In the “national competition” of the AI industry, large factories have obtained the qualification for guarantee early, while small and medium-sized enterprises do not even have the opportunity to “participate in the college entrance examination”. Computing power thresholds, data barriers, financing cooling…… Every threshold is quietly raising the cost of innovation. This article is not a complaint, but a calm analysis from a front-line observer: When AI entrepreneurship is no longer a game of “technology + imagination”, how should small and medium-sized enterprises break through? Who is deciding who is eligible to participate in the future?
As soon as the college entrance examination was over, it was busy with a number of big models at home and abroad.
Doubao, DeepSeek, ChatGPT, Yuanbao, Wenxin Yiyan, Tongyi Qianwen…… set off a wave of “AI catching up with the exam” war. It is reported that during last year’s college entrance examination, the results of the large model barely passed the first line, and this year it was collectively promoted to 985. It is reported that Doubao even crossed the admission line in Qingbei.
The head large model is having a lot of fun in the “examination room” of the college entrance examination, but the situation of small and medium-sized AI start-ups is becoming increasingly embarrassing.
In the first half of the year, a total of 13 AI executives from 5 companies in the once glorious large model “Six Little Tigers” changed; Waveform Intelligence has recently been disbanded; Bamboo Intelligence has been suspended for half a year due to high cash flow pressure; Overseas StabilityAI, with a loss of more than $30 million in the first quarter of 2024; ScaleAI is acquired by Meta……
The capital chain is broken, on the verge of bankruptcy, seeking to sell…… It has basically become the common status quo of small and medium-sized players in the global large model industry. According to a set of data last year, from the release of ChatGPT to July 2024, 78,612 newly registered AI companies in China have disappeared, accounting for 8.9% of the total new growth in the same period.
In 2025, with the further comprehensiveness of large factories, there are not many opportunities left for small and medium-sized AI start-ups.
Talent, capital, technology, traffic…… What is left for small and medium-sized AI companies?
As of the end of February 2025, the number of domestic AI native APP users reached 240 million, an increase of 88.9% month-on-month. As of March 2025, DeepSeek, the AI native app industry, topped with 194 million monthly active users, followed by Doubao and Tencent Yuanbao, with 116 million and 41.64 million monthly active users, respectively.
But in fact, in 2024, the top three applications are Doubao, Kimi intelligent assistant, and Wen Xiaoyan. Kimi has fallen particularly significantly, with only 18.2 million monthly active users, which is more than one and a half stars away from the bean bags that used to stand side by side. In fact, in the position change before and after, it is enough to get a glimpse of the situation of small and medium-sized enterprises.
Talent, capital, technology …… The three most indispensable elements of the AI track are passing in small and medium-sized enterprises.
In terms of talents, in the past year, the scarcity of talents in the large model track is comparable to that of chips. According to the data, our country’s artificial intelligence talent gap exceeds 5 million, and the ratio of supply and demand reaches 1:10, of which basic talents are the most scarce. However, AI startups represented by the “Six Little Tigers” continue to experience brain drain.
[fancyadid=“45”]
Incomplete statistics show that at least 22 executives have left the Six Tigers from 2024 to the present. What’s even more interesting is that many of the executives who left have rejoined the arms of big companies:
Several core members of Zero One Wanwu have joined ByteDance one after another; Duan Nan, head of the video generation model of Step Star, joined JD.com; Coincidentally, ScaleAI was acquired by Meta, and the founder joined Meta to be responsible for the AI business; Last year, Lei Jun personally pried DeepSeek’s “corner” and became a hot search……
Why have the talents that small and medium-sized AI start-ups can’t retain enter large factories? This is naturally closely related to “money”.
It is reported that large factories are generous in AI talents, Times Finance reported that the daily salary of intern positions in leading enterprises can be as high as 2,000 yuan, and overseas factories are more generous, it is reported that OpenAI, when poaching people, will give generous stocks, with a total median salary of 925,000 US dollars (about 6.7 million yuan).
In this already burning track of money, small and medium-sized enterprises obviously do not have such a thick family foundation, not to mention that there have been few opportunities for AI enterprises to be “invested” by capital.
According to the data, nearly 30% of the small and medium-sized enterprises of artificial intelligence have received investment, but only 2% of the number of invested enterprises has been “continuously invested”, and there are few enterprises that have received head capital investment. At present, capital’s investment in large models is becoming more and more calm, taking Zhipu, which is still financing in the first half of 2025, as an example, its single financing amount does not exceed 1 billion yuan.
Compared with last year, Zhipu’s cumulative financing amount was close to 7 billion yuan, and in December 2024, its single financing amount was as high as 3 billion yuan.
Without funds, not only can talents not be retained, but also have to temporarily stagnate in terms of technological development. It is reported that up to 72% of IT and finance leaders said that the cost of artificial intelligence is becoming “uncontrollable”, which has gradually led some companies to give up training super large models, such as at the beginning of 2025, Zero One Everything will take the lead in adjusting its team, abandoning super large models, and focusing on applications.
Computing power scale and model performance are no longer the goals of AI startups.
This change has also become the main reason for the departure of technical talents. Not only zero and everything, but about 25% of enterprises’ expenditure on large models came from the special innovation budget a year ago, and today it has dropped to 7%, which means that for most enterprises, AI is no longer an exploratory project.
The double shortage of talents and funds will eventually appear on the technical side one by one. According to the “Industry Express” report, in the artificial intelligence track, only nearly 11% of small and medium-sized enterprises have patent achievements in computer vision and natural language processing technology innovation, and only 18% of enterprises have patent achievements in large model innovation.
At present, the big factory is ambitious.
ByteDance, Tencent, Baidu…… It seems that AI is inevitable. Tencent Yuanbao spent about 300 million yuan on investment within a month, and promoted it in all Tencent ecological products; ByteDance has launched more than a dozen AI applications in one go, and the advertising of non-byte AI products has implemented strict restrictions.
As mentioned in Bilibili’s financial report, as early as the first quarter of 2024, Bilibili’s advertising revenue from the AI industry increased by 5 times year-on-year, and the CPA quotation given by the dark side of the moon on Bilibili was as high as about 30 yuan.
It’s just that now, these 30 yuan has become a huge amount of money that cannot be spent on the “dark side of the moon”, so Yuanbao replaced Kimi to enter the top three AI applications, and in 2025, the rich, powerful, talented, and technological manufacturers will only be more oppressive.
A capital game of “ghosts”?
According to the Zheshang Securities report, ByteDance’s capital expenditure is expected to reach 160 billion yuan in 2025, a full double of last year’s 80 billion. In February this year, Alibaba announced that in the next three years, Alibaba will invest more than 380 billion yuan in building cloud and AI hardware infrastructure, more than the past decade combined.
Under the premise that capital, talent, and traffic cannot compete with large factories, small and medium-sized AI start-ups have crossed the stage of struggling with technology and turned their direction to other subdivisions.
For example, Baichuan Intelligent has abolished the business group responsible for finance and education, betting on the medical AI track; MiniMax uses an open API platform; Step Leap Xingchen reduced and adjusted its AI social tool “Bubbling Duck” business, and cooperated with leading companies such as automobiles and embodied intelligence to develop vertical agents; The dark side of the moon bets on the content community seeks hematopoietic self-help.
It is not difficult to see that from the C-end to the B-end, and the increasingly vertical business is the main direction of a number of AI start-ups.
This is understandable, after all, by 2025, the outlet of the artificial intelligence industry will gradually shift from basic large models to embodied intelligence and Agent application fields. In order to survive smoothly, start-ups need to further concentrate their resources, and at present, the rate of return not only causes trouble for enterprises, but even capital must be weighed before entering.
It is reported that Zhipu’s commercial revenue in the first 11 months of 2024 will increase by more than 100%, and “Financial Magazine” reported that Zhipu’s revenue in 2024 will reach 300 million yuan, but the loss will be as high as 2 billion yuan, that is, its commercial income has not yet been able to cover the loss. In the second half of last year, Sequoia China has re-evaluated its investment in AI large models.
Whether commercialization can be implemented has become the first consideration of capital, small and medium-sized enterprises must live within their means, and there may still be a glimmer of opportunity hidden in the vertical track, which will be more obvious in 2025, and even the world is this trend.
According to data provided by venture capital firm UpHonest, in 2023, the proportion of AI track entrepreneurship projects in YC Startup Camp will be 19%, and the AI track entrepreneurship projects will account for 49%. However, by 2025, the proportion of AI entrepreneurship projects in vertical segments will increase to 40%, and the proportion of AI entrepreneurship projects in the general AI track will only be 26%.
However, what is interesting is that when the six tigers are worried about financing and listing, the domestic large model “top-stream” DeepSeek has always been secretive about capital. Since the beginning of 2025, the financing situation of DeepSeek has been very noisy, but the company has never made a clear response to the relevant financing.
On the evening of February 19, there was market news that DeepSeek is facing an urgent need for computing power expansion due to the explosion and is considering its first external financing. Some time ago, there was news that Alibaba planned to invest in DeepSeek at a valuation of $10 billion, and then the rumors were refuted.
So far, DeepSeek’s financing situation is still a mystery, and there are different opinions in the industry. Many people believe that DeepSeek has the quantitative endorsement of Magic Square and may not be short of money, but this view is a bit “naïve” compared to the speed of burning money in the entire AI track.
It is reported that OpenAI once had most of its revenue to flow back to the “financier” Microsoft cloud service business, and according to the profit sharing agreement disclosed by the two parties, Microsoft can get 75% of OpenAI’s profits until its principal investment is repaid.
In fact, as early as 2023, it was reported that Microsoft had pried away some major customers from OpenAI.
A typical example is Morgan Stanley, a Wall Street giant and one of OpenAI’s “core customers”, which has been reported to be purchasing some OpenAI’s model services through Azure. According to Microsoft’s financial report, up from 18,000 customers purchasing OpenAI software through Azure, up from 11,000 in August.
In 2025, there will be a rift in the relationship between OpenAI and Microsoft, which may be what DeepSeek is wary of.
For AI startups, it is becoming more and more difficult to share food from the mouths of giants, and those who are capable are no longer obsessed with capital protection, but pay more attention to independence.
Will small and medium-sized AI companies still have a “day to come out”?
According to QuestMobile data, the monthly active users of AI native applications will reach 270 million in March 2025, compared with the total monthly active users of China’s mobile Internet of 1.259 billion, the penetration rate is only 21.4%; The average daily usage time of users is only 123 minutes. In other words, AI products have not yet formed a rigid demand in the entire Internet industry, and with the further improvement of penetration, there may still be opportunities for new AI products in the future.
Small and medium-sized AI start-ups would rather bleed than exit the market, and should also be waiting for the next day. According to the Bureau of Labor Statistics, the U.S. AI market is expected to reach $126 billion by 2025, most of which is attributed to small and medium-sized enterprises.
But where is the future of small and medium-sized AI companies? Simply going to the B-end to find value may not be reliable, after all, the price war of large model ToB has been rolled to the extreme.
Taking Zhipu as an example, the price of its model call has been reduced to almost the lowest level in the industry, such as GLM-4-Flash, which is only 0.06 yuan/million tokens. In a statistics in the first half of last year, China Telecom, iFLYTEK, Baidu Cloud, and China Mobile won the most large model projects.
In fact, the prosperity of AI applications has never been able to be separated from the capabilities of the basic model, and at present, the model ability can still directly determine the user side, and even affect the fate of the entire enterprise, and in March, the monthly visits of DeepSeek’s chatbot have exceeded 16.5 million.
Compared with the traffic cost of other models, DeepSeek’s sudden rise is still due to its model capabilities. When small and medium-sized AI start-ups are confused, large manufacturers are still struggling with model capabilities, such as Byte has invested in the whole chain of computing power chip layer, cloud computing layer (volcano engine), model research and development layer, and application layer.
Secondly, in 2025, the entire AI industry will open the “era of agents”, and even several leading financial technology companies such as Ant Digital and Qifu Technology have officially released financial agent products. In the future, agents seem to be a major attraction that determines the large model competition.
As OpenAI CEO Altman publicly said: “We will have better and better models, but I think the next big breakthrough will come from agents.” However, a report released by Capgemini shows that only about 10% of enterprises have already started using AI agents, and 82% plan to integrate agents into their workflows in the next three years.
Small and medium-sized AI start-ups should seize this opportunity.
In addition, multimodal technology is crucial.
Google’s related report predicts that the global multimodal AI market will reach $2.4 billion in 2025 and $98.9 billion by the end of 2037. In the view of industry insiders, a unified multimodal large model can solve all understanding problems, and the main value of multimodal data is that it can improve the upper limit of intelligence.
At present, some of the six little tigers have already set their sights on this link. According to public information, in the 2 years since its establishment, Step Star has released a total of 22 self-developed base models, of which 16 are multi-modal models, accounting for more than 70%, and are known as the “multi-modal volume king” in the industry.
2025 is already halfway through, seeing Doubao, DeepSeek, and ChatGPT go straight into Qingbei, I hope that in 2026, AI start-ups can have the “college entrance examination” qualifications.
