Among the six rising AI stars, who can "leap over the dragon gate"?

Since the beginning of this year, a quiet race to go public has emerged in China's AI venture capital circle. In April, Zhipu took the lead in submitting an application for an initial public offering to the Beijing Securities Regulatory Bureau. In June, media reports claimed that MiniMax was preparing for an IPO in Hong Kong. Moreover, more covert news has been circulating in the AI venture capital circle: "Five out of the 'Six Little Dragons' are already preparing for listing."

Two years ago, ChatGPT ignited the global competition in large models. Zhipu, MiniMax, Dark Side of the Moon, Step Star, Baichuan Intelligence, and Lingyi Wanwu were hailed by capital as the "Six Little Dragons of AI."

However, the market has gradually become disillusioned with large models, and capital has returned to rationality. The stances of AI startups and internet giants have also changed. Some of the former AI startups have scaled back their operations, some have shifted to vertical fields, and some have been acquired by large companies... The development logic of the large model industry has been completely overturned.

As the large model elimination competition enters its brutal middle stage, going public has become a watershed between survival and extinction. However, just like a carp leaping over the dragon gate, leaping over doesn't necessarily mean becoming a dragon, and failing to leap might lead to a faster decline.

01 The Fading Aura of Capital Darlings

In 2023, ChatGPT swept across the globe in a phenomenon - level manner, and domestic technology companies and AI startups rushed to enter the field. Among them, the "Six Little Dragons" stood out in the fierce competition and became the capital darlings in the industry. In 2023, the "Six Little Dragons" raised a cumulative total of over 6 billion yuan, accounting for more than half of the early - stage financing of domestic large models.

The logic behind the crazy capital investment is clear. On the one hand, the "Six Little Dragons" seized the time window of the AI boom. Zhipu and MiniMax made forward - looking technological arrangements at the inflection point. As early as 2022, Zhipu had trained the first domestic open - source model with a scale of over 100 billion parameters, gaining a leading edge in model parameter scale and engineering capabilities.

Moreover, Baichuan Intelligence, Dark Side of the Moon, Step Star, and Lingyi Wanwu were all established in the first half of 2023, targeting the window period when the capital market had abundant liquidity.

Yang Zhilin, the founder of Dark Side of the Moon, once said, "We started to focus on the first - round financing in February 2023. If we had postponed it until April, there would have been almost no chance. If it had been too early, due to the pandemic, there would also have been no chance. The real window period was only one month."

On the other hand, the support of star teams is also crucial. In addition to companies like Zhipu and MiniMax that already have certain technological achievements, for other startups to gain investors' recognition, the presence of founders and top - notch technologists is a key factor.

Li Kaifu, the former president of Google Greater China, founded Lingyi Wanwu in March 2023. In just a few months, he attracted dozens of core members from domestic and foreign companies such as Alibaba, Baidu, and Google. Yang Zhilin, the founder of Dark Side of the Moon, is also known as a genius academic, having achieved remarkable results in academic research.

The rise of the "Six Little Dragons" was a perfect combination of timing, location, and human factors. However, for them, financing was just the "first hurdle," and the real market test was only beginning.

The inflection point arrived quietly in 2024. After a year of "burning money for scale," the large model industry began to realize that the era of blindly pursuing technical parameters had passed. For investors and consumers, technological iteration and commercialization models were more important, and the watershed in the large model industry began to emerge.

After that, the "Six Little Dragons" started to diverge. Based on their respective technological reserves and strategic focuses, they explored different paths and engaged in differentiated competition in terms of technological routes, product portfolios, and commercialization paths.

Since 2024, Baichuan Intelligence has stopped training ultra - large - scale general models and instead focused on the medical vertical field. At the end of last year, Lingyi Wanwu laid off its entire pre - training algorithm team and Infra team, and later the Infra team received an offer from Alibaba Cloud.

Dark Side of the Moon and MiniMax pay more attention to C - end products and applications. The former focuses on long - text and open - source technologies, while the latter continues to invest in video generation, visual multimodality, and overseas products. Only Zhipu and Step Star still adhere to the development of AGI large models, relying on the government resources of Beijing and Shanghai respectively and betting on the government and enterprise market.

Behind the divergence of AI startups is the overwhelming entry of internet giants. Industry insiders point out that a team capable of launching the next - generation base large model must have a top - notch talent team and a large amount of training data, which requires sufficient capital and a user ecosystem. Only internet giants can meet these requirements.

Pan Xin, the former chief architect of Lingyi Wanwu, once said that domestic first - tier large model companies need to burn about $1 billion a year, and international first - tier companies may need $5 billion a year. Training a multimodal large model once costs about $10 million. Although the emergence of DeepSeek has reduced the training and inference costs of large models, the industry still believes that the costs remain high.

In addition, the flow of talent is also crucial. Sam Altman, the CEO of OpenAI, recently revealed in an interview that Meta had made a high - paying offer to his team, even including a $100 million signing bonus. In 2024, 9 out of the 11 founding members of OpenAI chose to leave.

In the domestic large model industry, top technologists and corporate executives are also frequently changing jobs. According to statistics from Zhidongxi, in the first half of this year, 12 executives from the "Six Little Dragons" left their positions, and many talents have flowed to large companies such as ByteDance and JD.com. The siphon effect of large companies is very obvious.

Moreover, the ecological advantages of large companies cannot be underestimated. For example, in terms of data, they can obtain more real - world data from different scenarios through user interactions, and the quality of data annotation and real - time update capabilities far exceed the industry average. In terms of scenario applications, the existing business scenarios of large companies can provide a commercial closed - loop for AI.

02 The Hanging Sword of "Commercialization"

Therefore, after internet giants such as ByteDance, Alibaba, Baidu, and Tencent entered the market, the first - mover advantage of the "Six Little Dragons" has been gradually overtaken. In 2025, "commercialization" has become a sword hanging over their heads.

On the one hand, the open - source model represented by DeepSeek has directly impacted the original large model landscape. Based on lower costs and an open - source ecosystem, DeepSeek has demonstrated performance comparable to, or even better than, closed - source models such as OpenAI's GPT - 4 in multiple key tasks.

This means that the previous logic of the "Six Little Tigers" developing products in isolation will be completely overturned. The open - source strategy can attract more developers to participate in ecological construction, and the business model of traditional large model manufacturers relying on licensing fees for profit will become increasingly narrow.

In the debate over the advantages and disadvantages of open - source and closed - source models, domestic large model companies have quickly changed their attitudes. For example, Robin Li once said that "closed - source models will continue to maintain an advantage over open - source models," but Baidu has announced that the Wenxin Big Model 4.5 series will be freely open - sourced.

On the other hand, AI startups have to make choices in terms of products. DeepSeek has suddenly raised the "capability ceiling" of large models at the inference end, and AI startups have begun to actively or passively adjust their business directions.

In the C - end field, Dark Side of the Moon and MiniMax have been leading the way, but large companies have been closely following.

In February this year, it was reported that Dark Side of the Moon had significantly reduced its product promotion budget, and the download volume of Kimi also declined accordingly. Meanwhile, Tencent Yuanbao announced its integration with DeepSeek, invested about 300 million yuan in promotion within a month, and promoted it across all Tencent ecosystem products, helping Yuanbao top the download list.

Industry insiders point out that compared with large companies, AI startups do not have their own advertising platforms, and the traffic cost is extremely high. Once they compete directly with internet giants, they have little chance of winning in terms of capital, user scale, and ecological resources.

Although MiniMax focuses on the overseas market, after the "removal incident" of Talkie, its attempt to bypass the domestic market competition by going overseas is clearly full of uncertainties.

In the B - end field, the price war has put many large model companies in a dilemma of "no customers without price cuts, and no profit with price cuts." In May last year, ByteDance officially launched the Doubao large model, firing the "first shot" in the price cut by large companies. Zhipu, Doubao, Baidu, and Alibaba Tongyi followed suit and cut prices.

In 2025, the intensity of the price war in the large model industry further escalated. Alibaba actively cut prices in the cloud service segment to seize the market in conjunction with its new model. The prices of several billing models under Dark Side of the Moon dropped by up to 83.3%, and the price cut of Zhipu's large models reached over 90% at most.

However, according to a report by Zimubang, a marketing staff member of a large model startup said that "several AI startups were subsidizing the development of small - and medium - sized enterprise users last year." It's easy to imagine that if large companies increase subsidies and the customers of AI startups are not bound to cloud services, these customers are very likely to be snatched away by large companies.

The "Monitoring Report on China's Large Model Winning Bids" shows that among the top 6 general large model manufacturers in the winning bid ranking last year, except for Zhipu, the others are all large companies that have been established for many years or listed companies.

Therefore, the current commercialization achievements of the "Six Little Dragons" can be said to be still "on the surface." Even Zhipu, which firmly adheres to the TOB route, claims that its commercialization revenue increased by over 100% in the first 11 months of 2024. However, according to a report by Caijing Magazine, Zhipu's revenue in 2024 reached 300 million yuan, but the loss was as high as 2 billion yuan. The commercialization revenue still failed to cover the high R & D investment.

The capital market is quite familiar with such stories and can even be said to be "immune." SenseTime, one of the previous "Four Little Dragons of AI," has not made a profit since its establishment in 2018. After its loss - making listing, its market value has evaporated by 80% from its peak. If the "Six Little Dragons" cannot prove a sustainable profit path, going public will be like drinking poison to quench thirst.

03 The "First Stock" Is Not Just for Money

In fact, the "Six Little Dragons of AI" have gradually disappeared from the public eye, replaced by the "Six Little Dragons of Hangzhou." After all, the capital market is never short of hot spots. Therefore, the "Six Little Dragons" have launched their listing plans this year, with multiple considerations.

Firstly, raising funds through listing is the most important consideration. This is not only because the companies themselves have an urgent need for capital but also due to the pressure from investors.

Since the beginning of this year, artificial intelligence has once again become the focus of the investment market. Coupled with the further relaxation of the listing standards for technology stocks in China, it is a good time for the early investors of the "Six Little Dragons" to achieve returns through listing.

Secondly, becoming the "first stock in the large model industry" can help AI startups tell "new stories" and extend their survival time and space. Moreover, as the valuation of OpenAI has exceeded $300 billion, the Chinese large model industry needs a domestic benchmark to anchor market expectations.

Zhang Yaqin, a foreign academician of the Chinese Academy of Engineering, once said that only 10 companies will play important roles in the AI large model track in the end. Li Kaifu, the founder of Lingyi Wanwu, holds a more radical view, believing that only DeepSeek, Alibaba, and ByteDance may remain as the main large model companies in the Chinese market.

Anyway, it is a consensus in the industry that the final victory in the large model industry will only belong to a few players. Therefore, striving to become the "first stock in the large model industry" is just a ticket for them to enter the elimination round.

Li Kaifu once said in an interview that he believes 2025 will be the elimination year for the commercialization of large model companies. If a company cannot make its technology truly applicable in the market, it may face elimination. Each model has its own advantages, and the core pain point is not that one model outperforms others by 1% in performance but whether it can be truly applied.

So, the future of the large model industry comes back to the original question: What kind of products do users need? At present, the technology in the large model industry is still evolving, the application scenarios are still expanding, and the exploration of commercialization seems to have no end.

In 2025, the focus of the artificial intelligence industry has shifted from basic large models to embodied intelligence and Agent applications. Zhipu, MiniMax, Dark Side of the Moon, Step Star, etc. have also followed suit. After Zhipu, Dark Side of the Moon and MiniMax have both launched general Agent products.

It is not difficult to see that no company dares to relax on the path of large model technology iteration. After all, in this technological race, the outcome often lies in the unknown variables of the next algorithm iteration.

When everyone is going all - in, no one can guarantee whether the next card will be an ace or a joker. Therefore, for both the temporarily leading players and the chasing ones, anything is still possible.

So, going public is just the first step. If large model companies want to grab a larger share of the market, they need to spend their money wisely and face the efficiency battle in technology, application, and scenarios.

This article is from the WeChat official account "Bohu Finance", written by Kaikai, and is published by 36Kr with permission.