DeepSeek's $7.4 Billion Question: The Deal That Changed China's AI Map

*The Ulanqab data center in Inner Mongolia, where DeepSeek trains its models on clusters of Huawei Ascend chips — a physical manifestation of China's AI sovereignty strategy*

The Deal Nobody Saw Coming

On the morning of June 18, 2026, Liang Wenfeng walked into a conference room in Hangzhou's Binjiang District and signed a document that would reshape the global AI landscape. The document was a term sheet — the final one in a six-month negotiation that had drawn in every major capital pool in China and attracted the attention of regulators on both sides of the Pacific.

The numbers were staggering: $7.4 billion in fresh capital, at a pre-money valuation of $50 billion. If DeepSeek goes public, it will be the largest Chinese tech IPO since Alibaba's 2014 debut, and the largest pure-play AI company listing anywhere in the world. The round brought in a consortium of investors that reads like a directory of Chinese institutional power: the National AI Industry Investment Fund, several provincial SOE platforms, and a who's-who of cross-industry strategic investors including CATL and Geely.

But the headline number is not the most important part of this story. What matters is the deal structure — a complex arrangement that gives the Chinese state voting rights over DeepSeek's strategic direction while preserving Liang Wenfeng's operational control through a layered limited partnership architecture. Analysts are calling it the "VIE+" model, and if it survives regulatory scrutiny, it could become the template for every major Chinese AI company going forward.

Table 1: DeepSeek Funding Round Overview

*Sources: 36Kr (June 2026); Sina Finance; company disclosures via registration filings.*

From Quant Trading to AI Sovereignty: The DeepSeek Origin Story

To understand why this deal matters, you have to understand who Liang Wenfeng is and how DeepSeek was built. Unlike Sam Altman, who came from Y Combinator's venture capital world, or Demis Hassabis, who emerged from academic AI research, Liang Wenfeng is a quant trader. He founded Phantom Square Quantitative in 2015, a Hangzhou-based hedge fund that applied machine learning to Chinese equity markets. By 2023, Phantom Square was managing billions of RMB and generating returns that consistently outperformed the market.

Liang's interest in AI was initially practical. Better language models meant better sentiment analysis of Chinese financial news. Better vision models meant better pattern recognition in market data. But sometime in 2021, he realized that the models themselves were becoming more valuable than the trading strategies they enabled. He began redirecting Phantom Square's profits — and its GPU clusters — toward fundamental AI research. By 2023, DeepSeek existed as a side project within Phantom Square. By 2024, it was the main event.

What made DeepSeek different from China's other AI labs was its culture. While Baidu, Alibaba, and ByteDance were racing to build consumer-facing chatbots and competing for mobile app download rankings, DeepSeek focused on research. The company published papers. It released open-source models — DeepSeek-V2, DeepSeek-V3, and the breakthrough DeepSeek-V4 — that consistently ranked near the top of global benchmarks. It built a reputation as the "Chinese OpenAI that actually ships code."

The 36Kr analysis from late 2025 captured this persona perfectly: *"DeepSeek's persona was like a hidden sect in a martial arts novel — Phantom Square Quantitative behind it, Liang Wenfeng not lacking money, researchers buried in models, product and commercialization not urgent. Other startups outside were drumming up financing, listing, building applications, developing ecosystems, while it was like a silent compute monk, sitting in meditation, deriving formulas, training models."*

But the AI industry does not respect monks forever — especially when the monk holds the true scriptures. By early 2026, DeepSeek's models were being used by millions of developers, its API was processing billions of tokens daily, and the company's revenue had crossed $1 billion with profit margins above 50%. The monk had become a kingdom. And kingdoms need walls.

The "VIE+" Structure: How State Capital Meets Founder Control

The most significant aspect of the $7.4 billion round is not the size but the architecture. China's traditional Variable Interest Entity (VIE) structure — used by Alibaba, Tencent, and virtually every Chinese tech company to list abroad while maintaining domestic control — has been under regulatory pressure since 2021. The government wants to ensure that strategically important companies remain accountable to state interests. But outright state ownership kills the entrepreneurial incentive that made these companies successful in the first place.

DeepSeek's solution, according to multiple sources familiar with the deal, is a layered structure that analysts are calling "VIE+" or "Dual-Class LP." Here's how it works:

Layer 1: The Operating Company. DeepSeek's core AI research and operations remain controlled by Liang Wenfeng and the founding team through a domestic Chinese entity. This layer holds the models, the data, the talent, and the training infrastructure.

Layer 2: The Limited Partnership. The $7.4 billion investment flows into a limited partnership structure where the state-affiliated investors (National AI Fund, provincial SOEs) hold limited partner interests. As LPs, they have economic rights — returns on investment — but no direct operational control.

Layer 3: The Voting Rights Vehicle. Here's where the innovation lies. The LPs have negotiated a separate voting rights agreement that gives them veto power over a specific list of strategic decisions: foreign expansion plans, data-sharing agreements with non-Chinese entities, major M&A transactions, and any change in control that would result in foreign ownership above 50%. In exchange, Liang Wenfeng retains exclusive control over all operational decisions — hiring, research direction, product strategy, pricing.

Layer 4: The Public Listing Vehicle. When DeepSeek eventually IPOs, the listing entity will likely be a Hong Kong or offshore structure that draws revenue from the domestic operating company through contractual arrangements — similar to the traditional VIE, but with the additional LP layer providing state oversight.

Table 2: DeepSeek "VIE+" Governance Structure

*Sources: 36Kr analysis (June 2026); Sina Finance; structured from registration filings and investor disclosures.*

This structure is not just a financial innovation. It is a political innovation. It attempts to solve the central tension of China's tech policy in the 2020s: how to attract the massive capital required for frontier AI development while maintaining state oversight over technologies that Beijing considers strategic. The traditional answer was state-owned enterprises, but SOEs have proven incapable of the rapid iteration and risk-taking that AI requires. The "VIE+" model tries to combine the capital and stability of state backing with the agility and incentive alignment of private entrepreneurship.

Whether it works is an open question. The structure has not been tested in a crisis. What happens if Liang Wenfeng wants to open a research lab in Silicon Valley and the state LPs veto it? What happens if DeepSeek needs to license its technology to a foreign partner to access export-controlled chips, and the state views that as a security risk? The VIE+ model provides a framework for these conversations, but it does not resolve the underlying tension.

The Revenue Machine: Why DeepSeek Could Afford to Say No

One of the most striking aspects of the $7.4 billion round is that DeepSeek did not need the money. Unlike most startups that raise massive rounds because they are burning cash, DeepSeek was already profitable — and not just barely profitable, but generating over $500 million in annual net income on more than $1 billion in revenue.

Table 3: DeepSeek Estimated Financial Profile (2025–2026)

*Sources: 36Kr financial analysis; Sina Finance estimates; company disclosures.*

The profitability comes from two sources. First, DeepSeek's technical efficiency. The company's Mixture-of-Experts (MoE) architecture and Multi-Head Latent Attention (MLA) reduce training costs by roughly 90% compared to dense models of similar capability. A model that costs $50 million to train at OpenAI or Anthropic costs DeepSeek $5 million. This is not just a cost advantage — it is a strategic advantage that allows DeepSeek to iterate faster, experiment more broadly, and undercut competitors on API pricing.

Second, the Chinese market's scale. DeepSeek's API is the default choice for Chinese developers and enterprises, not because it is the best model in the world (though it is competitive), but because it is the best model that can be deployed without regulatory friction inside China. For a Chinese bank, a car company, or a local government building a smart city system, using DeepSeek is not just a technical decision. It is a compliance decision.

So why raise $7.4 billion if the business is already printing money? The answer is infrastructure. DeepSeek's models are growing exponentially in size and capability. The next generation — what the company is calling "DeepSeek-V5" internally — is rumored to require clusters of 100,000+ GPUs (or Huawei Ascend equivalents) running for months. That level of compute requires data centers, power plants, and cooling systems that cost billions to build. The $7.4 billion is not operating capital. It is a war chest for the next three years of infrastructure arms races.

The National AI Fund: China's New Capital Weapon

The lead investor in DeepSeek's round is the National AI Industry Investment Fund, established in 2024 with an initial corpus of 100 billion RMB (~$14 billion). This fund is not a traditional venture capital vehicle. It is a state-directed strategic investment platform that explicitly targets "AI companies of national significance" — a category that includes not just the obvious candidates like DeepSeek, but also chip designers, robotics companies, and data center operators.

The fund's investment in DeepSeek is its largest single deployment to date, and it signals a shift in how the Chinese state participates in technology financing. In the 2010s, state involvement in tech was largely indirect — through policy support, land grants, and tax incentives. In the 2020s, it is becoming direct equity investment with governance rights attached.

This is a significant departure from the American model, where frontier AI is funded primarily by private venture capital (Andreessen Horowitz, Sequoia, Thrive Capital) and corporate balance sheets (Microsoft, Amazon, Google). The US government has the CHIPS Act and the National Science Foundation, but it does not have a $14 billion fund that takes board seats in AI companies. China's approach is closer to the industrial policy models of South Korea and Taiwan in the 1980s and 1990s — direct state equity in strategic sectors, with the expectation of both financial returns and national competitiveness.

Table 4: Comparison of AI Funding Models — China vs. United States

*Sources: 36Kr; Sina Finance; Financial Times; comparative analysis.*

The trade-off is clear. China's model provides more capital, more infrastructure certainty, and more regulatory alignment for companies operating in the Chinese market. But it also introduces political risk that private capital does not face. If US-China relations deteriorate further, the National AI Fund could find its DeepSeek stake subject to sanctions, frozen, or otherwise impaired. The VIE+ structure does not protect against geopolitical risk — it arguably concentrates it.

Beyond the state funds, DeepSeek's round included strategic investors from industries not typically associated with AI venture capital. CATL, the world's largest battery manufacturer, and Geely, one of China's largest automakers, both invested. These are not financial investors seeking portfolio returns. They are strategic investors seeking AI integration — CATL for battery manufacturing optimization, Geely for in-car AI assistants. This pattern of strategic industrial capital investing in AI for competitive advantage is becoming common in China. AI is not just a software layer. It is becoming a physical infrastructure layer that touches batteries, cars, robots, and factories.

What This Means for the Global AI Landscape

DeepSeek's $7.4 billion round is not just a Chinese story. It has implications for the global AI competition in four dimensions.

First, it accelerates the bifurcation of the global AI ecosystem. DeepSeek's models are already competitive with OpenAI's and Anthropic's on most benchmarks. With $7.4 billion in fresh capital, that gap is likely to narrow further. But DeepSeek's models are increasingly optimized for the Chinese market — Chinese language, regulatory requirements, and enterprise use cases. The result is not a single global AI market but two parallel ecosystems.

Second, it raises the cost of entry for everyone else. The $7.4 billion round sets a new benchmark for what it costs to compete at the frontier. Startups in Europe, India, and the Middle East hoping to build "national champions" with a few hundred million dollars now face a reality where the leading players have war chests measured in billions. The AI industry is consolidating into a few mega-players, and DeepSeek just joined the club.

Third, it complicates the US-China technology relationship. Washington may view the National AI Fund's investment as a form of state subsidy that distorts competition, potentially leading to trade restrictions on DeepSeek's API services or foreign subsidiaries. The VIE+ structure, designed to navigate Chinese regulations, may create new vulnerabilities under American regulations.

Fourth, it validates the Chinese approach to AI development. For years, Western analysts dismissed China's AI industry as imitative — good at copying American models, bad at fundamental innovation. DeepSeek's technical achievements (the MLA architecture, the efficient MoE training, the competitive open-source releases) have already challenged that narrative. The $7.4 billion round validates a different model of innovation: one that combines entrepreneurial agility with state-directed infrastructure investment. Whether it produces better AI than the American VC-driven approach is an open question. But it is now clearly a viable competitor.

The IPO Question: When and Where?

Every major funding round raises the question of exit. DeepSeek's $50 billion valuation is a private market number. The public market will have its own view. If DeepSeek IPOs, it will likely choose Hong Kong or Shanghai's STAR Market — the "A+H" model that has become standard for major Chinese tech companies. Hong Kong provides access to international capital; STAR Market offers political safety.

The timing is less clear. Liang Wenfeng has historically been in no rush to go public. Phantom Square operated for a decade as a private company. But the pressure from state investors — who have mandates to deliver returns within specific timeframes — may accelerate the timeline. Most analysts expect a filing in 2027 or 2028.

Social Voices

知乎 (Zhihu)

"DeepSeek融资74亿美元，梁文锋还是实控人，但国家基金有投票权。这个结构很聪明——既解决了国家要'管'的问题，又保留了创始人要'干'的动力。问题是，当'要管'和'要干'冲突的时候，谁赢？"

*"DeepSeek raised $7.4 billion, Liang Wenfeng is still the controlling shareholder, but the state fund has voting rights. This structure is clever — it solves the state's need to 'control' while preserving the founder's incentive to 'build.' The question is: when 'control' and 'build' conflict, who wins?"*

微博 (Weibo)

"50亿美元估值，DeepSeek已经超过百度市值了。一个成立不到五年的公司，靠卖API就能比二十年互联网巨头值钱，这说明什么？说明AI时代，模型才是真正的基础设施。"

*"$50 billion valuation — DeepSeek is already worth more than Baidu. A company less than five years old, making money selling APIs, is more valuable than a twenty-year internet giant. What does this tell us? In the AI era, models are the real infrastructure."*

Twitter/X

"DeepSeek's $7.4B round at $50B valuation with state voting rights is the most important tech deal of 2026 that nobody in the West is talking about. While VCs debate whether AI is in a bubble, China just built a national champion with a war chest bigger than Anthropic's entire raise history."

小红书 (Xiaohongshu)

"刚用完DeepSeek写周报，然后看到它融了74亿美元的消息……我用的API是不是很快就要涨价了？😂"

*"Just finished my weekly report using DeepSeek, then saw the news about its $7.4 billion funding round... Is my API about to get more expensive? 😂"*

Hacker News

"The VIE+ structure is fascinating from a corporate governance perspective. It's essentially trying to create a 'golden share' arrangement without the poison pill. The question is whether international investors will accept it. If DeepSeek lists in Hong Kong and the structure holds, every other Chinese AI company will copy it."

豆瓣 (Douban)

"梁文锋的做法很'中国'——不争先，争的是滔滔不绝。当所有人都在做APP、做流量、做商业化的时候，他在做模型。等模型做好了，钱自然就来了。这不是运气，是战略定力。"

*"Liang Wenfeng's approach is very 'Chinese' — not competing for first place, but for sustained flow. While everyone else was building apps, chasing traffic, and commercializing, he was building models. Once the models were ready, the money came naturally. This is not luck. It is strategic patience."*

Related Articles:

- China's Humanoid Robot Industry Crosses the 100,000-Unit Threshold

- ByteDance's $200 Billion AI Infrastructure Gambit

- The AI Compute Crunch: China's Token Crisis and the Search for Efficiency

- Huawei's Pangu Ultra: 718 Billion Parameters of Chinese Silicon