DeepSeek’s sudden emergence has taken the West by surprise, triggered sudden fears that Silicon Valley is about to lose its global dominance, and put big tech stocks on course for a US$1 trillion wipeout amid a market panic.
US venture capitalist Marc Andreessen described the company’s latest model as “AI’s Sputnik moment” – a reference to the Cold War-era crisis caused by the Soviet Union launching the world’s first satellite, when America thought it was far ahead technologically.
Jerome Pesenti, a former head of AI research at Meta, said that DeepSeek was “almost on a par” with the top model from OpenAI, the maker of ChatGPT.
More alarming still, the company claims to have made its chatbot using elderly microchips and with just US$6 million of investment. If true, it threatens to destroy a business model built on the idea that AI is only possible with ultra-high-end hardware and hundreds of millions of dollars.
What is more, the public is convinced. DeepSeek’s smartphone app has shot to the top of the iPhone App Store by downloads, overtaking ChatGPT.
It is a remarkable success story for a company that was virtually unknown outside of the rarefied world of AI research until last week. Even there, DeepSeek was something of a mystery.
The company’s journey to AI dominance has been unorthodox. Unlike America’s labs, founded with the aim of developing a superintelligence (known as Artificial General Intelligence, or AGI) or Silicon Valley giants seeking to maintain their dominance, DeepSeek was launched as a spin-off to a Chinese investment fund, High-Flyer, launched in 2015.
Hangzhou-based High-Flyer has emerged as one of China’s best-performing “quant” funds, relying on mathematical expertise and computer programming to make investment decisions, under founder Liang Wenfeng, who had studied AI and computer science.
With roughly US$8 billion under management, Wenfeng, 40, spun up an AI team within his fund to work on AGI. His move was driven by “curiosity,” he told Chinese news website 36Kr, “curiosity about the boundaries of AI capabilities”. He said its AI lab was not launched as a money-making venture.
“Many people would think there is an unknown business logic behind this,” he said, “but in fact, it is mainly driven by curiosity”.
DeepSeek and High-Flyer are thought to only have a few hundred employees, compared to more than 3000 at OpenAI and hundreds of thousands across Google, Meta and Microsoft. Wenfeng was last week invited to meet the Chinese premier, Li Qiang, to brief him on its progress.
Aside from its strict censorship on Chinese politics, DeepSeek functions similarly to other “language models” such as ChatGPT, Google’s Gemini and Meta’s Llama. Ask it to write an essay, provide a recipe, write computer code or complete any of the other tasks that millions have come to rely on AI for, and it will oblige.
The difference is a stunning technical efficiency. AI giants are spending hundreds of billions of dollars a year on the high-powered computer chips used to “train” and then operate its most advanced systems.
DeepSeek says that its v3 model, released on Boxing Day, cost less than US$6m to train, less than a tenth of what Meta spent on its most recent system, and used underpowered Nvidia chips designed for the Chinese market. Unlike companies such as OpenAI, its systems are “open source”, meaning they can be downloaded for free by anybody.
Last week, it launched a new system, R1, which supposedly makes new breakthroughs in “reasoning”, the theory that AI tools can think critically and make nuanced decisions, rather than just regurgitating language it has harvested from the web.
Writing in his newsletter, Jack Clark, the co-founder of US AI lab Anthropic, said DeepSeek had “the best AI research team in China on a per-capita basis” and “the sort of talent and output that looks in-distribution with major AI developers like OpenAI and Anthropic”.
Some industry experts have questioned DeepSeek’s story, suggesting that the company may in fact have had Beijing’s help to source more powerful chips, but the claims have been enough to send shockwaves through the industry.
If a moderately resourced Chinese start-up with a fraction of Silicon Valley’s resources can go head-to-head with Western giants, it might collapse the argument that more money and more power will be the key to superhuman intelligence.
Initiatives such as Stargate, a US$500b OpenAI project championed by Donald Trump last week, might start to look obsolete, or at least less crucial.
DeepSeek has already caused alarm at Meta, whose boss Mark Zuckerberg on Friday announced plans to spend up to $65b this year on AI infrastructure. One employee claimed last week that the company’s AI team was in “panic mode”.
The crisis snowballed on Monday as shares in companies such as Nvidia, the Dutch microchip equipment maker ASML and power engineering company Siemens Energy slumped.
“It fundamentally breaks the AI ... narrative, if true,” said Mizuho’s Jordan Rochester, pointing out that vast amounts of spending on infrastructure might not be needed after all.
The surprising success of DeepSeek is likely to be closely scrutinised by Western officials. Both the UK and US have set up dedicated AI safety institutes to track the most advanced models.
A UK government spokesman said: “While we do not provide a running commentary on which models we are testing, the AI Safety Institute [AISI] engages with a range of companies to improve the science and practice of evaluating AI capabilities and will continue to strengthen these relationships.”
Professor Alan Woodward, a cyber security expert at the University of Surrey, says: “Although ChatGPT had first mover advantage, second movers can do things a lot cheaper.
“I suspect the US authorities will see it as a risk. It is a bit like TikTok. It is not necessarily a security risk, but you could bias the information you are giving to people.”
Only a few days ago, TikTok was effectively banned in the US, before it secured a last-minute reprieve from Donald Trump. Congress and US security officials have long worried that the app, which is owned by Chinese headquartered Bytedance, could pose national security concerns.
With DeepSeek, there will be obvious parallels if a popular AI app is designed to regurgitate CCP propaganda. The app’s privacy policy also states much of its data is stored and processed via China.
“We have seen it has been tuned and trained in accordance with Chinese political thought,” says Iain Mackay, the director of AI safety at UK start-up Faculty, which works with the UK Government. He adds he would “very strongly expect AISI to be interested in the capability” of DeepSeek.
The Daily Telegraph reported last year that the UK’s AISI had been contracting for translators fluent in Mandarin. Researchers from the UK institute visited China in 2024 and attended a briefing with the Beijing Institute of AI Safety and Governance.
There will also be cyber security concerns if DeepSeek’s software becomes popular among businesses or governments.
One AI expert muses that Chinese apps present a hacking risk, either deliberate or unintentional. Last year, a security researcher uncovered a bug in DeepSeek that allowed the app to reveal data that could potentially hijack a user’s account. It has since been fixed.
The more fundamental fear, however, will be that China has taken a major step towards usurping the US technological hegemony, in spite of America’s best efforts to crush its AI industry with sanctions.
The announcement “starkly demonstrates that China is very much keeping pace with the US in large-scale AI”, says Professor Michael Woolridge, of the University of Oxford.
With its engineers hamstrung by export controls, they may have been forced to get more creative than their Silicon Valley peers.
“Necessity is the mother of invention,” says Woodward. “The Chinese have asked, ‘How we can do this better, with less?’”
