In the past period, the industry's discussion about DeepSeek is no longer limited to the changes in the development paradigm of large - models brought by its technological innovation. More importantly, it has driven the extensive participation of China's AI ecosystem single - handedly. From the underlying computing power, cloud platforms and other infrastructure, to the intermediate - layer infra manufacturers, and then to downstream applications, including enterprises in fields such as the Internet, automobiles, intelligent hardware/smart home, finance, education, medicine, and media, from industry giants to startups, DeepSeek's circle of partners continues to expand. According to incomplete statistics, there are currently more than a hundred enterprises cooperating with DeepSeek.
On February 17th, the founder of DeepSeek attended a symposium for private enterprises, which was his second appearance at a high - level national meeting recently. Can the DeepSeek craze continue? What impact will its technological innovation have on the development of AI technology? Will domestic computing power accelerate the emergence of a Chinese - version NVIDIA? Recently, at an AI salon, Li Xingyu, the Chief Ecosystem Officer of Enflame Technology, Zhao Hongbing, the General Manager of the AI Cloud Business Unit of Parallel Technology, Mu Zelin, the CEO of Hao Wen Large - model, Chen Long, the CTO of Zhongke Jiahe, Chen Dazhi, the Managing Partner of Dingxing Quantum, and He Yihao, the Marketing Partner of Qingmao Intelligence, conducted in - depth discussions on these issues.
Unraveling DeepSeek's Breakthrough Code
When it comes to the popularity of DeepSeek, the guests were deeply impressed by its performance in aspects such as reasoning ability, comprehension ability, in - depth thinking ability, the detail and fluency of output, the transparency of the reasoning process, and multi - round dialogue ability. Behind this is DeepSeek's technological innovation in multiple aspects such as training, architecture, and algorithms, through which it forms a cost - effective advantage of low cost and high performance. Li Xingyu emphasized that DeepSeek's innovation lies in engineering. It has not changed in the underlying architecture. Like OpenAI, it is a model of engineering innovation. He believes that incremental engineering innovation conforms to the development rhythm of the technology cycle, and through a relay approach, new technologies can be continuously promoted into the commercialization process.
Mu Zelin also mentioned that the engineering innovation done by DeepSeek can solve the last - mile problem of applications. "This can give the entire Chinese AI industry more lasting vitality, allowing AI to reach applications faster and generate sustainable business models." In Chen Long's view, DeepSeek can significantly compress the training cost after deep accumulation and iterative optimization in the early versions. He also emphasized that engineering innovation is also very important in the computer field, and a large number of engineering practices will give rise to technological progress.
Zhao Hongbing said that DeepSeek may have achieved something subversive - it not only improved the AI ability level but also accelerated the popularization of AI, reaching 100 million users in the shortest time. However, He Yihao believes that in terms of technology and engineering, DeepSeek does have innovation, but it may not reach the so - called subversive level. "Innovation usually includes disruptive innovation and continuous innovation, and DeepSeek is more of continuous innovation." Anyway, DeepSeek's success represents, to a certain extent, the opportunity for China's AI to catch up or even lead.
In Zhao Hongbing's view, DeepSeek's popularity is inseparable from three factors: high talent density, ideals, and sufficient funds. Chen Dazhi believes that two characteristics of DeepSeek may be important factors for its success. One is that the nature of its funds is self - owned, which has higher flexibility and freedom; the other is its open - minded concept of employing people. "This makes DeepSeek not easily replicable." Chen Dazhi believes that if looking for a similar enterprise in the DeepSeek model, it is not certain to find a second one, and even if found, it may not reach the height of DeepSeek.
In Li Xingyu's view, DeepSeek's phenomenal success brings very meaningful inspiration to Chinese entrepreneurs. First, be driven by mission and vision; second, be down - to - earth and adhere to long - termism; third, think against the consensus; fourth, have an open and win - win mindset.
Open - source or Closed - source? DeepSeek Definitely Has More Tricks
DeepSeek's success also owes to its open - source strategy. Except for data, it has made public important indicators such as model codes, parameter weights, and algorithm architectures, and adopted a relatively lenient commercial open - source license. This has also made many companies reflect on or adjust their strategies. For example, some believe that a certain well - known AI company may be on the wrong side of history, and Baidu quickly announced that Wenxin Yiyan would be free and planned to open - source its next - generation model.
Chen Dazhi said from a market perspective that open - source will have more advantages in the future. "What is the ultimate goal of building large - models? It is to attract users. Therefore, open - source has unparalleled advantages." First, let customers use the models, so that more people can participate in improving the ecosystem. Chen Long predicts that open - source and closed - source will coexist. But in terms of social benefits or the degree of public benefit, the level of open - source sharing is higher, which indirectly reduces the overall social cost of repeated development. "Whether it is open - source or closed - source, the core issue is how to form a good business closed - loop, including how to indirectly promote the healthy operation of the entire industrial chain such as computing power." Chen Long said.
From the history of IT development, the coexistence of open - source and closed - source is the mainstream. Li Xingyu believes that the significance of open - source technology is to defeat competitors, while the significance of closed - source in business is to form a complete business model, and they can coexist. For example, Android is open - source while GMS is closed - source. He predicts that in the future, open - source will be a fundamental guarantee, and on the basis of open - source, a large number of value - added business models will be constructed, achieving a win - win effect.
Zhao Hongbing said that both open - source and closed - source have their own advantages and disadvantages, and there is still great uncertainty about whether a certain well - known AI company will open - source again. He Yihao's understanding of open - source is - showing off technological strength and not being afraid of being copied. "DeepSeek dares to open - source, which means it definitely has more tricks. I don't think anyone would show all their weapons at the beginning." He believes that open - source can better promote the development of the entire market and technology, and also stimulate closed - source and technological improvement, which is a healthy market competition.
Is It Meaningless for Giants to Hoard Computing Power? On the Contrary!
DeepSeek's low cost has also raised questions about the Scaling Law in the development of large - models. Is it still sustainable to develop large - model technology by piling up computing power? In Li Xingyu's view, DeepSeek and the Scaling Law are not in a subversive relationship but a complementary one. "The Scaling Law is a bit like the qi - sect in martial arts novels, while DeepSeek is a bit like the sword - sect. Which is more important? Actually, both are important. A great hero in martial arts combines qi and sword skills to reach the peak. In this sense, DeepSeek has opened up a second battlefield for the development of large - model technology."
He believes that the emergence of DeepSeek directly changes the computing power structure and predicts that in 2025, the inference computing power will exceed the training computing power. This does not mean that the training computing power will shrink. Although the marginal effect brought by piling up computing power is decreasing, after unlocking more applications, the training demand will be pulled up again. "However, the growth of inference computing power may be ten - fold or even more exaggerated." Li Xingyu believes that "the rapid growth of computing power will also lead to a decline in unit price, unlocking more applications and thus entering a virtuous cycle. So we are really entering the golden age of computing power and model applications."
Chen Long also believes that the computing power demand will shift from the training side to the inference side, and the computing power demand on the inference side may increase by an order of magnitude compared with the training side. This will promote the requirement for the diversity of computing power and give rise to the development of related industries such as applications, computing power operation, computing power optimization, and computing power integration, driving the industry division of labor to become more and more detailed. Regarding the reason why the training - side computing power will not shrink, Chen Long explained that the capabilities of large - models have not reached the expected ceiling. "I think the more computing power an enterprise can get, the better, and the computing power demand may continue."
In addition, Li Xingyu emphasized that this does not mean that it is meaningless for giants to hoard computing power. On the contrary, DeepSeek has further stimulated the giants' impulse to regain leadership through their computing power advantages. This may well explain why the CEOs of technology giants, including Microsoft, Google, Amazon, and Meta, are shocked by DeepSeek and have all said that they will increase capital investment in AI, data centers, and other infrastructure this year. Robin Li also said that he will not stop investing in AI. To some extent, DeepSeek's success has stimulated the giants' determination to maintain leadership through increased investment. But for many startups, it provides a development model worthy of reference. Li Xingyu mentioned that after DeepSeek made algorithm equality possible, more small players can gain a foothold in the market. Chen Long called on, "We should not artificially impose restrictions on ourselves and think that we are limited and thus not strive for greater and stronger development."
The Emergence of a Chinese - version NVIDIA May Be Accelerated
With the popularity of DeepSeek, more than a dozen domestic chip manufacturers have started to do adaptations. Enflame Technology launched the adaptation of the full - scale DeepSeek model on the second day after the Spring Festival holiday and, together with partners such as Parallel Technology and Zhongke Jiahe, continued to promote the system - level optimization of the DeepSeek model. The highly - anticipated domestic computing power has come into the spotlight. Whether domestic chip manufacturers such as Huawei and Enflame Technology can accelerate their breakthroughs has become the focus of the industry.
When talking about why to adapt to DeepSeek, Li Xingyu explained that this is the first time that domestic computing power has a reason not to follow NVIDIA at the technical level, but to conduct in - depth hardware - software collaborative design following DeepSeek. This gives domestic computing power the confidence to take an independent technological development path instead of completely imitating NVIDIA. "The biggest challenge for domestic computing power is the difficulty in commercial implementation, not the technology." Li Xingyu believes that the gap between domestic computing power and NVIDIA's computing power does not lie in performance but in the fact that domestic computing power has not established a good algorithm ecosystem. This has become the biggest problem in the domestic GPU industry in the past two years.
"The emergence of DeepSeek has greatly promoted the commercialization process of domestic computing power, and the downstream applications have also shown a blow - out development. It can be said that the real spring of domestic computing power has come." Li Xingyu said. "This is the first time that domestic computing power has been widely accepted, and it has unlocked the door for domestic computing power to enter the innovation field." Li Xingyu believes that in the future, more and more innovative companies will use domestic computing power for post - training, for thinking chains, and for various vertical model applications, which is a win - win situation for domestic computing power and domestic model players. He further said that now that the models are transparent, the future optimization path of domestic computing power will be much smoother than before. In this sense, it will definitely shorten the gap with foreign computing power.
Regarding whether there will be a Chinese - version NVIDIA, Chen Long's view is that there is enough data and a large enough market in China. We should learn from the inspiration of NVIDIA's rise, first improve the computing power, and then try to be as open as possible. "The vast data, users, and application markets together can greatly stimulate the industry's enthusiasm for optimizing around domestic computing power, thus enriching and perfecting the entire ecosystem." Mu Zelin said that in terms of training, NVIDIA is still preferred, but there will definitely be a Chinese - version NVIDIA at the edge side. "The difference between domestic and foreign inference chips is not very large, and the Chinese ecosystem provides a great opportunity for domestic enterprises making inference chips." Li Xingyu is more optimistic. "Once becoming the king at the edge side, it will naturally move towards the general - purpose field and finally become the general - purpose king." This is exactly the path that NVIDIA has taken.