How tech start-up DeepSeek emerged as unlikely game changer in US-China AI war

DeepSeek sent shock waves through the global tech market ahead of the Lunar New Year — sinking the value of semiconductor giant Nvidia and other large companies driving the artificial intelligence (AI) boom — as the Chinese start-up achieved a feat once-considered impossible by Silicon Valley.

The Hangzhou-based firm over the past several weeks released two powerful new AI models, DeepSeek-V3 and DeepSeek-R1, that were built at a fraction of the cost and computing power that major tech companies muster to build large language models (LLMs) — the technology underpinning generative AI services like ChatGPT.

On social media, the AI community expressed admiration for how DeepSeek's two open-source models either surpassed or matched the performance of rival products across a range of industry benchmark tests, in spite of tightened U.S. restrictions on China's access to advanced semiconductors and related technologies. In a post on X, OpenAI chief executive Sam Altman said: "DeepSeek's R1 is an impressive model, particularly around what they're able to deliver for the price."

The AI industry's buzz over DeepSeek appeared to reach a crescendo on Monday when news spread that its namesake chatbot, integrated with the R1 reasoning model, claimed the top spot among free-to-use apps on Apple's App Stores in the U.S. and China.

Here is what we know so far about the company and the discernible reasons for its success:

A visionary leader

DeepSeek founder and chief executive Liang Wenfeng rarely talks to the public. But in 2023 and 2024, he gave two interviews with Chinese tech media outlet 36Kr, which reported on Liang's vision for LLM training, the future of AI and his personal pursuits.

When he launched DeepSeek in May 2023, Liang said he was focused on artificial general intelligence (AGI) research. He intended to make the results of this research open to all, so that the technology will not be solely in the hands of "a small group of people and businesses." He said he looked for new recruits with a similar passion, as he built DeepSeek's team.

AGI refers to software with humanlike intelligence and the ability to self-teach, performing tasks that it was not necessarily trained for.

Following the launch of DeepSeek-V2 in May 2024, Liang expressed disappointment that Chinese firms were reluctant to conduct their own frontier research. "China has to be a contributor [to global innovation] and not just always get a free ride," Liang said. He pointed out that innovation is driven by curiosity and the desire to create, rather than just business needs.

The DeepSeek team chose to work on the "hardest things" in the industry, he said. This commitment to original research has paid off for the start-up, based on the wide acclaim received by its V3 and R1 models.

Liang was born in 1985 in the city of Zhanjiang, in southern Guangdong province. His father was an elementary school teacher. In 2002, Liang went east to study for his undergraduate and master's degrees in information engineering at Zhejiang University, where he graduated in 2010. Information engineering typically encompasses elements of computer science, electronic engineering, data science, information security, telecommunications networking and AI.

During that period in university, Liang developed his interest in applying machine-learning technology to quantitative trading. In 2015, Liang founded High-Flyer Quant, which uses deep-learning algorithms to run what has become one of the largest quantitative hedge funds in mainland China.

At the end of 2017, nearly all of High-Flyer Quant's operations were running AI-driven models, which led the firm to expand its hardware, software and algorithm teams. By 2019, total assets under the hedge-fund manager surpassed 10 billion yuan ($1.4 billion). In 2021, that total briefly exceeded 100 billion yuan.

Team of 'young geniuses'

DeepSeek's team of young scientists is made up almost exclusively of Chinese nationals from some of the top schools in the country, such as Tsinghua University and Peking University. Although DeepSeek mainly hires fresh graduates from these schools, Liang said the firm also recruits PhD candidates and young AI professionals with only a few years of experience.

The technical paper for the R1 model, for example, showed it was co-authored by about 200 researchers, including 18 core contributors.

A former DeepSeek employee, Luo Fuli, was recently under the national spotlight amid reports that Xiaomi founder, chairman and chief executive Lei Jun offered her an annual compensation package of 10 million yuan. Based on recent reports, Peking University alumnus Luo — dubbed an "AI prodigy" by Chinese media — has not yet accepted Xiaomi's offer.

Funding and computing resources

DeepSeek's total registered capital is just 10 million yuan, according to information from Chinese business tracking platform Tianyancha. But it benefits from the financial resources and tech infrastructure of hedge-fund manager High-Flyer Quant.

In July last year, Liang said High-Flyer Quant's problem is not money, but U.S. restrictions on exports of advanced chips to China. In May 2023, Liang revealed without elaborating that the company had accumulated more than 10,000 graphics processing units from Nvidia.

In a WeChat post in April 2023, High-Flyer Quant said it was shifting its focus to establish an independent research organisation dedicated to "exploring the essence of AGI" and hoped to recruit like-minded team members.

"We will make significant and sustained investments, avoiding mediocrity, and approach the biggest questions with a long-term perspective," the company's said. It later spun off DeepSeek as an independent enterprise.

Liang said DeepSeek has no plans to raise external funding in the short term. That strategy is in stark contrast to what other young Chinese AI enterprises are doing. For example, the country's six major LLM start-ups — Zhipu AI, Moonshot AI, MiniMax, Baichuan AI, StepFun and 01.AI — raised at least 20 billion yuan in 2024 through nearly a dozen financing rounds, according to Chinese media reports.

Management's start-up mentality

Liang said last July that DeepSeek has no limits in terms of deploying computing resources or in talent recruitment because the company has no internal layers or departmental divisions. For a young company, that appears to be an ideal set-up to rapidly pursue or respond to the latest tech developments.

What DeepSeek practices is a "natural" labour division among its employees, according to Liang. "Every individual has his or her unique life path, and own ideas, and there's no need to push them," he said.

Read the full story at SCMP.