DeepSeek Is Closing a $7.4 Billion Round at $74 Billion — and Pointing at Shanghai

The company that once bragged about doing more with less is now raising one of the largest private financings in Chinese technology history. DeepSeek is closing in on roughly 50 billion yuan — about $7.4 billion — at a pre-money valuation of about 500 billion yuan, or $74 billion, according to reporting from the South China Morning Post and The Wall Street Journal. A full raise at that mark would put the Hangzhou lab at roughly $81 billion post-money, and it would make DeepSeek the most valuable private AI company anywhere outside the United States.

The money is not the headline, though. The destination is. DeepSeek has begun preparing for an initial public offering on Shanghai's Star Market, the Nasdaq-style board China built for its strategic-technology champions. Two people familiar with the plans told the SCMP the company could file as early as the end of 2026, with a public debut targeted for 2027. Neither DeepSeek nor its investors have confirmed any of it.

The round is a striking step up from where the company sat only weeks ago. In June, DeepSeek closed a first outside financing of roughly 50 billion yuan at a valuation above $50 billion, a round that reportedly included Tencent, CATL, JD.com, NetEase and Monolith Management. The new raise re-prices the same company at $74 billion in under three months. Existing backers Monolith, Shixiang Capital and battery giant Contemporary Amperex Technology are participating again, per the SCMP, while new names in discussions include CPE, Legend Capital and Stony Creek Capital, a semiconductor-focused private equity firm. Funds tied to memory-chip designer GigaDevice and state investment vehicles from Hefei are also in the mix — a roster that reads less like a venture syndicate and more like an industrial policy alignment chart.

The efficiency story meets an infrastructure bill

What the capital is for is the most revealing detail in the whole story. DeepSeek is reportedly planning to add roughly 1 gigawatt of computing capacity, alongside spending on model R&D and on retaining researchers in a talent market that has gotten brutally expensive. That is a remarkable turn for a lab whose entire global reputation, forged in the January 2025 shock that knocked hundreds of billions off US chip stocks, rested on the claim that frontier models did not require frontier-scale spending.

The underlying business is genuinely working. Figures first reported by The Information put DeepSeek's revenue for the first seven months of 2026 at 475 million yuan, or about $70.7 million — roughly ten times what the company earned in all of 2025. Net loss narrowed to 715 million yuan over those seven months, against 935 million yuan for the whole of last year. Overall gross margin came in at 44.6%, but the number that will interest bankers is the API business: 82.9% gross margin, a figure neither OpenAI nor Anthropic comes close to disclosing. For context, OpenAI booked $5.7 billion in Q1 revenue at a 39% gross margin; Anthropic reported $11.5 billion in Q2 and guided to 63% gross margin for 2026.

So the picture is a company with world-class unit economics and a rounding-error top line. DeepSeek's models are cheap by design — V4 Pro charges $0.87 per million output tokens, roughly a seventh of what xAI's Grok 4.6 costs — and the company has been quietly raising API prices this year, as have Alibaba, Tencent, Baidu and Zhipu. The era of buying market share in China's model market at negative margins appears to be ending.

Why this matters

Read the round and the listing together and the strategy comes into focus: DeepSeek is converting itself from a hedge fund's research project into a Chinese public company. Founder Liang Wenfeng bankrolled the lab through High-Flyer Quant, the quantitative fund he built before DeepSeek, precisely so he could avoid the venture treadmill. A gigawatt of compute ends that arrangement. Outside capital, state-linked capital, and eventually retail capital on a Shanghai exchange are what a buildout of that size requires.

That transition carries a governance question the market has already started pricing. "DeepSeek's founding team, including Liang, are still traders at heart and inclined to chase maximum upside," Ke Zong, a portfolio manager at a Shanghai-based hedge fund, told CNBC — a polite way of noting that the people running China's most important AI lab are, by training, allocators. High-Flyer affiliates have been picking up pre-IPO allocations in hot Chinese listings including memory maker CXMT and Unitree Robotics, and DeepSeek itself took a 2.31% strategic stake in Unitree's offering with a 36-month lock-up.

Sigrid Wang, a tech analyst at Hutong Research, drew the distinction more sympathetically: "There's a genuine distinction between the quant funds seeking returns and DeepSeek selectively using its corporate balance sheet to build strategic relationships around the future AI stack." Both readings can be true, and public-market investors will eventually have to pick one.

The broader context is a Chinese AI listing wave that is already well underway. Zhipu went public in Hong Kong in January at HK$116.20 and has risen more than tenfold, reaching a market capitalization near HK$489 billion; MiniMax, which listed at the same time, sits around HK$124.2 billion. Both are now planning Star Market listings of their own. DeepSeek arriving at $74 billion is less an outlier than a confirmation that Beijing's capital markets have decided domestic AI labs are strategic assets worth financing at American multiples.

Three things to watch. First, whether the round actually closed — it was reportedly paused in late July after leaked remarks from Liang frustrated some investors, and neither the company nor its backers have confirmed the terms. Second, the filing itself: a Star Market prospectus would be the first audited look inside a Chinese frontier lab, and 44.6% blended gross margins are a very different story than 82.9% API margins. Third, where the gigawatt comes from. In an export-controlled market, the compute plan is the strategy, and the prospectus will have to say something about it.