appConfigEntityId

bodyType

customContents

entityId

entityUuid

isVideo

articlePageQuery

article

articleSchemaContent

content

hooksArticlePageAdvertisementConfigArticle

hooksArticlePageFloatingButtonConfigArticle

appConfig

filter

json

liveArticle

Zhou Xin

DeepSeek, the Hangzhou-based artificial intelligence start-up, may have made another breakthrough that could disrupt the way AI models are trained in China.
Its innovative data format, UE8M0 FP8, could pave the way for home-grown graphics processing units (GPUs) to be deployed in training powerful models even though local GPUs are not as powerful as Nvidia’s.
China’s stock market investors were excited about the breakthrough, pouring money into local GPU developers such as Cambricon Technologies, a Beijing-based start-up with the potential to take on Nvidia in the Chinese market.
The company’s Shanghai-listed shares have doubled in the past month, with a 20 per cent single-day gain last Friday and another 11 per cent gain on Monday following DeepSeek’s statement last week that it had changed the format of data to fit “next generation home-grown chips to be released soon”
The frenzy was akin to January when DeepSeek sent shock waves across the US equity market. The little-known company, whose head office is in a commercial office tower serving banks and investment firms, was regarded as “the biggest black horse” in China’s AI landscape after it unveiled its V3 model in December last year and the R1 model in January. The two models were on par with the world’s best in terms of performance but they were developed at a fraction of the cost.
For its V3 large language model, DeepSeek said it used only 2,048 Nvidia H800 GPUs and that the total training cost was US$5.6 million.

The H800 chips were initially tailor-made for the Chinese market to comply with US export restrictions but were subsequently restricted by Washington in 2023 under tighter US regulations.
The DeepSeek-R1 open-source reasoning model has demonstrated capabilities comparable to US start-up OpenAI’s o1 but also with significantly lower training costs. DeepSeek’s published technical papers on R1 did not mention the cost of development.
The news caused Nvidia’s stock price to slide 17 per cent on January 27. But the shock proved to be short-lived and Nvidia has regained the lost ground and more, with its shares 50 per cent higher now than on January 27.
Similarly, it is a leap of faith to believe that a technical change in how data is formatted in machine learning could see the immediate rise of China’s own AI hardware system and the end of US hardware dominance in AI.
Nevertheless, DeepSeek’s incremental progress is still significant as it has created a new trajectory for China to achieve AI supremacy. In conventional wisdom, the training of powerful AI models requires the best AI chips, the supply of which is dominated by Nvidia and at the mercy of the US government. That meant China would always be in a disadvantaged position due to its lack of access to the best chips and the equipment necessary to make advanced chips.
However, as Huawei Technologies founder Ren Zhengfei told the People’s Daily in a recent interview, although China’s home-grown AI chips are at least one generation behind the US, the country could achieve similar computing power and training efficiency by leveraging the synergies between hardware and software. The AI stacking might be clunky, but it would do the job.
DeepSeek’s highlighting of its data format innovation also showed that it is more than a for-profit commercial entity. It is more like a non-profit organisation that pursues a mission more important than profits or market valuation.
While OpenAI founder Sam Altman is under constant pressure to deliver new models and promote new offerings to justify the firm’s valuation, DeepSeek’s founder Liang Wenfeng acts more like a lab leader who is pursuing the goal of efficient and affordable machine learning to achieve artificial general intelligence.
In this sense, even though DeepSeek has disappointed some by not releasing a more powerful R2 reasoning model as many had expected, the firm’s significance for China’s AI development and even the world’s AI future has actually grown.


Chinese man sues AI firm after chatbot’s ‘auspicious’ date suggestion leads to disaster

Trump says he will create ‘AI Force’, name AI tsar

AI: 10 days that changed the course of artificial intelligence

People in China earn between US$7 and US$15,000 by licensing faces to AI content producers

Artificial intelligence (AI)

DeepSeek’s data format innovation for local AI chips fans fresh hopes of more disruption

The stock market frenzy in China last week was akin to January when DeepSeek sent shock waves across the US equity market. Photo: Shutterstock Images

A visitor looks at a DeepSeek poster during the Global Developer Conference in Shanghai, February 22, 2025.  Photo: VCG via Getty Images

Tech war: DeepSeek hints China close to unveiling home-grown ‘next generation’ AI chips

Visitors learn about Huawei’s Ascend AI chip-based CloudMatrix 384 system at the 2025 World AI Conference in Shanghai on July 27, 2025. Photo: Xinhua

Tech war: DeepSeek’s ‘UE8M0 FP8’ innovation seen as boost for China’s AI self-sufficiency

In this photo illustration, the logo of DeepSeek is displayed on a smartphone screen with DeepSeek-V3.1 in the background. Photo: VCG via Getty Images

A DeepSeek sign is seen near its office in Beijing, February 19, 2025. Photo: Reuters

The Nvidia logo is seen in this illustration taken January 27, 2025. Photo: Reuters

Opinion

China Opinion

DeepSeek’s incremental progress is significant as it has created a new trajectory for China to achieve AI supremacy


DeepSeek’s incremental progress is significant as it has created a new trajectory for China to achieve AI supremacy.


DeepSeek

United States

China

Liang Wenfeng

Sam Altman

Huawei Technologies

OpenAI

Nvidia

Ren Zhengfei

Expresses the views, ideas and conclusions of the author. (Eg. Enough of Crazy Rich Asians, think of crazy poor Hongkongers.)


Column

My Take

DeepSeek’s data format innovation for local AI chips fans fresh hopes of more disruption

.css-1c6uqr6{color:inherit;font-weight:inherit;font-size:inherit;font-family:inherit;line-height:inherit;overflow-wrap:break-word;}DeepSeek’s incremental progress is significant as it has created a new trajectory for China to achieve AI supremacy

DeepSeek’s incremental progress is significant as it has created a new trajectory for China to achieve AI supremacy