The global tech industry has been holding its breath: for several weeks, it has been watching for the next breakthrough from DeepSeek, a barometer of Chinese ambitions in the frantic race for artificial intelligence (AI).
At the start of 2025, the startup from Hangzhou (east) reshuffled the AI cards with the release of a “low-cost” conversational agent, capable of competing with Gemini, ChatGPT or Claude.
Since then, and despite a barrage of rumors about an imminent launch, DeepSeek’s new generation “V4” model is still awaited.
The choice of computer chips used to drive and power the new system is at the heart of speculation: are they world-leading American models or “made in China” alternatives that the country is striving to develop at full speed?
“This is important to know because, in a way, it testifies to China’s trajectory towards self-sufficiency in terms of AI”, deciphers for AFP Wei Sun, analyst at Counterpoint Research.
According to the specialist site The Information, the “V4” model can work with the latest chips manufactured by the Chinese giant Huawei.
Such a development would mark a decisive step for China, which still does not have Nvidia’s high-end H200 chip, despite a relaxation of export restrictions by the United States.
In anticipation of DeepSeek’s new launch, large orders for Huawei chips were placed by tech giants such as Alibaba, ByteDance and Tencent, The Information article claims, citing five people with direct knowledge of the orders.
When contacted by AFP, DeepSeek, Huawei, as well as Alibaba, ByteDance and Tencent, all declined to comment.
– Credibility at stake –
Born in 2023, DeepSeek was then just the side project of a speculative fund that had the flair to build up a massive stock of Nvidia processors.
In January 2025, its conversational agent R1, endowed with extraordinary reasoning capacity, shook Wall Street, causing a fall in American technology stocks.
Donald Trump himself saw it as a “wake-up call” for Silicon Valley.
R1 built on DeepSeek’s latest major AI model, V3, launched in December 2024. The company’s affordable and customizable AI tools have been widely adopted in China and are also popular in emerging markets such as Southeast Asia and the Middle East.
Presented as multimodal, that is to say capable of generating texts, images and videos, V4 could once again shake up American technological stocks, Stephen Wu, founder of the Carthage Capital fund, explained to AFP.
“I expect that the next version of DeepSeek V4 will not just be a software update, it will be a very capable open source model”, running “at a much lower cost”, he predicts.
But beyond performance, DeepSeek also plays on its credibility.
Until now, its prowess has relied on the power of Nvidia. Switching to Chinese national chips requires “deep re-engineering”, warns Wei Sun.
– Circumvent the ban? –
The United States cites national security reasons to justify its ban on exporting the most powerful AI processors from giant Nvidia to China.
“The extended wait for DeepSeek V4 highlights the challenges of scaling advanced models without unlimited access to Nvidia’s high-end hardware,” Wu said.
Yet some reports claim that Deepseek managed to circumvent the ban and trained V4 using thousands of Nvidia’s high-end Blackwell chips, dismantled in third countries and then smuggled into China.
Nvidia did not respond to AFP’s requests, but told The Information that “this type of contraband seemed unlikely.”
Another Chinese AI startup, Zhipu, unveiled an image generator in January that it said was trained entirely on Huawei chips.
Leaked data suggested that DeepSeek’s “attempt to train its models on domestic Huawei Ascend chips failed, forcing it to quietly return to restricted Nvidia chips,” Wu said.
But “if they managed to run V4 entirely on Huawei chips, it marks a significant change in the geopolitical landscape of technologies,” he warns.
