Recently, DeepSeek successfully concluded the largest financing round ever seen in China's AI sector.

While the financing figures are impressive, the terms are even more remarkable. External investors were granted no voting rights and are bound by a five-year lock-up period. Additionally, Liang Wenfeng insisted on a non-compete clause for DeepSeek's R&D personnel.

Out of the 51 billion yuan raised, Liang Wenfeng personally contributed 20 billion yuan, securing 84.29% of the beneficial shares and retaining 100% of the voting rights. DeepSeek's post-money valuation stands at approximately $50-59 billion (340-400 billion yuan), with Liang's 20 billion yuan investment translating into equity worth nearly 300 billion yuan.

The roster of shareholders in this financing round is equally noteworthy.

It includes CATL, the world's foremost power battery manufacturer, whose energy storage systems are crucial for AI data centers. The denser the computing power, the greater the power demand, making CATL's lithium iron phosphate energy storage cabinets indispensable for DeepSeek's data centers. Tencent, China's leading social and content distribution platform, has the capability to integrate APIs into products used by billions. JD.com and NetEase contribute supply chain data and content scenarios, respectively. IDG and Monolith provide cross-border capital channels, while Zhenxin Capital and Shixiang Technology invest deeply in the industrial chain. Guozhi Investment represents institutional presence.

To put it somewhat hyperbolically, Liang Wenfeng has strategically placed a representative at each critical node of the AI industrial chain—human resources, energy, distribution, data, capital, and policy—each occupying their respective positions. It appears as though Liang Wenfeng is orchestrating the entire setup.

With the exception of Guozhi Investment, other investors accepted the lack of voting rights not because they are in a weak position, but because DeepSeek is so scarce a resource that this scarcity itself constitutes Liang Wenfeng's greatest bargaining chip.

In 1987, Ren Zhengfei registered Huawei in a Shenzhen residential building with a capital of 21,000 yuan. He personally contributed only 3,000 yuan, with the remainder sourced from various channels. At the time, Ren even considered ventures like selling tombstones and weight-loss pills.

The telecommunications equipment industry is highly capital-intensive, necessitating massive ongoing investments in base stations, optical networks, and chip design. In the 1980s, no Chinese bank was willing to lend to a private company without a substantial background.

Thirty-nine years later, Ren Zhengfei holds just 0.52% of Huawei's shares, while 160,000 employees collectively hold 99.48%. This outcome is a result of the industry's capital structure rather than a founder's preference.

In China's tech history, a founder's ability to retain control is primarily determined by the capital intensity of the industry.

In 1999, Jack Ma met with Masayoshi Son in Hangzhou and secured $20 million in funding within six minutes. E-commerce, while asset-light, requires significant cash burn to capture market share. SoftBank's stake increased to 34%, diluting Ma's ownership to less than 7%.

During the 2000 dot-com bubble burst, Tencent's registered QQ users surged to 5 million, but the company had only 10,000 yuan in cash. Pony Ma approached Lei Jun, Zhang Chaoyang, China Telecom, Lenovo, and Zhonggong Network, but none were willing to acquire QQ for 600,000 yuan. Eventually, South Africa's Naspers acquired a 46.5% stake for $32 million. Social products' network effects necessitate user scale, which in turn requires cash burn and equity concessions.

Zhang Yiming retained voting rights but faces an impending threat. With TikTok facing a forced sale in the U.S., he wrote in an internal memo: 'We must find a balance between globalization and trust.'

Liang Wenfeng understands these nuances. AI large models represent the most capital-intensive industry in history, with training a GPT-4-level model costing over $100 million and maintaining global top-tier computing power burning billions annually. Fifty-one billion yuan may not be excessive—it might just be enough to secure a seat at the table.

Following this financing round, DeepSeek has established two systems. The inner system locks in hundreds of R&D personnel, ensuring uninterrupted model iteration roadmaps. The outer system allows each shareholder to contribute their remaining capabilities.

Born in 1985, Liang Wenfeng has surpassed his predecessors. How did he secure such immense control and orchestrate such a sophisticated strategy?

| Mathematical Belief |

In 2015, High-Flyer Quantitative Trading was founded. The name is derived from the Luoshu (magic square). The official introduction states: 'The company hopes to learn the natural and social order from the Luoshu and create a history as profound and wonderful as the magic square.'

The Luoshu is a 3×3 grid. When the nine squares are filled with numbers 1 to 9, the sum of each row, column, and diagonal equals 15.

8 1 6
3 5 7
4 9 2

The number 5 sits at the center. Remove 5, and the entire structure collapses.

Liang Wenfeng chose a pattern with 5 at the center as the company's name, reflecting his mathematical belief and operational logic.

In the laboratory at Zhejiang University's Yuquan Campus, he worked on target tracking for PTZ cameras using Kalman filters and Mean-Shift, rather than neural networks. He believed then that images possess an underlying mathematical structure—something more fundamental that could be excavated with algorithms and verified with data.

After graduation, Liang Wenfeng wrote code for quantitative trading, growing from managing 80,000 yuan to over 100 billion yuan in private assets in less than a decade.

From PTZ target tracking to quantitative trading, he consistently employed the same method: uncovering the mathematical structures beneath the surface, coding them into algorithms, and letting the system run itself.

In the global quantitative investment field, Renaissance Technologies is an unavoidable benchmark. Its founder, James Simons, is Liang Wenfeng's idol.

Simons earned a Ph.D. in mathematics from Berkeley at 23, became chair of Stony Brook University's math department at 30, and collaborated with Shiing-Shen Chern on the Chern-Simons theorem. In 1978, he founded Monemetrics (renamed Renaissance Technologies in 1982). Its Medallion Fund delivered a pre-fee annualized return of 66% from 1988 to 2018—the steepest return curve in human financial history.

But what truly inspires Liang Wenfeng about Simons may not be the 66%.

In 2003, Simons ousted all external investors and retired at the end of 2009. Eleven years after his departure, the Medallion Fund's pre-fee return reached 76%. The company employs 300 people, including 90 Ph.D.s, all sharing the same model with open-source code and a unified compensation pool. Simons set an ironclad rule: never override the computer. Once the system is built, the founder can be removed, like the 5 in the Luoshu—where the sum of each row, column, and diagonal remains 15.

Liang Wenfeng aims to do the same: build DeepSeek into a stable system.

In 2021, High-Flyer built the Firefly No. 2 supercomputer. In 2023, Liang moved 10,000 A100 GPUs into DeepSeek. By 2026, he used 51 billion yuan to pin every node of the industrial chain onto the chessboard. But DeepSeek's system has not yet learned to run itself—his 84.29% ownership represents the unfinished work.

So far, Liang Wenfeng has demonstrated sufficient resolve to optimize the system.

From late 2021 to early 2022, High-Flyer experienced its largest historical drawdown, triggering investor redemptions but maintaining a slight annual profit.

The media questioned: Had High-Flyer's AI model failed? Was the 'fully AI-driven quantitative trading' label just hype?

Liang Wenfeng did not respond or reassure investors. In July 2023, he registered DeepSeek. The 2022 crisis did not shatter his belief—instead, the system silently withstood the fiercest impact. Liang Wenfeng moved to a larger chessboard.

| Strengthening the System |

After the financing, Liang Wenfeng launched his fastest operational round to date.

DeepSeek began aggressively hiring. Cui Tianyi, head of the Harness team, publicly stated on X: 'Severely understaffed. Interviewing daily. Can't fill positions fast enough.'

Cui, a Zhejiang University computer science graduate with six ACM Asia regional gold medals, worked at Jane Street for nine years before joining DeepSeek in March 2026. He has three types of roles perpetually vacant: Agent Harness researchers, engineers, and product managers.

Harness is the 'control layer' connecting large models to intelligent agent products, managing context, tool calls, file read/write, terminal execution, and result self-correction. Everything beyond the model resides here. Unsurprisingly, DeepSeek is developing agents, possibly targeting Claude Code or Codex.

DeepSeek's hiring spree is also defensive. In the past six months, core member Guo Daya joined ByteDance's agent team—not because ByteDance offered a better package, but because agents were not a priority at DeepSeek.

Another core member, Ruan Chong, became chief scientist at Yuanrong Qixing, attracted by multimodal technology's more direct application in autonomous driving than in large model companies.

Luo Fuli joined Xiaomi and led the MiMo team through three product iterations in six months, slashing MiMo-V2.5 API prices by up to 99%, reducing output costs to one-eighth of R1's. Wang Bingxuan, a core author of DeepSeek's first-generation large language model, joined Tencent.

Hiring, retaining talent, building Harness, and establishing computing centers—all these actions aim to transform DeepSeek from a model-releasing lab into a self-sustaining system that generates revenue, iterates products, and retains talent.

With Claude Code, Anthropic's annual recurring revenue (ARR) reached $44 billion by May 2026. The entire Silicon Valley narrative has shifted from 'whose model is strongest' to 'who achieves commercialization first.'

In China, Zhipu's market cap briefly exceeded 1 trillion HKD six months after listing. MiniMax initiated an A-share IPO, while Jieyue Xingchen and Yuezhi Aimian await their turns.

Everyone is betting on a world-class Chinese AI company.

| The '5' at the Center of the Luoshu |

In China's tech history, the last person scrutinized in this manner was Ren Zhengfei. With 0.52% ownership and a 160,000-employee stock ownership plan, he proved a system could outlast an individual.

But Liang Wenfeng should be compared not to the telecommunications industry four decades ago but to fellow AI company founders at the same table.

On November 17, 2023, Sam Altman was ousted by OpenAI's board—the same board he helped establish. OpenAI's governance structure ceded control to a nonprofit board.

Investors had no board seats, and Altman lacked sufficient voting rights to protect himself. Ilya Sutskever, along with three external directors, removed him in one afternoon. Altman regained the CEO position after five days, 700+ employees threatening to quit, and Microsoft's intervention.

Dario Amodei took a different path. He left OpenAI in late 2020, disagreeing with Altman's commercialization pace. Founding Anthropic, he embedded safety into its governance, registering it as a public benefit corporation and establishing a Long-Term Benefit Trust. Five trustees with no financial ties to the company hold special shares, gradually gaining appointment rights for a majority of the board.

Within four years, these trustees will control most of Anthropic's board. Amodei used institutions to prioritize safety—but at the cost of potentially being constrained by them.

Altman was ousted by a structure he designed. Amodei created a structure that could eventually oust him. Liang Wenfeng rejected both paths, using 20 billion yuan, 84.29% ownership, and 100% voting rights to seal all doors that could expel him.

With 84.29% in his grasp, Liang Wenfeng leaves no room for challenges to his control. Yet, hearts and directions depend solely on him. If he ever veers off course, no mechanism can correct him. Burdened with public expectations, what will he do?

In The Hero With a Thousand Faces, the stage called 'Atonement with the Father' warns that the hero's greatest danger after acquiring ultimate power is not external enemies but themselves.

Simons took 26 years to remove himself from the system. Ren Zhengfei spent 39 years, using 0.52% ownership and a rotating CEO system among 160,000 employees to turn Huawei into a machine that no longer needs him.

The Luoshu grid is set. All 'numbers' occupy their rightful positions. Liang Wenfeng, 41, founded DeepSeek just three years ago.

He is the '5' at the center. The test has only begun.

Image sourced from the internet. Rights reserved.