DeepSeek, which once insisted on no external funding and no IPO, has just completed its first funding round exceeding 50 billion yuan. The process involved investors scrambling for access and intermediaries hawking deals, mixing truth with chaos.
A person involved in the transaction disclosed to the media that the first-round financing negotiations for the foundational AI large model company DeepSeek have concluded, with agreements signed by the end of last month. Reports indicate this round raised over 50 billion yuan, approximately $7.4 billion. Liang Wenfeng personally contributed 20 billion yuan, the largest single investment in this round. Tencent invested 10 billion yuan, CATL invested 5 billion yuan, while NetEase, JD.com, and IDG Capital each invested 3 billion yuan. The National Artificial Intelligence Industry Investment Fund contributed 1 billion yuan.
The source added that following this round, DeepSeek's pre-money valuation is about 350 billion yuan. However, with an additional 5% ESOP (employee stock ownership plan) issuance, the actual pre-money valuation is approximately 367.5 billion yuan, or about $54.3 billion.
The final list of investors shows significant changes compared to earlier market rumors, with several leading corporate strategic investors and market-oriented investment institutions joining. Notably, the share allocated to state-backed capital turned out to be the smallest, which surprised the aforementioned transaction participant.
Since April, the soaring valuation of DeepSeek has captivated the primary market. Its valuation was $20 billion in April and has now more than doubled to around $54 billion. Upon completion of this round, DeepSeek will become the company with the largest first-round funding in the history of China's AI large model sector.
This marks the company's first external fundraising in its three-year history. Previously, founder Liang Wenfeng stated DeepSeek would pursue "no funding, no IPO, no commercialization." Market interest in the final investor lineup was unprecedented, with various investors trying every means to secure an allocation. One investor reportedly paid a 5 million yuan "meeting fee," traveled to Hangzhou three times, and waited outside DeepSeek's office but failed to meet Liang Wenfeng in person. Another investor from a state-backed fund met with eight different financial advisors (FAs) claiming to have allocations, with almost no progress.
Although DeepSeek's first funding round is complete, people are still actively selling purported allocations to the company. Capital enthusiasm in China's AI field continues to rise. Following their Hong Kong listings, Zhipu and MiniMax have seen their market capitalizations climb steadily. Moonshot AI and StepFun have successively secured multi-billion-yuan funding rounds, with valuations of leading startups doubling within months. The DeepSeek funding process epitomizes the "fear of missing out" sentiment in the capital markets.
Intermediaries and Phantom Allocations
Many investment institutions have been searching for channels, leading to the emergence of various unverified intermediaries and middlemen around DeepSeek. Investor Li Jing was also tasked with finding an allocation. After news of the funding broke, he traveled to Hangzhou three times hoping to meet Liang Wenfeng, but returned empty-handed each time. In late May, Li Jing contacted an introducer who claimed to have access to Liang. This person demanded a 5 million yuan "meeting fee" to arrange a meeting. Li Jing arrived in Hangzhou two days early and waited near DeepSeek's headquarters. Later, hearing the funding situation had changed, he gave up on the meeting and successfully retrieved the 5 million yuan.
"Being able to pay the money is a kind of capability," Li Jing said. For most LPs (limited partners), the most challenging issue isn't how much to invest, but finding the right person to give the money to. With opaque channels, people are already profiting at every intermediate step.
DeepSeek is undoubtedly one of the hottest AI targets in China currently. Even though the first-round investor list is finalized, offers to sell DeepSeek allocations are still flying around on social media today.
In recent years, the global AI investment boom has persisted, driving up valuations for popular targets. After U.S. large model startup Anthropic completed its latest $6.5 billion funding round, its valuation rose to $96.5 billion, surpassing OpenAI to become the world's highest-valued AI unicorn. For hot targets like Anthropic, investing through an SPV (special purpose vehicle) was once common industry practice, where some GPs (general partners) package assets into tradable securities to bring in new LPs, who often lack the reputation for direct investment.
Pass-through investment via special purpose funds incurs additional management fees, a method less favored by top-tier funds. However, many investors told the media they were willing to pay extra channel fees to invest, generally believing DeepSeek's valuation would continue to rise and they could still profit upon an eventual IPO.
Posing as an LP, the media inquired with one channel contact. He claimed the current channel secured a 500 million yuan allocation from the "National Integrated Circuit Fund," limited to individual entry. Investors would enter as primary LPs directly into the state-backed shareholder structure, requiring submission of sufficient asset proof within two days, with subscription potentially completed within the week.
He also mentioned that the front-end management fees for the FA and GP in this transaction were 8% and 9% respectively, totaling a hefty 17% upfront. This means for a hypothetical 100 million yuan investment, 17 million yuan would first be deducted as fees—a rarity in standard private equity deals where GPs typically charge around 2% annually based on fund size. According to the media's understanding, intermediaries selling DeepSeek allocations often propose collecting management fees for 3 to 5 years in one lump sum.
These FAs also create an atmosphere of extreme scarcity. Another FA, claiming to have secured a 2 billion yuan DeepSeek allocation from a "local state-owned entity," told the media they had several channels, stating "a 2 billion yuan fund basically sells out in two or three days." However, when asked for the specific GP name, they became evasive, only mentioning a "long-term cooperative relationship." Yet, half a month later, when asked again, the same intermediary indicated that the same channel still had available quota.
Despite unclear source channels, some are still willing to take the risk. Investor Chen Duo, who works for a state-backed fund managing tens of billions, negotiated with eight different FAs in total. These FAs claimed allocations came from "certain local state-owned entities," but in most cases, communication ceased after Chen Duo provided a POF (proof of funds). Chen Duo stated that any channel unable to provide on-site verification of the underlying investment agreement or confirmation letter was essentially "bluffing" him.
Through these interactions, Chen Duo gradually understood the fundraising logic. Many institutions, to secure an allocation, first raise a fund, then use the money in the account to request an allocation from DeepSeek. FAs then create a seemingly reliable channel using rhetoric like "backed by a senior leader" to prompt investors to quickly show proof of funds. "(These institutions) want Liang Wenfeng to think they already have money in the account to get the allocation." This creates a chicken-and-egg cycle, as even with sufficient funds raised, securing a DeepSeek allocation is not guaranteed.
An investment manager at a local state-backed fund told the media that FAs still approach him to introduce LPs, which he finds absurd. "DeepSeek simply doesn't lack LPs and doesn't need FAs." He believes ordinary people wouldn't trust these unclear channels, except perhaps some very small institutions or individual LPs who might be persuaded.
A person familiar with DeepSeek's operations revealed to the media that all capital entering this funding round came from main funds, with no allocations through any special purpose funds. Reports indicate Liang Wenfeng and his team also required verification of all LP identities participating in this round to prevent allocations falling into unknown investors' hands. The person stated that FAs still fundraising ultimately won't succeed, questioning their motivation.
Who Gets a Seat at the Table?
A private equity fund partner, Tang Yu, has been following DeepSeek since its surge in popularity in February 2025. After news of open funding emerged, Tang Yu, privately commissioned by a state-backed fund with 3 billion yuan investment intent, searched everywhere for channels. Through personal connections, Tang Yu contacted a senior executive at High-Flyer Quant, a quantitative hedge fund controlled by Liang Wenfeng that has long funded DeepSeek's R&D costs.
Tang Yu and this executive have known each other for years. He initially thought the 3 billion yuan could open the door, but the executive merely left a company email address, asking him to send institutional introduction, letter of intent, and cooperation proposal via email, stating "a responsible person will review them uniformly." He never received a positive response thereafter.
Tang Yu claimed all interested parties able to directly contact this executive likely received similar replies. On May 28, he inquired again with the High-Flyer executive about funding developments. The response was that participating in this round was definitely too late, and the timing of the next round remained unknown.
Multiple investors involved in the transaction told the media that investors for this round were primarily selected personally by Liang Wenfeng, who favored institutions wishing for his success or having good relations with him. Contact was mostly initiated by Liang Wenfeng's invitation; those not invited found it difficult to establish truly effective connections, with repeated meeting attempts failing to yield cooperation. Another investor, Hu Jie, also noted that funds unfamiliar with Liang Wenfeng or lacking reputation basically couldn't get in.
Hu Jie works at a market-oriented investment institution where AI is a key focus area. However, when DeepSeek first opened funding, his interest was limited because initial understanding suggested this round might be state-capital-led, and the deal size was already large, making investment returns potentially not very high. Hu Jie wasn't alone in this thinking; many institutional investors initially believed this was a state-capital-led deal.
But after this year's May Day holiday, Hu Jie observed several investment institutions, including Monolith Capital and Hillhouse Capital, intensively engaging with DeepSeek. His institution subsequently joined the negotiations.
What surprised multiple transaction participants was that the final state-backed share was lower than expected. However, according to reports, only the National Artificial Intelligence Industry Investment Fund made a direct investment in DeepSeek in this round, free from lock-up restrictions and enjoying company voting rights. Other institutions invested in a limited partnership managed by Liang Wenfeng, not directly in DeepSeek itself, with shares locked up for five years, during which they cannot sell their holdings.
Public information shows the National Artificial Intelligence Industry Investment Fund Partnership (Limited Partnership) was established in early 2025 by Guozhitou Private Fund Management Co., Ltd. and the National Integrated Circuit Industry Investment Fund Phase III Co., Ltd. Guozhitou was founded in 2024, with its recent investments concentrated in AI, chips, and robotics, including Moonshot AI, MetaX, and CloudWalk Technology.
In this funding round, many state-backed funds were also actively seeking allocations. Chen Duo, serving a state-backed fund, contacted eight FAs before barely touching the edge of the deal. Another local state-backed fund investment manager actively engaging with AI companies told the media he started paying attention to DeepSeek during the 2025 Spring Festival. Other targets he favored included star model companies like Zhipu and Moonshot AI, but he failed to invest in any, even lacking access.
"Objectively speaking, state capital has its limitations," he said. "Large companies require fast capital deployment and single deal sizes are huge. For state capital, this means risk."
The investor list also includes several major tech giants. Among them, Tencent secured the largest share, followed by CATL. Previous media reports suggested Alibaba was also in negotiations with DeepSeek, but Alibaba has publicly denied this, expressing limited interest.
Multiple AI investors analyzed that between Tencent and Alibaba, Tencent had stronger motivation to join. Its foundational model capabilities have long lagged behind the first tier, giving it a stronger desire for strategic cooperation with DeepSeek. Alibaba, having built its own ecosystem from models to chips, has less strategic urgency to invest in DeepSeek.
A transaction participant believed the basis for cooperation between DeepSeek and Alibaba was also weaker than with Tencent.
The participation of CATL, a new energy giant, indicates to some extent that AI model vendors are strategically considering the energy variable behind computing infrastructure. A transaction source told the media that DeepSeek is already building its own data centers, planning to operate computing power directly. Public information shows that in late April, DeepSeek posted several computing-power-related job openings, including Senior Data Center Operations Engineer and Senior Delivery Manager, requiring candidates to manage the entire process from project initiation, construction to operation, with locations in Ulanqab, Inner Mongolia.
DeepSeek's Valuation Rationale
According to multiple investors in the AI field, DeepSeek's decision to open funding has at least two reasons. First, DeepSeek's model iteration requires continuous large-scale investment in training compute. Sufficient capital is crucial to support this long-cycle endeavor, as the self-generated assets previously relying on High-Flyer Quant may not meet future R&D demands.
Chen Shi, a partner at Ruifeng Capital, analyzed that for leading large model companies like Anthropic, their business model is the industry's "top-tier model" but also the "most pressured model" because the Scaling Law still applies. Each generation's training cost exceeds the previous, requiring continuous investment in compute and data. Once falling behind, they risk being forgotten by the market. As an AI startup, today's primary market can hardly meet the funds needed for model iteration; the next step can only be entering the secondary market.
The second reason is to provide market validation for employee stock options to retain talent. This round issued an additional 5% employee option pool, worth approximately 17.5 billion yuan. However, previous media analysis indicated that as of the V4 model release, DeepSeek's core research team had not faced talent attrition issues.
Hu Jie stated that DeepSeek employees also hope for a company IPO. An AI practitioner familiar with DeepSeek's R&D staff said it's unsurprising employees expect an IPO. "People aren't saints; those staying at DeepSeek can't rely on ideals alone."
Since 2025, several star Chinese AI large model companies have been intensively fundraising and pursuing listings, accelerating their capitalization. Besides Zhipu and MiniMax, which listed on the Hong Kong stock exchange earlier this year, StepFun's latest funding round is reportedly around $2.5 billion, with an expected cornerstone valuation of about $10 billion. Business information platform Qichacha shows Moonshot AI has raised approximately $4.4 billion cumulatively, with a latest pre-money valuation of $20 billion. Moonshot AI is dismantling its VIE (Variable Interest Entity) and red-chip structure in preparation for listing, while StepFun aims to file as early as the first half of this year.
In comparison, DeepSeek started fundraising the latest, yet its first round reached about $7.4 billion with a valuation of approximately $54.3 billion, far exceeding its competitors.
Internal and external factors have driven DeepSeek's continuously rising valuation above other model companies. Multiple transaction participants consider the current range reasonable because, in terms of foundational model capability, DeepSeek's inference efficiency is recognized as industry-leading.
Since its breakthrough R1 model last year, DeepSeek has maintained low cost and high performance. Taking the V4-Pro released in April as an example, official disclosures state that in a 1 million Token long-context scenario, the FLOPs per Token inference is only 27% of the previous generation V3.2. This efficiency advantage directly reflects in API pricing: V4-Pro costs 0.025 yuan per million Tokens, among the lowest globally for mainstream models. This extreme low price has made DeepSeek widely popular among global developers. On the latest leaderboard of the AI model aggregation platform OpenRouter, DeepSeek V4 Flash ranks first in Token consumption this month.
A professional from a U.S. dollar fund familiar with the AI field told the media that on the OpenRouter platform, the proportion of Chinese model calls has surged from 1% three years ago to 60%. Chinese model companies have successfully extended the advantage of extreme supply chain cost reduction to the AI field, gradually making large models a standardized commodity, with DeepSeek executing this most effectively.
Furthermore, multiple investors and industry experts analyzed for the media that another reason for DeepSeek's high valuation is its rise as a national-level AI strategic platform.
External factors also play a role. The secondary market performance of Zhipu and MiniMax has stimulated pricing for later entrants. For instance, Zhipu's IPO market capitalization was 57.9 billion Hong Kong dollars, and its latest market cap has risen to 740.1 billion Hong Kong dollars. A state-backed fund professional familiar with AI venture capital analyzed that after this year's Spring Festival, the rapid surge of comparable model companies, including Zhipu and MiniMax, created a strong capital market wealth effect.
This fund professional also believes the current moment is a key node for discussing large model commercialization because, after three to four years of development, technology can now support mature product deployment, and the market urgently needs to verify AI's commercial value. Against this backdrop, more primary market capital is flowing into large model companies. In the past, these model startups were generally viewed by state-backed funds as targets "lacking certainty."
Unlike the overall market's "fear of missing out" (FOMO) sentiment, some investors who never had contact interest believe DeepSeek's current valuation is too high, "hard to comprehend." Participating in negotiations consumes significant time and monetary costs, with a low probability of successful investment, making the transaction not worthwhile considering investment returns. Hu Jie views DeepSeek as a "white horse stock," referring to those with high market capitalization and relatively stable performance trends.
Facing capital market scrutiny, DeepSeek cannot avoid the challenge all tech startups face: how to balance technological ideals with commercial returns long-term. How to commercialize is a crucial question hanging over DeepSeek's high valuation. According to previous media reports, in March, a DeepSeek executive revealed in a meeting that the company is aware of external focus on its business model implementation and technological progress. DeepSeek has been making many efforts and attempts and has preliminarily validated some paths.
Compared to other model companies and internet giants, DeepSeek's progress on the product front has been slow, with previous focus largely on model training. Regarding whether DeepSeek should develop products, Hu Jie told the media that Liang Wenfeng likely realized the importance of productization around late last year, but the team hasn't found a suitable leader, struggling with the product roadmap. "The product matter requires Liang Wenfeng's own reflection." Recently, public recruitment websites show DeepSeek opening many product-related positions, including Model Strategy Product Manager, Agent Data Strategy Engineer, etc.
However, this doesn't necessarily mean DeepSeek has changed its course. According to a May 22 media report, Liang Wenfeng stated in a meeting with investors that the company will continue advancing open-source AI models, aiming to achieve Artificial General Intelligence (AGI). In his statement, technological breakthrough is DeepSeek's core mission to push technical boundaries, not pursue profitability.
Hu Jie noted that fundraising is merely a capital operation; DeepSeek's daily operations and technological roadmap are not necessarily bound by capital markets.
An investor on the transaction list told the media that for DeepSeek, he believes short-term commercialization isn't that important; achieving AGI is paramount, and China must have its own AGI.
Comments