"During the Spring Festival alone, nearly 100 investment institutions asked people to introduce them to see if there was an opportunity to invest in DeepSeek."
In the face of the phenomenon of DeepSeek, a large model released by domestic AI companies, an angel investor admitted to the Surging news reporter, "We need to think about why projects like DeepSeek were previously missed by us."
DeepSeek technology boom triggered a global shock, the stock prices of many technology giants on the other side of the ocean plunged, and artificial intelligence leader Nvidia fell 4.3 trillion yuan in market value overnight.
"DeepSeek does not have a promotional budget, nor does it have an annual salary of 10 million people, it is a clear goal to invest in research and launch products." The above investor comments. Another large model unicorn told reporters that "DeepSeek founder Liang Wenfeng is a person with AI beliefs, and the technical romanticism he insists on is very well-known in the industry."
"Don't miss the next DeepSeek because you follow DeepSeek, what we need is not a rush to pursue and imitate, the era of artificial intelligence will force humans to return to the source of value." Fudan University School of Computer Science and technology professor, Shanghai Key Laboratory of data science director Xiao Yanghua told reporters.
In his view, behind the rise of DeepSeek is the epitome of China's AI power, and a group of artificial intelligence startups similar to DeepSeek are standing on the world stage.
Why DeepSeek broke through
"DeepSeek was out of my reach last year. Now it's out of my reach." Tao Ru, an algorithm engineer at an AI company in Shanghai, told Surging news reporters with a smile that as an algorithm graduate of a top domestic university, he had received an olive branch from DeepSeek last year, but finally gave up because he was worried that the company was not well-known enough and would not focus on AI.
On social media, many fresh graduates have shown DeepSeek's job invitation, and the words are quite regrettable.
Also "missing" DeepSeek there is a circle of investment institutions, "DeepSeek company executives are clearly not interested in commercialization, just want to do technical research." Investment institutions need to commercialize the company, have revenue profit and capitalization requirements for the company, and also need the founder to surrender a certain degree of equity and freedom." An investor admitted that, in fact, no practitioner of AI investment does not know DeepSeek, and some people extended an olive branch before the explosion.
But the end result is that no VC (venture capital firm) has successfully invested in the company.
Some people who know DeepSeek told reporters that they had interviewed some talents in the field of AI, and finally rejected their own company and went to DeepSeek, on the grounds that they had a good scientific research atmosphere and were a team that really did things.
"In terms of salary package, DeepSeek is only in the middle of the industry, not the highest." He admitted, "The density of talent may not be as good as the head of the big factory, not that the people in the big factory are not smart, but the smart people in the big factory spend too much energy outside of technology." "Big companies pay well, but their internal struggles are fierce, and the desire to focus on technology is not as pure as a technology company like DeepSeek."
"DeepSeek's popularity was due to chance, but more to necessity." Xiao Yanghua told reporters.
"The parent company behind it, Magic Capital, has strong technical strength and computing power foundation in the field of quantitative trading and intelligent finance. At the end of 2022, when OpenAI just launched ChatGPT, there were few domestic clusters of Wanka, except magic Square. In addition, the convergence of a large number of AI-related talents in the financial field also gives DeepSeek a deep talent advantage."
"It's more about the change in mindset." Xiao Yanghua admitted that most AI companies in the past were eager to achieve success, busy with brush list, publicity, realization and capital accounting, while DeepSeek was calm and focused on technical exploration, allowing the team to drive research and development based on curiosity, and was not eager to commercial realization. In terms of environmental factors, Hangzhou, where DeepSeek is located, has an advanced innovation environment. The government has created an atmosphere of tolerance, trial and error and exploration, and only built platforms without interfering with the innovation direction of enterprises, which is very conducive to the development of enterprises.
Tan Jian, associate professor of intelligent interaction design at Beijing University of Posts and Telecommunications, believes that the key changes brought by DeepSeek mean that, as the cost of the model decreases, in the future, high-level AI applications will be promoted by small and medium-sized enterprises, and form a "hundred flowers" situation, and in the short and medium term, cloud computing, edge computing, and server three types of manufacturers. "At present, the three major operators and many Internet computing service platforms have connected to DeepSeek and provided Internet access, and it can be predicted that the revenue of these traditional cloud services and computing platforms will rise steadily as the entire population registers AI services."
DeepSeek is not the only Chinese
DeepSeek's explosion also lets the outside world see that China has formed a number of powerful and influential companies in the large model industry, including ByteDance, Ali, Tencent and other large factories, and there are start-ups such as Moon's dark side, Wisdom spectrum and MiniMax.
On the first day of the New Year after DeepSeek hit the whole network, Ali Yuntongyi team released its flagship model "Qwen2.5-Max", becoming the second Chinese large language model that can match the o1 series of OpenAI company in the United States, which once again caused a shock.
According to the ranking of the third-party platform, "Qwen2.5-Max" ranked 7th in the overall list with 1332 points, surpassing the deep search "DeepSeek-V3" and OpenAI's "o1-mini". In math and programming, "Qwen2.5-Max" ranked first and second in Hard prompts.
The AI unicorn "Dark Side of the Moon" was established in April 2023, and its legal representative Yang Zhilin graduated from Tsinghua University. He earned a doctorate from Carnegie Mellon University in the United States and started his own business in Beijing. According to third-party data, as of January, the company valuation on the dark side of the month has reached $3.3 billion.
MiniMax, a Shanghai-based AI unicorn, was founded in December 2021 with large multi-modal models of text, voice, music, images and video. A highlight worth noting is that MiniMax is at the forefront of the country in terms of AI going to sea. The latest data show that MiniMax's Conch AI overseas version topped the global AI video list in December last year, with more than 27 million monthly visits.
Liu Hua previously told surging news reporters that the United States is in a leading position in AI large model technology as a whole, and in voice, video and other segments, China's large model catch-up speed is fast, such as conch AI and Kuaishou's large model are widely used in the United States, and at present in these fields, the technical level of the two countries has reached a stage of equal.
"In fact, the rate of iteration and evolution of large model technologies in the United States today is actually slower than before." Liu Hua said, "At present, the leading enterprises in the United States have or are building 100,000 card clusters. There are even plans to build million-card clusters to train more advanced models. However, the realization of the million-card cluster faces many challenges, among which whether the local large-scale power facilities support is one of the key factors."
Against this backdrop, Chinese companies are rapidly catching up with their American rivals. DeepSeek is like a "catfish", once again stimulating the vitality of AI competition.
"The large model field has not yet formed an absolute moat, the industry is still in the early stage of development, and there is still a long way to go from the mature stage." Chen Cheng, a veteran observer of the AI industry, told Surging news reporters that he guessed that competition and internal volume in the large model industry would be further intensified after DeepSeek exploded.
"The fierce competition between manufacturers, the most benefit is undoubtedly the user of the large model, that is, the ordinary user, they will enjoy the continuous evolution of the large model capability, the continuous improvement of the cost of the dividend."
A well-known AI company told reporters that DeepSeek is not perfect, v3 model mainly in mathematics and code and other capabilities are more prominent, other general class text generation, understanding and other effects have room to improve, on its particularly low training costs, does not include all costs. (Editor's note: v3's published training cost is about $5.576 million. A third party report points out that the $5.576 million figure mainly refers to the cost of the GPU for model pre-training, and does not include other important costs such as R&D, data collection, and cleaning.)
"After the explosion of DeepSeek, it will undoubtedly stimulate all parties in the industry to further invest in benign technology competition, which is extremely beneficial to the development of the entire industry." At present, in the large model circuit, companies are working hard to achieve better performance, and this dynamic competitive atmosphere is very rare." Another AI startup insider admitted to reporters that "at this stage, who will fall behind and who can stand out are still unknown." In order to survive and develop in the competition, we will definitely invest more resources, and the whole industry is full of vitality and development potential."
How does the DeepSeek effect work
Hu Yanping, chief expert of FutureLabs FutureLab, said that DeepSeek has become an effect, including four aspects, namely, the cost effect of computing power, the user detonating effect, the confidence boosting effect and the open source ecological effect: "Next, there will be a new phenomenon, many slightly strong related enterprises will be based on the base of the large model to do a variety of post-training, distillation fine-tuning, combined with knowledge base, etc., and then to face thousands of industries, forming a large model industry in the AI 2.0 era of the back of the market."
Based on this observation, Hu Yanping believes that the AI industry has three potential directions next: the first direction is the end of the first cycle of AI 2.0 represented by the large language model, and the second cycle represented by multi-modal, embodied intelligence, space-time intelligence, etc.; The second direction is the aftermarket of AI large models, that is, the ecological emergence of the waist and long tail; The third direction is end-to-end AI agents, especially those that can be integrated with workflows and individual needs.
In Xiao Yanghua's view, many AI startups have a good university background, and there is no shortage of talent and funds, but there is a general problem of mental anxiety, too eager, but not conducive to original innovation.
"Companies need a more relaxed development atmosphere and develop steadily according to their own pace and strategic direction." He believes that, in fact, governments around the world are now very concerned about AI companies, but the lack of truly excellent, take out of hand enterprises, "the government's concern should be moderate, after building a good environment and platform, there will be no more intervention." "Caring too much can disrupt the pace of business. It's more important to be a smart carer."
In addition, the emergence of DeepSeek, proved that AI enterprises rely on burning money volume "investment flow" "customer" path is not feasible, the past Chinese AI large model "volume" computing power, "volume" price, "volume" customer, "volume" liquidity ability, now people more recognize the original innovation of long-term, enterprises should think about structural innovation and low-cost research and development, Instead of burning money.
"Big AI models are a very risky investment track with huge amounts of money, and only a few companies will survive, which means that many investments in companies will fail." The vice president of a large model enterprise told the Surging news reporter that in the current environment, the US dollar fund cannot achieve the traditional "investment and financial tube withdrawal" path because of the reason of fund-raising limitation, "the large model industry must accept a reality, and the government will guide the foundation to play a more important role."
From a formal point of view, he suggested that you can refer to the current domestic and international computing power coupons. For example, now the state-owned enterprises can establish a computing power cluster to provide computing power for large model enterprises, and after the investment, most of the investment funds will be returned to the hands of state-owned enterprises in the form of computing power leasing fees.
Xiao Yanghua believes that the wave of entrepreneurship set off by the large model industry means that private enterprises and small and micro enterprises play an important role in the national science and technology innovation system, and will drive more enterprises to flourish in the future. "Enterprises are often the most curious and creative in the start-up and small and micro stage. Curiosity is like a precious spark. It is so precious that the whole society should take good care of it so that the seeds of innovation can continue to take root and grow in the right soil."
Your message must be between 20-3,000 characters!
