The biggest news in the Chinese AI circle this weekend is that DeepSeek is rumored to be releasing about 3% of its equity financing at a valuation of US$10 billion. For a company that has long insisted on "self-supplying blood" and whose founder Liang Wenfeng himself directly and indirectly holds 84.29% of the shares and has almost 100% of the voting rights, this news itself is enough to ignite industry discussions.
However, it is worth noting that the news has only been brewing for two days, and the information returned by various channels is highly consistent: a person from a large state-owned equity institution said that the news is "probably true", but "it is completely uninvestable at the moment"; many venture capitalists also admitted that for popular projects such as DeepSeek, the financing share usually needs to be "grabbed". In other words, even if the financing news is true, there is a high probability that there will be very few external institutions that can actually get a share.
Focusing on this rumored incident, we put forward four levels of logical judgment, which are broken down layer by layer as follows.
01
The first level of logic: The essence is the structural design of equity incentives for unlisted companies
The particularity of DeepSeek is that it has never received any external equity financing since it was incubated by Magic Square Quantitative in 2023. This fact brings about a structural problem that is easily overlooked: options in the hands of employees lack a market-based pricing anchor.
An investor who has invested in large models hit the nail on the head in his analysis of China Business News: Even if DeepSeek is open to financing, it is not a game for most people, and according to Liang Wenfeng’s ideas, the terms must be extremely strict. Regarding this financing shift, the investor judged that it was most likely for the purpose of pricing and cashing out employee options, and that it was "too late."
The logical deduction is as follows.
In the employee equity incentive system of unlisted companies, the value of options needs to be confirmed by external market pricing. The absence of external financing means that there is no valuation anchor verified by real money. The equity commitments in the hands of employees cannot be converted into clear wealth expectations, and there is a lack of sufficient liquidity and premium reference in the eyes of top talents.
The competition for talents in the AI field has reached a fierce stage: Luo Fuli, a key contributor to the DeepSeek-V2 architecture, joined Xiaomi, Guo Daya, the core author of the GRPO algorithm, joined ByteDance, and Ruan Chong, a core researcher in multi-modality, joined Yuanrong Qixing. The salary packages offered by these competitors can double or even more than DeepSeek's existing salary.
Introducing a round of small financing is essentially using a market-based transaction price to complete an official pricing for the option pool of all employees. US$300 million corresponds to about 3% of the equity. This transaction volume is enough to generate a legally binding and market-referential price anchor, but not enough to shake Liang Wenfeng’s absolute control. From this perspective, the first function of this round of financing is to "explain internally", so that past contributors have clear expectations of returns and future talents have clear incentives.
This also explains why “share is hard to grab”. If the core purpose of financing is pricing rather than introducing strategic resources, then Liang Wenfeng will tend to choose the investor with the highest degree of agreement on terms, the weakest strategic appeal, and the lowest willingness to intervene in business decisions. In the eyes of an idealistic founder, the entry of external capital itself is a necessary compromise, and his natural tendency must be to minimize its impact.
But there is a deep logical consideration here:Is the introduction of external financing really the only way to solve the option pricing problem?
In fact, option pricing of unlisted companies does not necessarily rely on equity financing. Under a mature legal and financial framework, the company can hire a third-party evaluation agency to conduct an independent valuation, or use Huanfang Quantitative Investment to establish an internal repurchase fund to repurchase employee options at fair value. These paths can also provide liquidity exits for options without diluting the founder’s control at all.
So why did Liang Wenfeng choose the path of financing? The possible answer lies in the essential difference between "market-based endorsement" and "internal valuation."
No matter how high the price of internal repurchase is, it is still essentially the company using its own money to buy its own stocks, and there is no transaction behavior of external market entities as support for fair value. In the perception of top talents, the "confidence" of this arrangement is far lower than that of introducing strategic investors, which means that an independent third party has confirmed the market value of the company's equity with real money. In other words, financing is not the “only solution” to option pricing, but it is the most credible “optimal solution”.
02
Second logic: The valuation of US$10 billion is an unreasonably low price, and it is unlikely that “outsiders” will get a share
If the essence of this round of financing is the structural design of equity incentives, then pricing is the most worthy of careful study. 10 billion US dollars, this number is too low to comply with common sense logic in the current AI valuation coordinate system.
Let’s look at the horizontal comparison first. In January 2026, Zhipu AI was listed on the Hong Kong stock market, with a market value of approximately US$6.8 billion on its first day, and a latest market value of approximately US$50.7 billion; MiniMax had a market value of approximately US$13.7 billion on its first day of listing, and its latest market value was approximately US$34.4 billion. As a large model unicorn that has not yet been listed, the valuation of Dark Side of the Moon has risen from US$4 billion in November 2025 to US$18 billion.
Let’s look at vertical logic again. Huanfang Quantitative, the parent company behind DeepSeek, has an average return rate of 56.6% in 2025 and a management scale of over 70 billion yuan, ranking second in the tens of billions of quantitative private equity performance list. According to a rough estimate based on industry practice of "1% management fee + 20% performance compensation", in 2025 alone, Huanfang Quantitative will bring approximately 5 billion yuan in revenue to Liang Wenfeng, equivalent to more than 700 million US dollars.
What’s more worth pondering is the “profitability consideration”: Magic Square Quantitative is essentially a financial machine with stable profitability. If there is a clear value connection between DeepSeek and Magic Square Quantification - whether it is financial channels or technical synergy - then the price-to-earnings ratio corresponding to a valuation of US$10 billion is only more than ten times. For a complex that possesses both top AI research and development capabilities and top quantitative trading capabilities, this pricing is difficult to reconcile under any reasonable financial model.
In both horizontal and vertical dimensions, the valuation of US$10 billion is significantly lower than the market reference frame. This naturally leads to a deeper question: Why is Liang Wenfeng willing to introduce external capital at such a low price?
A reasonable explanation is:Low valuations themselves are screening mechanisms.In a financing where "outsiders cannot get a share", pricing is not the primary consideration for founders. On the contrary, an obviously low valuation can effectively filter out those investment institutions with demanding financial return requirements and strong willingness to negotiate, and screen out partners that truly accept the rules of the game set by Liang Wenfeng. In other words, this price is not the result of a market game, but an entry barrier actively set by the founder.
But this explanation can only answer "why it is low", not "why it is this number".
So, what is the true anchor of the number 10 billion? The answer is probably hidden in the magic square’s quantified ledger.
Since DeepSeek was incubated in 2023, R&D investment, computing power procurement, and team compensation have all been borne by Huanfang Quantitative. This is an internal transfer cost that can be accurately calculated. According to cross-verifiable data in the industry, Magic Square’s cumulative investment in DeepSeek over the past three years is approximately in the order of hundreds of millions of dollars.
A $10 billion valuation release of 3% just means:The US$300 million raised in this round is approximately equal to Magic Square’s total investment in DeepSeek over the past three years.
We can regard this as a very accurate financial signal: if US$300 million corresponds to Huanfang’s cumulative investment in DeepSeek over the past three years, then after this round of financing is completed, DeepSeek will be officially independent from Huanfang in a financial sense. This means that DeepSeek’s continued losses in the future will no longer be filled by Magic Square’s profits. It will need to face the capital market on its own. This financing will be the starting point for independent operations.
03
The third level of logic: dimensionality reduction and anchoring, using equity swap to lock in structural advantages
The above two logics explain "what" this round of financing is and "why it is priced like this", but do not fully answer the question "what will the money be used for". The financing scale of US$300 million is a drop in the bucket in the current AI computing power competition.
Let’s do a simple calculation. OpenAI completed US$122 billion in financing in March 2026, with a post-investment valuation of US$852 billion; Anthropic completed US$30 billion in Series G financing in February this year, with a post-investment valuation of US$380 billion (it should be noted that despite Anthropic’s recent growth Swiftly, its annual revenue has exceeded US$30 billion and has overtaken OpenAI. At the same time, it has received valuation offers of approximately US$800 billion from investors. However, its post-money valuation in its last round of formal financing is still US$380 billion, which does not exceed OpenAI’s current valuation of US$852 billion.)
No matter which set of data you refer to, the single-round financing scale of leading players can easily reach tens of billions or even hundreds of billions of dollars, and DeepSeek's $300 million in financing is not even enough to purchase a medium-sized Wanka cluster. Not to mention, the upcoming DeepSeek V4 has a total parameter volume of one trillion and will face the exponential increase in computing power and power call requirements in the Agent era.
Large model training follows the Scaling Law, and performance improvement requires exponential computing power investment. Electricity costs account for as much as 60% to 70% of the operating costs of large AI models. Under this structure, Token can be regarded as a "power derivative" to some extent. With the release of V4 and the opening of Agent capabilities, DeepSeek will face an exponential increase in call volume, which will lead to a simultaneous surge in power costs.
This leads to an inference: US$300 million in cash financing is a drop in the bucket for computing power procurement, but if part of it is locked in power infrastructure partners through equity swaps—for example, part of the equity is exchanged for long-term low-price power supply agreements with power companies or data center operators—then the strategic value of the deal is completely different. The cost of electricity in China is only less than one-fifth of that in the United States. If this comparative advantage can be locked and amplified by DeepSeek through equity ties, it will be an infrastructure layout that is far more important than the financing amount itself.
Looking further, electricity may just be the entry point. This logic can be extended to "dimensionality reduction anchoring"General model: After large model competition enters the era of intelligent agents, the competition dimension is expanding from the model capability itself to the infrastructure level.
DeepSeek can completely use its own equity as a "high-dimensional currency" to anchor any "low-dimensional node" with structural cost advantages in the industry chain. Electricity is only the most conspicuous one. Potential targets also include domestic chip production capacity, data center cabinet resources, cross-border network bandwidth, etc. The essence of equity financing is redefined here: it is no longer just exchanging equity for cash, but exchanging equity for structural barriers.
Frankly speaking, this part is deductive conjecture and lacks solid information support. Among all the reports currently publicly available, none point to DeepSeek using the financing for power infrastructure replacement. Linking electricity costs to equity structure has not yet formed a precedent in the AI industry. Therefore, this judgment is closer to a logical derivation of possibility rather than a factual assertion.
04
The fourth level of logic: signal hedging, the balancing technique between deterministic narrative and uncertain reality
Back to the most fundamental question: Why did Liang Wenfeng choose financing at this point?
One dimension that is ignored by most analyzes is “signal hedging.” The multiple delays of DeepSeek V4 have created an accumulation of negative expectations in market opinion. It has been 15 months since the release of R1. During this period, competitors have iterated many rounds. Doubao has firmly ranked first in domestic AI applications with more than 331 million monthly active data. V4 was postponed from the original scheduled February to March this year, and then to the current rumored late April. Each postponement is eroding the market’s deterministic narrative of DeepSeek’s “always leading”.
In this context, launching a first-round financing is itself a powerful hedging signal. Its subtext is: We are evolving from a pure research institution to a commercial company with a capital governance structure, not because the technology has encountered a bottleneck, but because the organization needs to enter the next stage.
Use the financing narrative to hedge the product delay narrative, and use the certainty of "organizational evolution" to hedge the uncertainty of "technological rhythm." This level of signal value may be far more strategic than $300 million in cash.
This also explains why the financing news was released at such a low valuation. If Liang Wenfeng’s goal is just to raise money, he has every reason to wait until V4 is released and market confidence is restored before setting a price. But the value of "signal" lies precisely in the front-end. Releasing a positive structural signal when market expectations are most fragile is far more powerful than adding icing on the cake when market confidence is high.
05
Conclusion
Based on the four-fold logic, the picture of DeepSeek’s current round of financing gradually becomes clear:
It is a highly restrained equity structure design: small-amount equity transactions are used to achieve market-based pricing for employee options; investors with high cooperation are screened at obviously low valuations; equity is used as a "high-dimensional currency" to anchor dimensionality reduction at the infrastructure level; and "organizational evolution" signals are used to hedge against negative narratives about product delays.
These four logics collectively point to one conclusion: the rules of this round of financing are set entirely by the founders, and the role of “outsiders” has been carefully limited from the beginning. For investors who rush to book flights to Hangzhou over the weekend, the real test is not whether they can meet Liang Wenfeng, but whether they are willing to accept a set of game rules completely defined by the other party.