DeepSeek, a "technological idealist" in China's AI field, is undergoing a radical transformation. This is reflected not only in its upcoming next-generation large-scale model V4, but also in fundamental shifts in its strategic logic, supply chain choices, and capital paths.
After nearly a year of silence, DeepSeek is expected to unveil its "sword" in late April 2026. This is not just a technological iteration, but also a "coming-of-age ceremony" concerning geopolitics, domestic substitution, and commercial survival.
01
From "Technological Surprise Attack" to "Domestic Symbiosis"
Nearly a year and a half has passed since the release of DeepSeek V3, which shocked Silicon Valley. In that moment of glory, DeepSeek achieved extremely high performance at extremely low cost, causing a stir in US tech stocks. However, since then, DeepSeek has entered a long period of "silence." While Zhipu, MiniMax, and even Kimi are releasing version numbers in rapid succession, even monthly, DeepSeek's update frequency appears "Zen-like" and restrained.
But silence often precedes an explosion.
According to multiple media outlets, including CLS, DeepSeek V4 will be officially launched later this month. The biggest difference from previous releases is that this time, DeepSeek has made a highly symbolic decision: a complete "de-Americanization" of its technological foundation.
According to *The Information* and several other tech media outlets, during the development of V4, DeepSeek broke with industry practice of prioritizing cooperation with Nvidia and AMD, refusing to grant these two American chip giants access to pre-release versions. Instead, it partnered with Huawei.
According to current plans, DeepSeek has granted early testing and optimization permissions for its V4 models to domestic hardware suppliers such as Huawei. V4 is planned to be released in two versions: a full version with over one trillion parameters, optimized for advanced inference and complex code tasks, and designed for Huawei's Ascend chip; and a lightweight version with approximately 200 billion parameters, designed for general dialogue and API services, and capable of running on other domestically produced chips. To prepare for this moment, DeepSeek engineers have spent the past few months deeply immersed in Huawei's Ascend ecosystem, migrating the underlying code from NVIDIA's CUDA platform and adapting it to Huawei's CANN framework.
If V4 achieves competitive performance on Huawei chips, it will be the world's first cutting-edge AI model independent of NVIDIA.
It is understood that V4 will run entirely on Huawei's latest Ascend 950PR chip. This chip uses SMIC's 7nm process, and its computing power density at certain precision levels has reached 2.87 times that of NVIDIA's H20. To meet this change, major domestic AI model manufacturers have already ordered hundreds of thousands of Ascend chips from Huawei.
02
From Price War to Ecosystem War
Looking back at DeepSeek's development history, its impact on the domestic AI industry has been disruptive.
In its first release (the V3/R1 era), DeepSeek played the role of both a "price killer" and a "technology democrat." It proved that top-tier models could be trained without spending trillions of dollars. This forced major domestic internet companies to follow suit with price reductions and even open up some model weights, directly ending the early era of "bubble-like pricing" for large-scale AI models in China.
With the release of V4, DeepSeek plays the role of a "verification stone for domestic computing power."
JPMorgan Chase previously analyzed that the new version of DeepSeek brought a "second wave of impact." But this time, the impact wasn't on price, but on the supply chain. If V4 can achieve stable performance on the Ascend chip, it will directly prove that the domestic AI computing power platform has the ability to handle large-scale, highly complex inference tasks.
This is a "decoupling" for domestic AI companies. In the past, the fate of domestic AI companies largely depended on Nvidia's supply list. AI technology media outlet Houchangcun stated: Once DeepSeek proves the feasibility of the "Huawei + Ascend" path, the entire Chinese AI industry will break free from excessive dependence on overseas computing power and begin a true "internal circulation."
03
Financing, Talent, and the IPO Strategy
Technological shifts are often accompanied by adjustments to business models. DeepSeek, once known for its abundant cash reserves and aversion to fundraising, has finally opened its arms to capital.
Why now?
DeepSeek is currently facing both internal and external challenges.
The external challenge lies in the fierce competition. Competitors like Zhipu and MiniMax have already secured hundreds of billions of Hong Kong dollars in market capitalization through IPOs or fundraising, and are rapidly iterating their products to capture market share. If DeepSeek continues its slow, "research institute" approach, it risks falling behind in the application deployment phase.
The internal challenge is the loss of talent. This was the key factor that ultimately led Liang Wenfeng to decide to seek funding. With DeepSeek's explosive growth, core talent has become a prime target for headhunters and competitors. Wang Bingxuan, the core author of the first-generation large language model, went to Tencent; Luo Fuli, a key contributor to the V3 model, jumped ship to Xiaomi; and Guo Daya, a post-95s R1 researcher, joined ByteDance, with rumored transfer fees in the hundreds of millions of yuan.
To retain young, talented individuals born in the 1990s and 1995s, relying solely on quantifiable salaries is insufficient; they need to see the potential for stock options. Fundraising and pursuing an IPO have become essential means to retain core assets.
According to disclosures, DeepSeek is in talks for its first round of external funding, targeting at least $300 million, valuing the company at over $10 billion.
While this valuation lags behind OpenAI ($850 billion) and Anthropic ($800 billion), considering its extremely low operating costs and immense strategic value, it is considered to have significant potential.
Comparison with domestic AI companies already listed or preparing for IPOs:
Zhipu Huazhang: As the first A-share listed AI company, with a market capitalization of HK$397.2 billion;
MiniMax: Market capitalization of HK$269.4 billion;
Kimi (Dark Side of the Moon): Valuation of $18 billion (approximately RMB 120 billion);
Jieyue Xingchen: Valuation of $5-6 billion.
Compared to these "workaholics," DeepSeek's advantages lie in its strong technological moat and brand reputation, but its disadvantages lie in slow product iteration and insufficient ToC commercialization. Following this round of financing, DeepSeek will inevitably shift from a "laissez-faire" approach to "high-frequency competition," accelerating the iteration of derivative versions after the V4 release and utilizing the funds to reserve more computing power (even domestically produced computing power) to support a larger user base.
04
Conclusion
DeepSeek has finally found Huawei, not only a chip supplier, but also an opportunity for the entire domestic technology industry chain to unite against a common enemy. The release of V4 was a display of technological prowess; the start of financing is a rite of passage for its business.
When DeepSeek goes from "not lacking money" to "needing money," and from "technology first" to "commercial success," the real heavyweights have only just taken their seats at the Chinese AI table.
