The Deal That Broke Three Years of Principle

DeepSeek was not supposed to raise money. Since its founding in July 2023, the Hangzhou company had operated entirely on capital from High-Flyer Quant, the quantitative hedge fund run by its founder Liang Wenfeng, a fund that delivered a 56.55% annualized return in 2025 and manages over 70 billion yuan. Liang had publicly maintained what industry insiders described as a philosophy of "no fundraising, no IPO, no commercialization," a stance that distinguished DeepSeek from every other major Chinese AI lab.

The decision to raise external capital was driven primarily by talent retention pressures. As competitors like Zhipu AI and MiniMax went public or saw valuations surge, DeepSeek employees holding unvested options faced growing incentive gaps. Without a fundraising event, the valuation would not move up, and employee options would not appreciate. The capital raise solved an organizational problem before it solved a financial one.

In April, when funding discussions first circulated publicly, DeepSeek's informal valuation was around $10 billion. By late May, the company had completed its first external round at a post-money valuation of approximately $52 billion, a fivefold increase in roughly six weeks, reflecting both the urgency of the process and the scale of investor appetite for the only major Chinese AI lab still in private hands.

The final figures: 51 billion yuan, approximately $7 billion to $7.4 billion, raised from investors including Tencent, CATL, NetEase, JD.com, IDG Capital, and Monolith Capital, alongside state-backed AI investment funds. DeepSeek has not issued an official statement confirming the complete investor list or final terms, and precise figures vary slightly across sources.

The Governance Structure: One State Fund, All the Votes

The most significant structural element of the deal received less attention than the headline valuation, but carries more long-term consequence.

Most investors placed their capital into a limited partnership managed by Liang rather than acquiring direct equity in DeepSeek itself. Those investors face a five-year lock-up period and do not receive voting rights. China's National Artificial Intelligence Industry Investment Fund, a state vehicle Beijing uses to fund strategic technology sectors, is the only participant reported to hold direct equity with full voting privileges and no lock-up. Liang himself contributed roughly 20 billion yuan, approximately $3 billion, accounting for nearly 40 percent of the total raise.

The dual outcome of that structure is clear. Liang retains operational control: the limited partnership majority means no consortium of outside investors can override his decisions. The state fund retains governance access: its direct equity and voting rights give Beijing a formal seat at the table that no commercial investor has been granted. The structure is not unprecedented in Chinese technology deals, but it is unusually explicit about the hierarchy it establishes.

For Western enterprise customers and governments assessing DeepSeek's open-source models, the structure matters because it determines who ultimately controls decisions about the model's development direction, training data, and alignment parameters. DeepSeek's V3 and R1 model weights are publicly available and can be self-hosted, which reduces data-sharing risk for enterprise users. The governance of the company that develops future versions is a separate and less resolved question.

The Business Model: Cheaper, More Profitable, and Growing Fast

What the funding round revealed for the first time is that DeepSeek has a genuine business behind its research profile.

DeepSeek's annualized revenue is approaching $500 million, according to The Information, the first concrete revenue figure publicly attached to the company. The company has maintained gross margins of 70% to 80% on access to its V4 flagship model despite charging substantially less than leading US competitors. Infrastructure improvements have allowed DeepSeek to process more queries with fewer chips, reducing per-query costs.

The pricing comparison makes the margin figure more striking. DeepSeek's V4-Flash model, launched in public beta on July 31, runs at $0.14 per million input tokens. OpenAI's GPT-5.5 charges approximately $5 per million tokens. Anthropic's Claude Opus 4.7 charges at a comparable premium tier. DeepSeek is charging roughly 35 times less per token than comparable US frontier models, and generating 70-80% gross margins while doing so. That combination has no precedent in the AI API market.

Vercel's data from June 2026 showed DeepSeek processing nearly 23% of all enterprise tokens through its AI gateway platform, compared with Anthropic's 32%. The US-China competitive gap in enterprise deployment, measured by actual usage rather than benchmark scores, has narrowed from essentially zero to a single-digit percentage point gap in the roughly 18 months since DeepSeek's R1 release.

The cost efficiency thesis Liang has championed since R1's January 2025 release, that competitive AI does not require frontier capital intensity, now has revenue evidence behind it. A company approaching $500 million in annualized revenue at 70-80% margins, built on a fraction of the compute budget of its Western counterparts, is generating more than $350 million in annual gross profit. That number supports a credible IPO story in ways that the research-lab-with-a-hedge-fund-backer framing could not.

The Valuation Gap That Defines the Context

At $50 billion post-money, DeepSeek's valuation sits at roughly 7.4% of Anthropic's $965 billion valuation and approximately 5.9% of OpenAI's $852 billion, despite V4 competing directly on key benchmarks against both companies' flagship models. The valuation gap reflects several compounding factors: US capital market depth, the premium assigned to American AI labs' safety governance and enterprise trust, and the structural discount applied to Chinese technology companies by international investors navigating geopolitical risk.

Whether that discount is appropriately calibrated is the analytical question DeepSeek's IPO preparation will force to a public test. A mainland Shanghai STAR Market listing, targeted for 2027 with a filing potentially as early as late 2026, would value DeepSeek in yuan for a Chinese investor base, one less constrained by the national security considerations that limit American institutional participation in Chinese AI equity.

The valuation trajectory since April suggests the discount is compressing fast even without that test. From approximately $10 billion in April to $52 billion post-money in June to a reported $71 billion pre-money target for a second round in July, a sevenfold increase in roughly three months, the market is repricing DeepSeek's value faster than any other AI company in the current cycle, including SpaceX in its IPO week.

What the Capital Will Build: Infrastructure, Chips, and Agents

The original article noted that capital would fund research, talent, and hardware. The subsequent month's announcements provide more specific detail.

DeepSeek has begun constructing gigawatt-scale proprietary data centers, developing in-house AI inference chips to reduce dependence on Nvidia hardware subject to US export controls, and rolling out AI agent products that represent its first significant commercial product extension beyond API access to its base models. The hiring program announced after the round closed targeted at least a doubling of headcount across seven categories, from algorithms and R&D to operations, product management, and corporate functions including HR, legal, and finance.

DeepSeek also invested approximately 140.8 million yuan in robotics company Unitree's Shanghai STAR Market IPO, signing a joint development agreement to combine DeepSeek's model expertise with Unitree's hardware capabilities for humanoid AI applications. The Unitree investment is the first disclosed equity stake DeepSeek has taken outside its own operations, a signal that the company is beginning to build an ecosystem of complementary hardware capabilities alongside its core model business.

The in-house chip development effort is the most strategically consequential capital allocation. US export controls have restricted DeepSeek's access to advanced Nvidia H100 and H200 GPUs. The company's response, demonstrating that competitive model performance is achievable with constrained compute, then investing in domestic chip development to close the hardware gap from the supply side, reflects a dual-track strategy that both works around the current restriction and builds toward independence from it.

The Second Round, Paused

A development that post-dates the article's original publication date but belongs in any complete account of the DeepSeek story: the company moved immediately to raise again.

Less than one month after closing the $7.4 billion round, DeepSeek began discussions for a second round targeting approximately $1.5 billion at a $71 billion pre-money valuation, a 37% increase from the June post-money figure. Bloomberg subsequently reported the company was seeking close to $8 billion at a $74 billion pre-money valuation.

On July 25, DeepSeek suspended the second round days after comments widely attributed to founder Liang Wenfeng about US-China AI competition circulated publicly, generating significant media attention and apparent concern about the fundraising's sensitivity at that moment. Bloomberg reported on August 6 that the company had resumed discussions for close to $8 billion at approximately $74 billion, citing the V4 Flash launch as a catalyst.

The pause-and-resume sequence reveals something about the political sensitivities surrounding the company's capital-raising. The same Liang Wenfeng comments that briefly halted the round were apparently related to statements about the competitive dynamics between American and Chinese AI labs, a topic that, given DeepSeek's unique position as the primary Chinese challenger to US frontier models, carries implications beyond ordinary commercial commentary.

Conclusion: The IPO That Will Define China's AI Market Position

DeepSeek's first funding round is a milestone in Chinese AI financing. Its second round discussions, its V4 model launch, its Unitree investment, and its IPO preparation together constitute a strategic pivot from a research lab operating as a philosophical project into a commercial entity preparing for public accountability.

A mainland listing at 480 billion yuan would be the first major AI-native IPO on a Chinese exchange during this cycle and would dwarf the Zhipu and MiniMax Hong Kong listings that preceded it. It would also provide the first publicly tradable benchmark for how investors price Chinese frontier AI capabilities against American counterparts, a test the current valuation gap, at 7-8% of OpenAI and Anthropic, may not survive unchanged.

Whether the $74 billion second round closes, whether the IPO files in 2026 as targeted, and whether V4's agent capabilities prove competitive with what OpenAI and Anthropic are shipping this year will determine how that repricing resolves. What is already established, as of the June close, is that DeepSeek is no longer a hedge fund's research project. It is a $500 million annualized revenue business with 70-80% gross margins, state-fund governance access, and an IPO on the calendar.

Loading Article...

Loading Article...