This May, China's AI industry saw exceptional activity in the primary market.
First up is DeepSeek, a domestic AI unicorn that is advancing its first external funding round since its founding. Recent information shows that investors such as the National AI Industry Investment Fund and Tencent have entered the negotiation list.
A month ago, DeepSeek was rumored to be raising funds at a valuation exceeding $10 billion; a few weeks later, its valuation has been pushed up to the range of $45–50 billion. Based on current information, DeepSeek’s current funding round may reach $3–4 billion, with the proceeds primarily allocated to computing infrastructure, research and development, and employee incentives.
Shortly after, Moonshot AI was revealed to have completed or nearly completed a $2 billion funding round, pushing its post-money valuation beyond $20 billion. This round was led by Meituan Dragonfly, with participation from China Mobile, CPE, and others.
Almost simultaneously, another AI “Little Six” company, JieXing星辰, also reported that a nearly $2.5 billion funding round is about to be finalized. Notably, this round includes capital from companies in the mobile and consumer electronics supply chain, such as Ascend, Longqi, OmniVision, and ZTE. Hong Kong Investment Management Limited, often referred to as the “Hong Kong version of Temasek,” has also been reported to join the shareholder list.
Three leading AI unicorns are simultaneously advancing their fundraising efforts—not by coincidence, but as a result of a collective reassessment of China’s AI industry. Represented by DeepSeek, the valuation threshold for China’s top foundational model companies has now reached over RMB 10 billion.
Keep in mind that a few months ago, Zhipu and MiniMax, the first to pursue a listing on the Hong Kong stock market, had pre-IPO valuations of around HK$50 billion each.
After listing on the Hong Kong stock market in January, these two companies quickly became rare benchmark stocks in the secondary market, with their market capitalizations multiplying significantly. As of this month’s 9th, Zhipu’s latest market capitalization has surpassed HK$400 billion, while MiniMax remains steadily at HK$230 billion.
The secondary market has already set the example; the enthusiasm in the primary market now only needs a spark to ignite. With DeepSeek’s funding signal, the next wave of financing and IPOs for foundational model companies has finally reached the remaining AI unicorns.
A
No one expected that the pricing model of China’s AI unicorn would be overturned in just the past few weeks.
Three weeks ago, media reports surfaced that DeepSeek was in talks for its first external funding round since its founding, aiming to raise at least $300 million with a valuation exceeding $10 billion—but this signal was quickly picked up by the capital markets and accelerated.
A few days ago, a major name appeared on DeepSeek’s list of investors: China’s National Integrated Circuit Industry Investment Fund, also known as the "National Big Fund," is in talks to lead DeepSeek’s seed round, with a potential valuation of up to $45 billion.
According to the latest reports, DeepSeek’s valuation may have reached $50 billion, with fundraising expected to amount to $3–4 billion, and Tencent is also in discussions to participate.
In other words, DeepSeek's valuation was rapidly multiplied fivefold in the primary market over the course of a few weeks.
This round of valuation increase is partly due to DeepSeek being viewed by both capital and industry players as a core competitor in China’s domestic AI infrastructure. Even though DeepSeek’s revenue remains limited, investors still regard it as a key company in China’s strategic AI initiative.
On the other hand, the market has long been anticipating DeepSeek opening its doors to funding, but for years, DeepSeek has maintained the image of a "technological idealist" rejecting external capital.
Over the past two years, DeepSeek primarily relied on its founder, Liang Wenhong, and funding from its parent company, Quantex, making it difficult for external investors to enter. However, starting in 2026, investment in the AI industry has begun to grow exponentially, with rising costs for research and development, computing power, and talent.
DeepSeek also faces competition from companies like ByteDance and Alibaba, not only in the model market but also in the talent acquisition arena. For example, Guo Daya, a core researcher at DeepSeek, recently joined ByteDance.
On the other hand, expanding computing power is a practical justification for fundraising. The parent company, Qianwen, previously reserved an NVIDIA A100 cluster of approximately 10,000 units for DeepSeek; however, with the transition to newer models like V4, both training and inference have significantly increased computational demands. Even as the company accelerates adaptation to domestic computing resources, further expansion of its computational infrastructure still requires external funding.
Notably, Alibaba once attempted to join DeepSeek’s funding round.
Several weeks ago, reports revealed that Tencent and Alibaba had both engaged in investment discussions with DeepSeek. However, recent rumors suggest that Alibaba and DeepSeek have failed to reach an agreement. According to the latest report from Economic Daily News, market participants stated that Alibaba "likely did not enter into negotiations."
On the other hand, DeepSeek’s latest funding round coincided almost exactly with the launch of its new-generation model, V4.
Two weeks ago, DeepSeek-V4 Preview was officially launched and open-sourced. According to the official documentation, the V4 series consists of two versions: V4-Pro and V4-Flash, both with a maximum context length of 1M; the total parameters of V4-Pro reach an impressive 1.6T.
The external response to V4 has upheld DeepSeek’s longstanding value-for-money approach. The pricing for V4 Pro is just $0.435 per million input tokens and $0.87 per million output tokens; on the developer community Hacker News, a user noted that while running the same code review took about twice as long as with Opus/GPT, the cost was less than one-tenth.
In contrast, Anthropic's latest enterprise developer cost for Claude Code has reached $13 per day, or $150–$250 per month.
Therefore, the market feedback for V4 resembles a one-time value-for-money regression: Artificial Analysis ranks V4 Pro in the top tier of open-source weighted models, just behind Kimi K2.6.
More importantly, the significance of V4 lies not just in model upgrades, but in its tight integration with domestic computing power.
A key change in DeepSeek V4 is its optimization for Huawei’s most advanced Ascend AI chips. Huawei immediately announced that its Ascend super-nodes will fully support DeepSeek V4; shortly afterward, industry reports indicated that leading AI companies such as ByteDance, Tencent, and Alibaba are accelerating their efforts to secure Huawei’s Ascend 950 series chips following the V4 release.
This has placed DeepSeek at the center of China's domestic AI computing ecosystem.
In the past, Chinese large model companies found it difficult to bypass NVIDIA's CUDA ecosystem—the more advanced the models became, the more they relied on high-end GPUs. The validation of DeepSeek V4 within Huawei's Ascend ecosystem signifies a crucial alignment between domestic foundation models and domestic computing power.
This is also why DeepSeek has become the leader in this round of funding.
As DeepSeek's valuation rises, other leading foundation model companies now have a new benchmark for valuation.
B
After DeepSeek raised the valuation benchmark, several unlisted AI unicorns, including Moonshot AI and Jiepao星辰, quickly underwent a reassessment of their valuations.
Public information shows that Moonshot’s recent funding round amounted to approximately $2 billion, pushing its post-money valuation beyond $20 billion. The round was led by Meituan Dragonfly, with participation from China Mobile, CPE, and others, among which Meituan Dragonfly invested over $200 million.
This was the largest single investment since Moonshot AI's founding and the highest private funding amount for a large model startup in China at that time.
As for why we say "at that time," it's because Jiepao Xingchen and DeepSeek soon broke this record.
However, setting aside the single-round amount, Moonshot's fundraising pace has been intensely active this year. In the first two months of this year, Moonshot completed three consecutive funding rounds, and with this most recent $2 billion round in May, it has raised over $3.9 billion in less than six months.
In other words, the dark side of the moon has more than quadrupled in valuation since the end of last year.
Notably, behind every funding round of Moonshot AI stands a significant presence from the Meituan ecosystem.
Meituan Longzhu was established in 2017 as an industry investment platform rooted in the Meituan ecosystem, initially focusing primarily on consumer and local life investments, but has recently increased its emphasis on cutting-edge technology. In July 2023, Longzhu partner Wang Xinyu participated in the financing negotiations for Moonshot AI, making an early major investment prior to Kimi's rise to prominence.
After the most recent round of increased investment, Wang Xinyu publicly appeared to support Moonshot AI. In a recent interview, he revealed that Kimi’s ARR surpassed $100 million in early March following the K2.5 update and rose to over $200 million in April, with both paid subscriptions and API calls accelerating.
Wang Huiwen, co-founder of Meituan, is another key figure in Moonshot AI’s funding journey. Public reports show that he has invested in Moonshot AI multiple times in his personal capacity and continued to participate in last year’s $500 million Series C round, bringing his total investment to approximately $70 million.
On the model front, Moonshot AI recently released its latest model, K2.6, which enhances programming capabilities, long-range task execution, and agent cluster performance, supporting up to 300 sub-agents working together. Kimi is also beginning synchronized testing of the Claw group feature.
Test data shows that K2.6 can encode continuously for 13 hours and write or modify over 4,000 lines of code, making it the most powerful code-generating model Kimi has developed to date. After Zhipu and MiniMax went public, Moonshot AI was once regarded as the most likely among the remaining members of the "Six Little Dragons" to pursue an IPO.
At the same time, another "dragon" company, Jieyue Xingchen, quickly took action.
Over the past two days, multiple media outlets reported that Jiepoin Stars is set to complete a nearly $2.5 billion funding round. In addition to various local state-owned capital firms, key players from the mobile phone industry supply chain—including Accton, Longcheer, OmniVision, and ZTE—have collectively invested, covering segments such as device manufacturing, core components, and communication terminals.
Among them, Accton and Longcheer are leading smartphone ODMs, OmniVision is a core manufacturer of image sensors, and ZTE is a provider of terminals and communication equipment. Their investment in Step Function somewhat signals industry partners lending their support.
On the other hand, Jiepao has been highly active in capital maneuvers this year. In January, Jiepao Xingchen completed a B+ round of financing exceeding RMB 5 billion, with investors including Shanghai State-owned Assets, China Life Equity, and Pudong Venture Capital. On the same day, Yin Qi, Chairman of Qianli Technology, was appointed Chairman of Jiepao Xingchen.
Yin Qi’s involvement rapidly shifted Step AI’s external perception. Jiang Daxin remains Step AI’s founder and CEO, representing model development and company operations, while Yin Qi serves more as a bridge connecting capital, industry, and end-use scenarios. Meanwhile, Yin Qi’s other role as co-chairman has deepened Step AI’s ties with Qianli Technology and its underlying Geely-affiliated capital.
On April 22, Qianli Technology announced a comprehensive strategic partnership with Jieyue Xingchen, with both parties collaborating to build the "Native Autonomous Driving Base Model."
Former Huawei Honor CEO Zhao Ming made a public appearance after joining Qianli, stating at the launch event that Qianli Tech and Jieyue will deeply integrate from the foundational model pre-training stage, feeding common corpora alongside real-world autonomous driving perception and control data from the same source.
This explains why a group of end-market industry capital appears on Jiep's funding list.
Step’s recent business initiatives have also focused on edge-side agents, voice models, and multimodal capabilities. On the 8th of this month, Step released StepAudio 2.5 Realtime, positioned as a new-generation real-time large voice model. Earlier, Step launched the Step-GUI series, aiming to extend GUI agent capabilities to mobile devices and other endpoints.
On the other hand, despite continuous fundraising efforts by Step Intelligence and Moonshot, a crucial step remains to successfully complete a Hong Kong IPO: dismantling the red chip structure.
A red chip structure is a corporate ownership framework in which a domestic Chinese operating entity channels overseas financing and listing arrangements through an overseas holding company. In the past, many internet companies used red chip/VIE structures to list in the U.S. or Hong Kong.
However, in recent years, red chip companies in industries involving data, AI, and hard technology have faced increasingly stringent requirements for overseas listing registration and regulatory review. For companies preparing to list in Hong Kong, dismantling the red chip structure often means bringing equity and control relationships under domestic regulatory registration.
Public reports indicate that Jiepao Xingchen has already dismantled its red chip structure, which is considered a key prerequisite for a Hong Kong IPO; in contrast, Moonshot AI is still in the negotiation phase, evaluating timing windows and structural options, with no public information suggesting it has initiated equity restructuring.
It remains to be seen who will secure the AI Hong Kong stock window first in this round.
Another former "Six Dragons" company, Zero1 AI, has also been reported to be pursuing an IPO in Hong Kong, with the latest reports indicating it is currently raising pre-IPO funding and planning a listing on the Hong Kong Stock Exchange.
Yiwan Wuxing responded that it is currently focused on enterprise agents and market deployment, with a capital strategy that is “open but cautious,” and no further disclosures are available at this time. Regarding valuation, Yiwan Wuxing was valued at over $1 billion in 2023, but recent funding rounds and new valuation information have been limited. On the business front, the company has scaled back large-scale foundation model pre-training and shifted its focus toward enterprise agents, vertical use cases, and industry-specific deliveries.
In reality, the concept of the "Six Little Dragons" is already a thing of the past. Today, leading foundation model companies are highly sought after by investment institutions, especially following the industry’s reassessment triggered by DeepSeek’s recent developments. On the large model table, only players with assets exceeding tens of billions are in the game.
C
In this capital-rich May, the valuation landscape for the large model industry in the primary market has completely changed.
According to the latest reports, DeepSeek has an estimated valuation of $45–50 billion, equivalent to approximately RMB 320–360 billion; Moonshot AI has a valuation of $20 billion, equivalent to over RMB 140 billion; while Jiepang Xingchen, while raising $2.5 billion, has already set its target valuation at $10 billion.
In other words, the primary market is defaulting to a new standard:
China's leading foundational model companies are essentially valued or market-capitalized at a threshold of tens of billions of RMB.
Multiple factors are driving this change, beginning with the advancement of domestic AI infrastructure.
The integration of DeepSeek and Huawei Ascend on V4 has begun to align the paths of domestic foundation models and domestic computing power, prompting a swift response from capital markets.
On the first trading day of May, China's computing power sector surged across the board, with the cloud services and semiconductor sub-industry indices rising 5.77% and 5.48% respectively, ranking as the second- and third-largest gains among all industries. A combined 26 stocks from these two sectors hit trading limits or rose over 10%. The primary driver behind this movement is the signal released by DeepSeek V4 in the domain of domestic computing power.
Analyses from multiple securities firms show that domestic AI chip and solution providers generally achieved several-fold revenue growth in the first quarters of 2025 and 2026, with many companies turning profitable or reporting significantly increased profits, indicating the industry is transitioning from the R&D phase to the commercial returns phase.
In other words, this round of revaluation encompasses not just foundational companies, but the entire application ecosystem loop of models, chips, and cloud platforms.
On the other hand, under the wave of AI agents, large models are beginning to achieve commercial success in certain sectors.
From the early-year craze around "crayfish," to a surge in global token usage, and the widespread adoption of Vibe Coding, AI programming, and agent tools, some large model companies have already found their commercial breakthrough.
Zhipu and MiniMax were among the first companies to benefit from this opportunity in the Hong Kong stock market. Amid the early-year "lobster" craze, their flagship models temporarily surpassed overseas models like Gemini and Claude on the OpenRouter model usage leaderboard, driving continuous growth in their stock prices.
In other words, commercial returns are already beginning to emerge in scenarios such as Vibe Coding and Agent tools—areas where DeepSeek and Moonshot excel.
The most critical point is that in today’s large model market, fewer powerful general-purpose base model players remain, and industrial capital is beginning to secure entry points in advance.
Not every company can continue burning cash on a general-purpose foundational model—this has long been an industry consensus.
Li Kaifu, founder of 01.AI, has bluntly stated, “Only big companies can afford to burn cash on massive models.” He believes that, in the Chinese market, only three companies—DeepSeek, Alibaba, and ByteDance—may ultimately remain as major foundation model providers. Regardless of whether this prediction ultimately proves accurate, it reflects a reality: foundation models will be reserved for only a few companies with substantial capabilities.
Under these circumstances, high-quality pre-listed base companies will become even scarcer.
The list of investors in Jiepao Xingchen, including numerous end-device manufacturers, reflects a deeper integration between large model companies and the external industrial chain. Smartphones, PCs, automobiles, and wearable devices are all seeking new interaction methods in the AI era—exploring voice, GUI agents, multimodal understanding, and on-device inference to define the next generation of endpoints.
Industrial capital investing in AI is essentially investing in the future "native AI brain" for their own industries.
However, this round of funding also has clear uncertainties.
The secondary market has already provided an initial warning. After Zhipu and MiniMax went public, their prices surged significantly, raising the valuation ceiling for AI assets in Hong Kong’s stock market.
However, a high valuation also means higher volatility. After the release of DeepSeek V4, the stock prices of Zhipu and MiniMax experienced significant corrections, as the market feared that new models could trigger price competition.
On the other hand, Zhipu and MiniMax are expected to be included in the Hang Seng Tech Index and the Stock Connect program in June, but the real stress test will come in July with the lock-up expirations—nearly 50% of MiniMax’s shares are set to be unlocked, putting the liquidity premium to the test.
This precisely shows that the valuation of large model companies is not guaranteed to rise indefinitely. Capital markets can price in the future in advance and will quickly adjust expectations—this applies equally to DeepSeek, Moonshot AI, and Step AI.
In addition, the AI cloud market representing the B-end base and C-end traffic are still firmly controlled by major companies such as ByteDance, Alibaba, Baidu, and Tencent.
According to QuestMobile’s Q1 report, the monthly active users for Doubao, Qwen, and DeepSeek reached 340 million, 170 million, and 130 million respectively, with both user base size and engagement rising across the industry. Among these, DeepSeek stands out as the exception—it has no support from major tech companies’ advertising budgets, yet still maintains a leading position among domestic C-end AI applications.
But this also highlights how difficult it is to replicate DeepSeek’s uniqueness. Most AI unicorns must still carve out their own space amid intense competition from big tech companies, continuously refining their foundational models while exploring viable commercial pathways.
At this “Crazy May,” it’s still far from the time for several foundation model companies to deliver answers.
This article is from the WeChat official account "Letter List" (ID: wujicaijing), authored by Li Zhaofeng.