DeepSeek’s “Catfish Effect”

Wen| Tidal Business Review

On the eve of the Year of the Snake, DeepSeek shocked the global artificial intelligence field like an atomic bomb.

This proves that it is not feasible for AI companies to rely on burning money, investing, and gaining customers, and highlights the importance of long-term original innovation and low-cost R & D.

This also broke the market monopoly of OpenAI and Nvidia, which were technically closed, and triggered fluctuations in the stock price of global technology giants. Nvidia’s market value evaporated by 4.3 trillion yuan overnight.

Currently, DeepSeek has more than 125 million cumulative users. Like catfish, it stirs up the global artificial intelligence field and restores confidence in domestic AI infrastructure. It is known as a national sport-level existence.

01 New path curve rush

The AI model is an asset-heavy industry. Scaling Law will not expire in 2024. As the model continues to grow larger, training new models requires more data and more computing power.

This means that sufficient financial support is needed. A senior executive of a large factory made a judgment that if the model is large, the annual consumption will cost US$2 billion to US$3 billion.

The emergence of DeepSeek proves that it is not feasible for AI companies to rely on the path of burning money, investing, and acquiring customers. In the past, China’s AI models have computing power, price, and ability to acquire customers, and monetize customers. Nowadays, people more recognize the original innovation of long-term doctrine. Enterprises should think about architectural innovation and low-cost R & D rather than simply burning money.

In the past 2024, Kimi on the Dark Side of the Moon and Doubao on ByteDance have taken turns to become popular AI applications. Behind them are heavy promotion and crazy computing power.

Take Douyin as an example. In 2024, ByteDance began to work on the large model, relying on its previous volcanic cloud infrastructure and talent team accumulation, as well as large-scale capital investment and poaching people everywhere. Among the four Internet giants, ByteDance has invested as much as RMB 80 billion in the AI field, far exceeding Tencent, Ali, and Baidu.

ByteDance’s Volcano Engine supports multi-core, multi-cloud architecture, has ultra-large-scale computing power, supports tens of thousands of card cluster networking, and trillion-parameter MoE large model; provides ultra-high performance networks, supports 3.2Tbps RDMA networks, and has wide global network POP coverage, latency optimization up to 75%. It also has obvious advantages in computing power resources.

The emergence of DeepSeek has promoted the popularization and efficiency of AI technology. Its rise will promote the promotion of a miniaturized model, namely edge AI, and make the hybrid AI model of cloud and edge processing considered the future development direction of AI.

02 Make friends with friends and businessmen”

DeepSeek has solved the dilemma in the domestic AI field, and the implementation of applications has activated the entire industry. Domestic chips, software and hardware companies, and cloud manufacturers have connected to DeepSeek to make friends.

The demand for computing power among large model companies continues to rise, and the demand for cloud services has increased significantly. Dozens of cloud computing service leaders such as Tencent Cloud, Alibaba Cloud, Huawei Cloud, Baidu Smart Cloud, and Volcano Engine have demolished their walls and launched options such as limited time discounts and annual and monthly subscriptions.

In the chip field, more than 10 domestic AI chip companies such as Huawei Shengteng, Muxi Technology, Moore Thread, and Biren Technology have all indicated that they have adapted to the original DeepSeek version or the distilled small size version.

At the same time, domestic mobile phone brands and new energy vehicle brands have also connected to DeepSeek.

In addition, DeepSeek also stimulated Alibaba Cloud Tongyi’s team to release its flagship model Qwen2.5-Max, becoming the second domestic large-language model that can compete with the O1 series of OpenAI Company in the United States, exciting it inside and outside the industry.

DeepSeek’s effect is still continuing and will attract domestic computing power manufacturers represented by Huawei (also including Cambrian, Haiguang, Moore Thread, etc.) to invest more resources to support the adaptation of domestic large models, and the support of government resources will also be increased, which is conducive to the better development of the domestic artificial intelligence industrial chain.

At the same time, competition will also be intensified, and manufacturers with leading advantages will become stronger. A number of startups and Internet companies that originally targeted general models will withdraw from the competition because resources cannot support them, and instead implement artificial intelligence applications in specific directions. This will contribute to a more reasonable distribution of resources in our domestic artificial intelligence industry.

The long thought chain model verified by DeepSeek will help the entire industry chain reduce investment in data inference path labeling. In the future, the implementation process of large models in applications requiring complex reasoning will be further accelerated.

From the launch of DeepSeek to the roll-out of the application side, to investors ‘reunderstanding of the technological potential of China technology companies and revaluation of the value of technology stocks, this transformation is dazzling. DeepSeek’s national games are no exaggeration at all.

03 Domestic large models have a new head

DeepSeek’s popularity has also shown the United States and Europe that China has formed a number of powerful and influential companies in the big model industry, including major manufacturers such as ByteDance, Alibaba, and Tencent, as well as start-ups such as Dark Side of the Moon, Smart Spectrum, and MiniMax.

Objectively speaking, whether it is DeepSeek V3/R1, iFlytek Spark X1, or domestic large models such as Byte Bean Bag, and Ali’s Tongyi Thousand Questions, compared with OpenAI’s o3, there are still gaps in the understanding and application of realistic physical laws., solving ultra-complex scientific problems, and understanding multimodal video stream input and other difficult tasks.

General artificial intelligence is not limited to this. In many aspects such as multimodal large models, embodied intelligence, world models, and environmental simulators, China’s artificial intelligence has huge space to explore, especially in high-process artificial intelligence chips. In terms of mass production and evolution capabilities.

Of course, China’s artificial intelligence will still face some challenges.

DeepSeek’s great achievements based on Nvidia’s computing power will inevitably lead to the continuous escalation of overseas blockages on chip supply, bringing more restrictions to the development of domestic large models.

In particular, there is a high probability that the United States will impose more stringent restrictions on my country’s Nvidia computing power acquisition and domestic computing power overseas projection. H800, A100, H20 and other cards will be completely banned, and the restrictions imposed on DeepSeek will increase day by day.

Therefore, it will be even more critical to achieve independent control of domestic computing power. Fortunately, one company among all domestic large models has achieved results with domestic computing power.

The training and reasoning of the iFlytek Spark model are carried out on domestic computing power, and only 10,000 910B domestic computing power cards are used. IFlytek has not only achieved some results in the research and development of large models with several times less computing power than its friends, but also made adaptation and efficiency optimization of a large number of domestic computing power platforms. It can be said that it is truly a domestic computing power has achieved extreme efficiency, which requires more technical strength and courage than its realization on Nvidia, and is also more strategic.

If iFlytek Spark can succeed, it will, like DeepSeek, give our colleagues more confidence and courage to do bottom-level technological innovation than ever before.

From this perspective, DeepSeek has only opened the curtain for China’s artificial intelligence in 2025, and the real explosion is coming.

Jesse of Tidal’s editorial department has become accustomed to a man-machine conversation after the manuscript is completed in the past two weeks, in order to gain some unexpected ideas. This time, when she typed the question, the dialog box rotated for a long time, and a line popped up on DeepSeek: The server is busy, please try again later.& rdquo;

DeepSeek broke the monopoly with cabbage price technology and condensed the industrial chain with an open source ecosystem, which finally triggered a comprehensive upgrade of domestic AI from technology to business. But all this comes too fast. Cooking oil with fierce fire always gives Deepseek and the entire AI industry some time to adapt.

As DeepSeek founder Liang Wenfeng said in an interview, there will be more and more hard-core innovations in the future. When this society makes hard-core innovators successful, group ideas will change. We just need a bunch of facts and a process.& rdquo;

You see, business is like this.