DeepSeek's "Coming-of-Age Ceremony": Ditching NVIDIA, Embracing Huawei

Editor: Captain Teemo

DeepSeek, a "tech idealist" in China's AI sector, is undergoing a transformative metamorphosis. This is evident not only in its upcoming next-generation large model V4 but also in its fundamental shifts in strategic logic, supply chain choices, and capital pathways.

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 a "coming-of-age ceremony" concerning geopolitics, domestic substitution, and commercial survival.

01

From "Technological Surprise" to "Domestic Symbiosis"

Nearly a year and a half has passed since DeepSeek V3's release, which shocked Silicon Valley. At its peak, DeepSeek achieved high performance at extremely low costs, once causing volatility in U.S. tech stocks. However, DeepSeek then entered a prolonged "silent period." While companies like Zhipu, MiniMax, and even Kimi were aggressively releasing new versions on a quarterly or even monthly basis, DeepSeek's update frequency appeared "Buddhist" and restrained.

But silence often precedes an eruption.

According to multiple media sources, including Cailian Press, DeepSeek V4 will officially debut later this month. The most significant difference this time is that DeepSeek has made a highly symbolic decision: a fully "de-Americanized" technology foundation.

As confirmed by The Information and multiple tech media outlets, during the development of V4, DeepSeek broke away from the previous industry practice of prioritizing partnerships with NVIDIA and AMD, refusing to provide pre-release access to these two U.S. chip giants. Instead, it turned to Huawei.

Under current plans, DeepSeek has granted early testing and optimization access for the V4 model to domestic hardware suppliers like Huawei. V4 is scheduled to be released in two versions: a full version with over a trillion parameters, optimized for advanced reasoning and complex coding tasks, targeting Huawei's Ascend chip; and a lightweight version with approximately 200 billion parameters, designed for general dialogue and API services, capable of running on other domestic chips. To achieve this, DeepSeek's engineers have spent the past few months delving into Huawei's Ascend ecosystem, migrating and adapting the underlying code from NVIDIA's CUDA platform to Huawei's CANN framework.

If V4 demonstrates competitive performance on Huawei's chips, it will be the world's first cutting-edge AI model not reliant on NVIDIA.

It is reported that V4 will run entirely on Huawei's latest Ascend 950PR chip. Manufactured using SMIC's 7nm process, this chip's computational density at specific precisions has reached 2.87 times that of NVIDIA's H20. To accommodate this shift, major domestic AI large model companies have placed orders with Huawei for hundreds of thousands of Ascend chips.

02

From Price War to Ecosystem War

Reviewing DeepSeek's development history, its impact on China's AI industry has been disruptive.

During its first release (V3/R1 era), DeepSeek played the role of a "price disruptor" and "technology democratizer," proving that top-tier models could be trained without piling on trillions of dollars. This forced domestic internet giants to follow suit with price cuts and even open-source some model weights, directly ending the early "bubble pricing" era of China's AI large models.

Now, with the release of V4, DeepSeek is assuming the role of a "validator for domestic computing power."

JPMorgan Chase has analyzed that DeepSeek's new version brings a "second wave of impact." However, this time, the shockwave lies not in pricing but in the supply chain. If V4 can deliver stable performance on Ascend chips, it will directly prove that domestic AI computing power infrastructure is capable of handling large-scale, high-complexity inference tasks.

This represents a "liberation" for domestic AI companies. Previously, their fates were largely tied to NVIDIA's supply lists. AI tech media outlet Houchangcun stated: Once DeepSeek proves the viability of the "Huawei + Ascend" path, the entire Chinese AI industry will shed its excessive reliance on overseas computing power and embark on a true "internal circulation."

03

Financing, Talent, and IPO Ambitions

Technological shifts often accompany adjustments in business models. DeepSeek, once the "least cash-strapped" and "least financing -averse" company, has finally opened its doors to capital.

Why now?

DeepSeek is currently facing "internal and external challenges."

The external challenge lies in fierce competition. Peers like Zhipu and MiniMax have secured market capitalizations in the hundreds of billions of Hong Kong dollars through listings or financing, iterating products at breakneck speeds to capture market mindshare. If DeepSeek continues its "research institute"-style slow pace, it risks falling behind during the application deployment phase.

The internal challenge is talent attrition. This was the critical blow that prompted Liang Wenfeng to finally decide on financing. As DeepSeek gained prominence, its core talent became prime targets for headhunters and competitors. Wang Bingxuan, the lead author of the first-generation large language model, joined Tencent; Luo Fuli, a key contributor to the V3 model, moved to Xiaomi; and Guo Daya, a post-95 R1 researcher, joined ByteDance, with transfer fees reportedly in the hundreds of millions.

To retain these young post-90s and post-95s geniuses, salaries from High-Flyer Quant alone are insufficient—they need to see hope for "option realization." Financing and moving toward an IPO have become necessary means to retain core assets.

According to disclosures, DeepSeek is in talks for its first external financing round, targeting at least $300 million with a valuation exceeding $10 billion.

While this valuation lags behind OpenAI ($850 billion) and Anthropic (up to $800 billion), considering DeepSeek's extremely low operating costs and immense strategic value, this valuation is seen as highly imaginative.

Comparing with listed or IPO-preparing domestic AI companies:

Zhipu Huazhang: As the first A-share AI company, with a market cap of HK$397.2 billion;

MiniMax: With a market cap of HK$269.4 billion;

Kimi (Yuezhi Anmian): Valued at $18 billion (approximately RMB 120 billion);

Jieyue Xingchen: Valued at $5-6 billion.

Compared to these "workhorses," DeepSeek's advantage lies in its deep technological moat and brand reputation, but its disadvantages include slow product iteration and insufficient To C commercialization momentum. After this financing, DeepSeek will inevitably shift from "Buddhist updates" to "high-frequency competition," accelerating derivative version iterations post-V4 release and using capital to secure more computing power (even domestic) to support larger user scales.

04

Conclusion

DeepSeek has finally awaited Huawei—not just a Huawei that can supply chips but also a moment when the domestic tech ecosystem stands united. The release of V4 is a technological show of strength; the initiation of financing is a commercial coming-of-age ceremony.

As DeepSeek transitions from "cash-rich" to "capital-needing" and from "tech-first" to "commercial success," the true heavyweight players have just taken their seats at China's AI table.