DeepSeek's Reported US$7–10 Billion Funding Round: Liang Wenfeng Can't Escape the Capital Game
A translated summary of 36Kr's deep dive into DeepSeek's record-breaking (and still unconfirmed) first funding round
Crossing the River bridges the information gap between China's entrepreneurial ecosystem and the West — translating, summarizing, and contextualizing the Chinese-language content that rarely crosses the Pacific.
This post is a translated summary of "DeepSeek融资500亿，梁文锋难逃资本局", originally published on 36Kr by 光锥智能 (Guangzhui Intelligence).
Note: As of publication, this funding round has not officially closed. The details below are based on reporting from multiple credible sources — including 36Kr, Bloomberg, the South China Morning Post, and Caixin — but key terms including valuation, round size, and investor commitments remain subject to change. Notably, Chinese-language sources report a round size of ¥50 billion (~US$6.9B), while Bloomberg cites US$10 billion — the discrepancy likely reflects evolving terms or differing treatment of founder commitments.
Liang Wenfeng came from the quantitative finance world, but he has always been deeply skeptical of outside capital.
Which is why nothing in China's tech world this May has been more shocking than the news that DeepSeek is raising money.
According to multiple sources, DeepSeek is seeking to close a round of US$7–10 billion (Chinese media report ¥50 billion; Bloomberg reports US$10 billion). Founder Liang Wenfeng reportedly plans to personally invest ¥20 billion (US$2.7B) — roughly 40% of the total. Tencent is said to be committing ¥6 billion (US$820M) for approximately 2% equity, with China's National Integrated Circuit Industry Investment Fund (the "Big Fund") in talks to lead.
If completed, this would shatter the record for the largest single funding round by a Chinese AI company, with a post-money valuation of approximately US$45–50 billion.
Just 21 days earlier, in early April, the market valued DeepSeek at US$10 billion, with an initial funding plan of just US$300 million — primarily to establish a price for employee stock options. Three weeks later: valuation up more than 4x, round size up dramatically.
For anyone who knows Liang Wenfeng, this is seismic. This is the man who repeatedly declared "no fundraising," who turned down investment overtures from both Tencent and Alibaba. Why the sudden reversal — and at this scale?
From Quant Trading to "Technical Puritanism"
To understand this fundraise, you first have to understand the man behind it.
Liang Wenfeng's starting point was finance. In 2008, fresh out of Zhejiang University with a master's degree, he led a seven-person team using machine learning models for quantitative trading, achieving 500% returns in three months. By 2015, he had formally established High-Flyer Capital Management (幻方量化), which quickly became one of China's top quant funds.
By 2021, High-Flyer's assets under management exceeded ¥100 billion (US$13.7B). In 2025, the fund averaged 56.6% returns, generating over US$700 million in personal income for Liang. When people say DeepSeek "doesn't need money," it has never been an empty claim.
But Liang's ambitions were never confined to finance.
As early as 2020, he and his team became convinced that AI would be the core driver of the future — and that compute was its foundation. While most quant firms focused on strategy and scale, High-Flyer made an extraordinarily capital-intensive bet: pouring profits into building its own massive GPU clusters. In March 2020, "Firefly One," a supercomputer packed with over a thousand high-end GPUs and costing over ¥100 million (US$13.7M), went online. This was years ahead of the industry, and it laid the groundwork for everything that followed.
In April 2023, High-Flyer announced the creation of DeepSeek, formally entering the AI arena. Liang set three iron rules: no fundraising, no IPO, no commercialization.
In 2023–2024, as capital flooded into Chinese AI — as ByteDance, Alibaba, and Baidu spent billions on models while simultaneously racing to monetize; as Zhipu AI, Baichuan, and Moonshot launched round after round of fundraising — DeepSeek stayed behind closed doors. No roadshows, no press events, no commercial contracts. Just High-Flyer's unlimited funding pipeline and pure research.
Liang's rationale was simple: "DeepSeek's goal is to build a world-class general AI model — not to make money, not to go public. Capital demands short-term returns. Commercialization compromises technical direction. Both interfere with the core mission."
That purity paid off. DeepSeek-V2 launched in May 2024 with an innovative architecture at breakthrough price-performance. V3 was open-sourced in December 2024 with 53 pages of technical detail. R1, released in January 2025, matched OpenAI's o1 on math, code, and reasoning — at a disclosed training cost of just US$294,000, with the base model costing roughly US$6 million. A fraction of what American peers were spending.
DeepSeek had become China's technical benchmark for large language models. Liang's "three iron rules" became an industry legend — proof that you could build world-class AI without outside capital or commercial compromise.
But the undercurrents were already shifting.
Three Pressures That Shattered the Ideal
Liang Wenfeng's insistence on staying outside the capital ecosystem underwent a "strategic adjustment" in April 2026. A man from quantitative finance doesn't make impulsive decisions. His calculations were driven by three converging pressures.
Pressure one: talent.
DeepSeek has always been the most unusual company in Chinese AI — a tiny team with extreme talent density, almost no lateral hires, built primarily from fresh graduates and retained interns. But that model means every core researcher carries outsized weight.
Over the past year, multiple technical leads have been poached. AI researcher Luo Fuli was recruited by Lei Jun to Xiaomi. Researcher Guo Daya joined ByteDance as an agent team lead. Since the second half of 2025, at least five core R&D staff have departed. Every major tech company in China is spending without limit to recruit DeepSeek talent.
The industry consensus: Liang's decision to fundraise was driven first and foremost by the need to price employee options and retain his team.
Pressure two: compute costs.
If two years ago you could compete in LLMs through algorithmic ingenuity — "four ounces deflecting a thousand pounds," as the Chinese saying goes — today it is a raw compute war. V4.1 is scheduled for June 2026, targeting improved reasoning, multimodal capabilities, and stability. Behind that: training clusters scaling from thousands to tens of thousands of cards, spanning both Nvidia and Huawei Ascend hardware. Every step demands astronomical investment.
High-Flyer is wealthy, but not wealthy enough to sustain an arms race against global giants indefinitely. And in AI, standing still means falling behind.
Pressure three: the market itself has changed.
Domestically, ByteDance and Alibaba are pouring enormous capital into AI. Moonshot just closed US$2 billion at a US$20 billion valuation, with ARR exceeding US$200 million as of April. MiniMax and Stepfun are also raising aggressively.
More critically, the IPOs of Zhipu AI and MiniMax — and their subsequent surges on the Hong Kong exchange — have reset valuation benchmarks for the entire sector. Zhipu's market cap exceeded HK$370 billion (US$47B); MiniMax topped HK$210 billion (US$27B). If DeepSeek doesn't establish its price during this window, its primary market valuation risks being instantly overtaken.
Taken together, this explains why Liang decided to raise — and why it had to be now.
But characteristically, his approach to entering the capital game has upended industry norms.
The Control War: Liang Wenfeng's Capital Chess Game
DeepSeek's reported US$7–10 billion round was never simply about "needing money."
This is a meticulously designed capital chess game. Liang's singular objective: bring in outside money while maintaining absolute control over the company's direction.
On April 27, 2026, Chinese corporate registry records show DeepSeek increased its registered capital from ¥10 million to ¥15 million. Liang personally invested ¥5 million, raising his direct stake from 1% to 34%. The original majority shareholder — Ningbo Cheng'en, one of Liang's affiliated entities — was diluted to 66%. No external shareholders were added. Through direct and indirect holdings, Liang controls approximately 84.29% of total equity.
This internal restructuring was completed before any external capital entered. By the time the Big Fund and Tencent sit down at the table, Liang's control position is already locked in.
Two features make this round unlike any other in Chinese AI:
First, Liang personally committing ¥20 billion (US$2.7B) — 40% of the total — making himself the round's largest investor. This "founder mega-lead" model is virtually unprecedented in global AI history. The message is unambiguous: I will take your money, but I will not accept your control.
Second, DeepSeek is reportedly accepting only two categories of capital: state-backed strategic funds and industrial capital. Traditional financial VCs — Sequoia, Hillhouse, and the like — have been shut out entirely. Alibaba reportedly failed to reach agreement on terms.
The logic: traditional VCs seek 3–5 year exits, which would force rapid commercialization and an IPO timeline incompatible with DeepSeek's long-term research orientation. State and industrial capital prioritize strategic value over short-term returns, won't interfere with technical direction, and can provide policy access, compute resources, and deployment scenarios.
Liang's personal ¥20 billion (US$2.7B) buys him the right to dominate the conversation at this valuation. If he had merely followed with a token investment, outside investors would hold the pricing power and strategic leverage. Instead, his commitment sends a clear signal: my conviction in this company exceeds that of any external investor. You may participate, but on my terms.
It is an extraordinarily sophisticated founder strategy.
The Price of Entering the Arena
DeepSeek's reported valuation of US$45–50 billion would make it the second-highest-valued LLM company in China, behind only Zhipu AI (US$48B). It surpasses MiniMax (US$28B) and Moonshot/Kimi (US$20B). A 4–5x valuation increase in 21 days is virtually without precedent in venture history.
But the global gap remains stark. OpenAI's valuation of US$852 billion is roughly 17x DeepSeek's. Anthropic's target of US$900 billion is 18x. DeepSeek's valuation sits at approximately 6% of either American leader. This reflects two realities: capital is concentrating into a handful of global leaders, and American companies benefit from global capital flows while Chinese companies rely more heavily on domestic and strategic capital.
Still, the strategic dimension matters. With the Big Fund's involvement, DeepSeek is positioned to play a central role in China's domestic AI chip ecosystem and model self-sufficiency — a differentiation that pure commercial valuation doesn't capture.
Risks remain. DeepSeek's commercial model is relatively thin, primarily API-based. Its open-source strategy has built a massive developer ecosystem — V4-Flash cache hits are priced at ¥0.02/million tokens (US$0.003), essentially the global floor — but many enterprise clients simply self-deploy, diverting paid API revenue. Whether a US$45–50 billion valuation can be supported by corresponding revenue remains the biggest open question.
There is also the question of whether Liang's absolute control can survive the introduction of powerful new shareholders. And talent pressure may not have disappeared just because options now have a price. Meta has reportedly offered packages of US$200–300 million over four years to poach top AI researchers — contracts that exceed what the world's highest-paid athletes earn.
In an era where AI talent can be precisely priced, stock options are only part of the retention equation. Research culture, intellectual freedom, and growth opportunity matter just as much. Whether DeepSeek can maintain its distinctive purity after entering the capital game is the deeper question.
What Comes Next
DeepSeek V4.1 is slated for June — less than two months after V4's April 24 release. It will reportedly feature native multimodal input (images, audio), enhanced MCP protocol support for enterprise tools, and continued cost reductions on the existing hybrid attention architecture, with million-token context as standard.
That pace of iteration demands sustained capital. And in this particular capital game, there are no winners who take all — only those who adapt and survive.
From refusing all fundraising to personally committing US$2.7 billion, Liang Wenfeng's transformation appears sudden but was in fact inevitable. The AI competition in 2026 has evolved from a contest of model technology into a total war across compute, talent, product, and ecosystem. In the face of that war, any attempt to build a "technical utopia" in isolation leads, eventually, to a walled city.
This article originally appeared on 36Kr (36氪), originally published by 光锥智能 (Guangzhui Intelligence) via WeChat. Translated and adapted for a Western audience as part of the Crossing the River project — bridging the information gap between China's entrepreneurial ecosystem and the West.
Original: https://www.36kr.com/p/3806156817243912