Whose job was destroyed by DeepSeek open source?

The open source of DeepSeek is bringing benefits to a wave of mid-tier AI companies and local chip companies.

Yang Fangxian, founder of 53AI, a company that implements large enterprise models, has been several times busier than last year. Just after the Spring Festival, many clients who had only consulted him before came to Yang Fangxian and asked him to build AI applications as soon as possible. "More than half of our existing customers have switched their base models to DeepSeek." It is reported that most of these customers are medium and large private enterprises.

The founder of another company that provides model scene landing services told Huxiu that after DeepSeek was open sourced, several AI application entrepreneur forums became very active, and people generally believed that the spring of AI application entrepreneurship had arrived early. This wave of enthusiasm also blew into the chip circle. People from local chip leading companies told Huxiu that the number of people seeking cooperation with them increased by two times month-on-month after the Spring Festival.

But for the six big model companies, after a year of exploring business models, they have not yet found a mature route by 2025, and open source has dealt them a huge blow.

"The only ones who are doing the same thing as DeepSeek will be hit the hardest, such as the big model 'Six Little Tigers' that was often mentioned in the past year," Feng Bo, partner manager of Changlei Capital, revealed to Huxiu.

Inside a certain company, there has been a heated debate over whether to open source DeepSeek since the first day of its release. The dilemma facing them is: open source may mean the subversion of the entire business model; if it is not open source, the overall momentum will be greatly left behind.

This is just the beginning.

On February 21, DeepSeek officially announced on the overseas social media X that "starting next week, we will open source five code bases one after another to share our small but real progress in a completely transparent way." DeepSeek also named this "Open Source Week" plan. As of press time (February 27), DeepSeek has open sourced four code bases, namely FlashMLA, DeepEP, Deep GEMM and Optimized Parallelism Strategies.

In fact, long before the "Open Source Week", the "open source trend" triggered by DeepSeek had already affected almost all AI companies in the world. Even Baidu, which was once the most staunch closed-source company in China, announced that it would soon open source the Wenxin Big Model.

But at the same time, the wide scope of this open source trend has led to questions about whether this open source trend is beneficial to all companies. Among the upstream and downstream companies in the AI ​​ecosystem, which companies are truly benefiting from it? And which ones are just following the trend?

Relevant industry insiders revealed to Huxiu, "This open source week, including code optimization, did reveal more of their (DeepSeek) innovative algorithms, but it has little to do with most people. The impact is relatively large for base model manufacturers, while other application companies, chip companies, and all-in-one machine companies are mostly beneficiaries."

It is worth mentioning that during the DeepSeek "Open Source Week" plan, DeepSeek personnel also "low-key" attended the closed-door meeting of this year's Global Developers Conference (GDC). The timing of the two events coincides perfectly.

Yang Fangxian said that Liang Wenfeng might be working on something bigger than just AGI. "If we compare the big model to a lower-level operating system, then OpenAI might be 'IOS', and DeepSeek is 'Android'."

01DeepSeek open source benefits all-in-one machines, chips, and application circles

On February 24, DeepSeek’s “Open Source Week” plan was just launched; on the other side of the Atlantic, investors collectively sold off technology stocks.

The next day, as of the close of Tuesday, the market value of the seven major U.S. technology giants (Tesla, Nvidia, etc.) evaporated by about US$1.4 trillion, among which Nvidia's stock price closed down 2.8%.

Although the two cannot be forcibly linked, it is difficult to say that there is no impact. Previously, Huang Renxun said that investors have misread the development of AI, and DeepSeek will accelerate the adoption of artificial intelligence, which does not mean that computing resources are no longer needed. He emphasized that "late training is still important and requires a lot of resources. And reasoning itself is a computationally intensive process."

In this regard, Huxiu learned from relevant industry insiders that "in fact, the computing power required for reasoning is not much, and these domestic chips like Ascend can fully support it. In addition, DeepSeek is not really short of resources. As a research-based company, there are many forces willing to support it, and it does not need to be financed."

From this perspective, the biggest beneficiary of DeepSeek open source is the computing power layer, including all-in-one machines and chip companies.

Feng Bo told Huxiu, "With DeepSeek, domestic chips like Suiyuan, Muxi, Tianshu, Zhixin, Moore Thread, and Biren can also be more involved in training and reasoning business. Therefore, domestic chips are definitely the beneficiaries. The second is the hardware at the basic layer."

He also made an interesting analogy, "If anyone doesn't provide DeepSeek with an all-in-one machine now, it's like not providing a computer for a child to go to college."

As of press time on February 27, the four code libraries that DeepSeek has released during its "Open Source Week" - FlashMLA, DeepEP, Deep GEMM, and Optimized Parallelism Strategies - have improved large model reasoning, MoE models, and FP8 computing performance. In general, through these innovative algorithms and code optimizations, DeepSeek has maximized GPU computing power, allowing small and medium-sized developers to build high-performance models at a low cost.

But in reality, it is difficult for most companies or developers to make DeepSeek usable out of the box, and most of the time, companies at the model service layer are required to do the setup work.

Therefore, in addition to the chip and hardware companies at the bottom level, the upstream is definitely the group that benefits the most, such as the AI ​​middle-layer companies represented by 53AI mentioned above, and the AI ​​application layer.

Yang Fangxian told Huxiu that "from this year on, it has become standard for customers to ask whether to connect to DeepSeek," and many of them come for DeepSeek.

The Open Source Week is just DeepSeek sharing some of the latest code optimization results with everyone, which does not mean that everyone can learn everything. With DeepSeek, this trump card, it may bring a survival crisis to the base model manufacturers.

02DeepSeek wants to be “Android”?

In addition to AGI, Liang Wenfeng may have even greater ambitions.

During the DeepSeek "Open Source Week" plan, DeepSeek-related personnel "low-key" participated in the 2025 Global Developers Conference. Behind the perfect coincidence of the two times, perhaps Liang Wenfeng is still planning something big.

First of all, the "Open Source Week" plan itself seems to be about sharing innovative algorithms and codes, but in fact, long before the "Open Source Week", the open source trend set by DeepSeek had spread to almost all large model companies, forcing some manufacturers to open source. And launching the "Open Source Week" plan at this time point will further promote the open source trend and make the "open source wave" more fierce.

In Yang Fangxian's view, "If the big model is compared to a lower-level operating system, then OpenAI may be 'IOS', and DeepSeek is 'Android'. Because what is more important than technology is the open source ecosystem and collaboration."

Huxiu learned from an upstream AI company that the access to DeepSeek has helped them to get more orders from major customers. Before this, these major customers still had many concerns.

However, Huxiu has learned that DeepSeek's official API service is still supported by DeepSeek's self-built servers. So, one question is, when more companies join, will commercialization become the next big challenge facing DeepSeek?