to get a piece. But a source close to the company — an investor at a top-tier firm — tells us none of those numbers were right. DeepSeek's actual pre-money valuation is 300 billion RMB, or roughly $44 billion.
That puts it ahead of publicly traded AI model company Minimax (240 billion RMB as of April 23) and close to Zhipu AI's 380 billion RMB.
Ever since DeepSeek R1 took the world by storm in 2025, everyone has been waiting for the next big leap — R2 and V4. Neither has shipped yet, but that hasn't cooled the market's appetite one bit.
The sheer size of this fundraising round — and the steep requirements it imposes on investors — is rewriting the record books for Chinese tech. According to our source, DeepSeek plans to raise 50 billion RMB in total: 20 billion from internal sources and 30 billion from outside investors. Every participating institution must put in at least 5 billion RMB. Even limited partners face identity checks. One top-tier VC firm was turned away because its fund included too many individual investors who didn't meet the bar.
DeepSeek has never taken outside money before, relying entirely on founder Liang Wenfeng and his affiliated firm, High-Flyer Quant (Huanfang Lianghua). The company's valuation was a black box. This rare public fundraising round is the largest single raise ever by a Chinese AI model company.
Still, it's a different league from US giants. OpenAI just closed a $122 billion round in April; Anthropic raised $20 billion in its Series G back in February.
DeepSeek's pivot from self-funded to seeking external capital is, in some ways, a concession to reality. Building large language models is an all-out war of capital, compute, and top talent.
For a long time, DeepSeek cultivated the image of a pure research lab chasing AGI ideals. Liang and his team stayed deliberately in the shadows. Beyond the occasional technical paper, the outside world could barely glimpse what the star company was up to. The same photo of Liang giving a talk years ago was recycled in every article.
But since the second half of 2025, the landscape has shifted dramatically. The LLM arms race has entered a new phase of capital frenzy. Zhipu AI went public and its market cap surged 7x. Kimi kept raising money, its valuation quadrupling to $18 billion in the three months around Chinese New Year, with talk of a Hong Kong IPO. Even Stepfun (Jieyue Xingchen), the quietest of China's "AI Foundation Model Four Little Dragons" (AI jimo si xiaolong), landed over 5 billion RMB in January. We've learned that three months later, Stepfun is already in a new Pre-IPO round at a $9 billion valuation.
While every other model maker is stockpiling ammunition and jockeying for position, DeepSeek — in the same race — risks losing its first-mover advantage if it doesn't act, even with High-Flyer's financial backing.
More critically, the external competition is destabilizing DeepSeek's internal talent. Liang fostered a free, flexible research culture. But when faced with uncertain stock options versus the multi-million-dollar annual packages (nianbao) offered by big tech, it's hard for team members to stay indifferent.
Publicly reported departures already include Wang Bingxuan, a core author of the LLM (now at Tencent working under Yao Shunyu); Wei Haoran, a key contributor to the OCR series; and Guo Daya, a core author of R1. All have been outed as joining big companies.
And it's not just talent. Compute shortages have been a constant headache for Chinese AI companies. The "Xiaolongxia" (Little Lobster) trend that exploded earlier this year — seen as a paradigm shift — dramatically increased token consumption for long-chain AI agent tasks, pushing inference costs close to training costs and driving compute expenses exponential. Training costs are climbing too: as model parameters move from hundreds of billions to trillions, the compute required for a single training run is skyrocketing. Right now, model generalization is at a critical breakthrough point that demands massive compute — and compute is ultimately a battle of capital.
Beyond fundraising, other signals of change are emerging inside DeepSeek. On April 20, the company posted new job openings for a "Corporate Culture Manager" at its Beijing and Hangzhou offices. The primary responsibility: planning internal events and dail care activities to foster a sense of belonging. The goal, obviously, is to retain employees.
A few days earlier, DeepSeek also listed a "Business Procurement" role, tasked with gathering internal ToC cloud business requirements — essentially bridging internal tech needs with external suppliers to ensure enough compute while controlling costs. It's being read as a sign that this engineering-focused company is starting to think about products and serving end users.
Then there's the new model. We exclusively reported in March that DeepSeek V4 would launch in April. Other media have since gotten more precise word that the release could come as soon as this week.
Whether it's the new model or the fundraising, every move DeepSeek makes is being watched closely. From rejecting capital to opening the floodgates, from open-source dark horse to a $44 billion valuation, each step is rewriting the capital narrative and technical landscape of China's homegrown large language models.