The AI landscape has been irrevocably altered in the 100 days following the meteoric rise of DeepSeek R1 on January 20, 2025. This period has witnessed DeepSeek emerging as a dominant force, not only showcasing China’s AI prowess but also shaping the trajectory of global AI development. This report delves into the transformative impact of DeepSeek’s ascent, exploring its influence on AI ventures, investment strategies, and the competitive dynamics within the industry.
A Lifeline for AI Startups
For Zhang Yang, founder of an AI hardware firm, DeepSeek’s emergence was nothing short of a lifeline. On the brink of collapse in late 2024, Zhang Yang’s company, which focused on companion AI hardware, had struggled to secure adequate funding despite his extensive efforts. The demand for their product was limited, and the company was rapidly depleting its resources on research and development.
The launch of DeepSeek R1 in early 2025 proved to be a turning point. Zhang Yang’s company received a significant order from a well-known brand seeking to develop a smart home device for children based on the DeepSeek R1 model. This unexpected opportunity transformed Zhang Yang’s business from the verge of bankruptcy to a state of rapid growth. “I really want to thank Liang Wenfeng,” Zhang Yang quipped, expressing his gratitude.
This transformation is not an isolated case. Li Po, who ventured into AI entrepreneurship after leaving Microsoft in July 2023, shared DeepSeek’s app with his parents during the Lunar New Year. He noted that DeepSeek was the first AI application that truly resonated with the general public.
Li Po’s startup, which was in the seed funding stage last year, found itself facing the stringent requirements typically associated with Pre-A and A round financing. This year, however, his team has been engaged in promising discussions with several leading US dollar funds. For Li Po, who had previously focused on overseas markets, DeepSeek’s international recognition has served as a calling card, demonstrating not only the advanced level of AI technology in China but also the potential of Chinese AI teams to drive down costs and accelerate commercialization.
The Shifting Sands of AI Investment
The 100 days following the release of DeepSeek R1 have witnessed a significant shift in the AI venture capital landscape.
In August 2024, Zhu Xiaohu, Managing Partner at GSR Ventures, predicted that no independent large model companies would survive beyond five years, suggesting that the AI “Little Six Tigers” would ultimately be acquired by larger corporations. However, within 60 days of DeepSeek’s breakout success, Zhu Xiaohu posted on his WeChat feed in March 2025, expressing his willingness to invest in DeepSeek at any valuation if the company were to open up for financing.
Despite this renewed confidence in AI, investment patterns have shifted away from large model development and towards AI applications and embodied intelligence. For example, Huarong Capital invested in Silicon-based Fluidity, whose reasoning acceleration services for DeepSeek were crucial in its widespread adoption. ZhenGe Fund and others have invested heavily in embodied intelligence, with YuShu Technology’s B-round financing of 1 billion yuan setting a new record in the field. Private equity firms havealso begun to explore vertical areas such as “AI + Healthcare” and “AI + Law.” For example, the valuation of Harvey, a legal contract generation tool, doubled in just six months.
DeepSeek’s emphasis on low-cost open source solutions has not only lowered the barrier to entry for AI application development but also intensified competition in the AI industry.
For example, YueZhiAnMian, one of the AI “Little Six Tigers,” reportedly cut 70% of its marketing budget for the Kimi chatbot to focus on model training, attempting to replicate DeepSeek’s technical approach. Zhipu AI, which had previously laid off 20% of its staff in 2024 due to commercialization pressures, announced in March that it would open source its GLM series models, joining the open source movement. It also received over 1.5 billion yuan in funding from state-owned assets in Hangzhou and Zhuhai, seeking to assert its influence as a “national team.”
Li Po’s AI Agent, “WorkfxAI,” focuses on vertical scenarios such as e-commerce, banking, and healthcare. By fine-tuning models with data from these specific domains, it has generated nearly one million yuan in revenue. “The competition is no longer about technical parameters but about who can more quickly and effectively bind models to specific scenarios,” Li Po said.
DeepSeek has also continued to iterate and innovate. On March 25, DeepSeek officially announced a minor version upgrade to the V3 model, and on April 30, it released the DeepSeek-Prover-V2-671B large model, which focuses on mathematical proof. Regardless of whether the effects are positive or negative, the "unintentional" earthquake triggered by DeepSeek will undoubtedly have even more profound and far-reaching consequences.
The Allure of Overseas Funding
“Domestic financing looks overseas. The emergence of DeepSeek has made overseas investors more interested in Chinese AI teams.”
Li Po was one of the earliest members of Microsoft’s Bing search team. From Alibaba to Microsoft, and then to starting his own AI venture in July 2023, he has built teams both in China and abroad. However, last year, he was unable to secure a satisfactory amount of funding.
Before DeepSeek’s breakout success in January 2025, AI entrepreneurs like Li Po faced a depressed investment market.
In Li Po’s field of Agent, securing angel round funding in China required achieving an annualized revenue of $1 million within six months, starting from zero. At the time, there was no clear understanding of AI among either B-end or C-end users in China, which made the requirements seem harsh to startups like Li Po. “When raising angel or seed rounds, investors basically used Pre-A or A round standards. It wasn’t enough to have a lot of users to get funding. They demanded A-round/Pre-A level revenue and payment rates.”
In foreign countries, the keywords associated with Chinese teams were “technologically backward, copycats.” Li Po’s team also failed to gain attention.
After DeepSeek’s explosive success during the Lunar New Year, the wind changed.
Li Po no longer needed to painstakingly explain what AI is to his relatives and friends. DeepSeek’s success made it understandable even to his parents back home. At the same time, DeepSeek’s technological breakthroughs made overseas investors realize that “China’s AI technology has caught up. Although it may not be as innovative as OpenAI in terms of performance, DeepSeek has proven that we can lower prices.” Overseas funds that had previously ignored Li Po have now become more inclined to look at Chinese teams.
Today, Li Po’s seed round financing has reached the final stage, “The goal is to raise $5 million.”
Fang Cheng, a partner at a leading domestic venture capital fund, also told Alpha that DeepSeek’s price reduction will lead to an explosion in the AI application market, and there is a consensus to actively invest in the upstream (such as chips) and downstream (such as applications and hardware) of AI. “DeepSeek’s emergence has indeed led to a convergence of everyone’s perceptions.”
Fang Cheng discovered that DeepSeek’s emergence has not only led large companies to rush to launch deep-thinking large models, but also AI entrepreneurs are learning from DeepSeek.
Li Po’s “WorkfxAI” has also begun to add a thinking demonstration process throughout its product design and various links. DeepSeek’s popularity made him realize that whether it is ToB or ToC, users want more transparency and explainability. Users want to know what the AI is actually doing.
Embracing Transparency and Explainability
Like Li Po, after DeepSeek’s success, Cheng Sen not only replaced a portion of his product’s models with DeepSeek but also modified the product’s internal processes based on the inspiration he gained.
After leaving a leading Internet company to start his own business in September of last year, Cheng Sen launched his own Agent tool in January of this year. When analyzing the reasons for DeepSeek’s success, Cheng Sen realized that DeepSeek’s high generation quality was due to its internal thinking being displayed externally, allowing C-end users to directly issue instructions and constrain the large model in the CoOT (Chain of Thought) process.
Therefore, Cheng Sen also added nodes within the product to incorporate the machine’s thinking process into the next round of output. “The output quality is indeed higher after adding the thinking process.” Although the monthly token consumption cost has increased by 2-3 times compared to before, the number of registered users has also increased from 4,000 in February to over 30,000 today.
However, Fang Cheng admitted that although the team has seen twice as many projects recently compared to November of last year, there is a tacit understanding in the industry to avoid the large model base field where DeepSeek is located, and instead focus on investing in AI applications and embodied intelligence tracks.
“DeepSeek has proven that large models do not necessarily have to be stacked with computing power to improve performance, but it has also raised the barriers to large models again. Now it is difficult for funds other than state-owned assets to take over the follow-up investment in the Little Six Tigers.” Fang Cheng said that even if investors have loosened up their rhetoric towards entrepreneurs, there are not many who are actually taking action.
The Reality of Fundraising in a Competitive Landscape
For entrepreneurs, the current investment market seems like the spring sun, looking bright but unable to transmit heat.
“Compared to last year, there are more investors looking at the Agent direction, but it is more difficult to raise funds.”
Cheng Sen told Alpha that last year he, like Li Po, was blocked by almost harsh financing conditions. The title of a leading Internet company did not smoothly “monetize.” In the end, he chose to take money from individual investors as a start-up fund.
With DeepSeek’s popularity, investors are looking more active than last year, but DeepSeek has further lowered the technical threshold for ordinary developers, and the result is that there are more competitors, and it is more difficult to raise money.
In fact, in the 100 days since DeepSeek’s success, the large model track has become increasingly crowded.
AppGrowing data shows that since February, Kimi’s investment has declined sharply, from over 100 million yuan in monthly investment in December 2024 and January, to 44.25 million yuan in February. As a relatively high-profile technical school among the Six Little Tigers, YueZhiAnMian has also revealed news of a significant reduction in product investment budget.
At the same time, when DeepSeek announced that its theoretical profit margin was 545%, even with extremely low token prices, it could still make a profit. Zhipu, valued at over 20 billion yuan and having just received over 1.5 billion yuan in financing in various places, had sales of 300 million yuan in 2024, according to Phoenix.com, but still had a loss of 2 billion yuan. In front of DeepSeek, which is more open source and has higher visibility, AI entrepreneurs, even the top Six Little Tigers, are facing a fierce elimination round in the commercialization of B-end businesses.
MiniMax has also recently been rumored to have the departure of Wei Wei, the head of the open platform. In response, MiniMax said that the domestic B-end business has ushered in a new stage of development and will be led by other leaders. The commercialization pressure of large model companies brought about by the DeepSeek craze seems to be unavoidable for MiniMax as well.
Even in vertical scenarios, the level of competition is only increasing. What lies in front of AI entrepreneurs is the urgent commercialization problem that needs to be answered.
Zero One Wanwu, founded by Li KaiFu, has stopped the “pre-training” of large language models and instead focused on selling customized artificial intelligence business solutions using DeepSeek’s models. Baichuan Intelligence has turned to the healthcare market.
Cheng Sen chose to develop customized services for smaller B-end businesses, targeting two age groups: the old and the young.
Under the Deepseek craze, Cheng Sen announced on Xiaohongshu that his Agent had access to Deepseek, and the number of likes could reach more than 4,000, which is the most popular post since he started his business. However, as the number of users increased from 4,000 in February to more than 30,000, and the monthly token consumption increased by 2-3 times, Cheng Sen still needs to find a way out for the team.
However, whether it is in the form of team residency support or exploring the charging model of commercial SaaS, Cheng Sen admitted that these are only possible sources of cash, “After all, the domestic commercial SaaS has been developed for many years, and large companies have not yet formed a particularly good model.”
Different from the confusion of looking for investors when starting a business at the beginning of last year, Cheng Sen’s team decided to narrow the user base to two age groups, the old and the young. Although the consumption power of these two groups is questioned, Cheng Sen decided to first do “valuable things, it doesn’t matter if we can’t raise money this year.”
The Unwavering Pressure of Technological Competition
It seems that the waves stirred up by DeepSeek’s pebble are still limited.
For entrepreneurs, what has not changed in the 60-plus days and 100 days since DeepSeek’s breakout success is the pressure brought about by the never-ending technological competition.
Li Po’s WorkfxAI maintains an update frequency that is iterated every 1-2 months. The functional updates of large factories such as Google, OpenAI, ByteDance, Alibaba, DeepSeek, etc., will become a reference for WorkfxAI’s update progress, “Once Gemini and OpenAI are updated, we will update immediately.”
In order to maintain the technical barriers of providing an Agent platform for B-end users, Li Po’s team has to walk on the road of continuous iteration.
And just recently, DeepSeek Vv3 has achieved functional updates, and the R2 model, which focuses on reasoning ability, is also expected to be released before May. In March, OpenAI made major updates to GPT-4o and Sora, and launched a new text-to-image model. Not only does it support continuous questioning, style conversion, and image PPT, but even surpasses Midjourney, the leading platform in the field, in some performance aspects.
A flutter from a technology giant may be a storm for entrepreneurs. “The algorithms and models that took so much time and manpower to fine-tune may be replaced by a large model update once.” Cheng Sen added.
In order to keep up with technological iteration, entrepreneurs need to compete with large companies for people and resources. Although DeepSeek’s popularity has led to strategic adjustments by large companies such as ByteDance, Alibaba, and Tencent, which has resulted in a number of excellent technical talents suddenly appearing on the market, “there are still not enough people. The annual salary packages of good algorithm talents are not something that ordinary start-ups can afford.” Fang Cheng added.
It should be noted that in this competition full of unknowns, the determination of both large companies and entrepreneurs to continue investing in AI will not change, and has even been further strengthened in the 60-plus days since DeepSeek’s popularity.
In addition to keeping up with DeepSeek and seizing spillover dividends, Tencent is not only actively promoting Yuanyuanbao, but also launched its own T1 deep reasoning large model at the right time. At the same time, Alibaba has also recently increased the importance of Quark again, positioning it as Alibaba’s AI flagship application.
According to AppGrowing data, since the end of February, Tongyi Qianwen has gradually increased its investment in advertising, and its investment volume entered the TOP10 after releasing the open source inference model QwQ-32B on March 6. Byte’s Doubao large model department (Seed) also proposed at the all-hands meeting that the long-term application of the model should be done according to the model’s capabilities, and the Doubao inference model will be fully updated.
“Large models are like nuclear weapons, and large companies must have them no matter what. Because if there is a major breakthrough in model capabilities and you don’t keep up, your users will still be snatched away.” Fu Sheng’s judgment on the current AI application competition also echoes the actions of large companies.
Overall, “DeepSeek is more like an inspiring, its emergence has given the market a consensus and made entrepreneurs more excited, but the real practical effect is not obvious. DeepSeek may need to match OpenAI in all aspects of general capabilities. If it can still have such a low price at that time, it will completely lead an AI revolution.” Li Po added.
Whether it is large companies, the AI Six Little Tigers, or even DeepSeek, this battle around base models and AI applications has not yet reached its end. The 100 days of DeepSeek’s popularity may be more like a technical footnote, and Chinese entrepreneurs are starting to ride the wave. Next, there are more unknowns waiting for them.