On Monday, the world watched as $1tn was wiped off the stock market in a single day, a huge bonfire kindled by the little-known Chinese artificial intelligence start-up DeepSeek.
Its release of a new AI model, known as R1, upended assumptions about US supremacy in AI and raised the prospect that some in China are learning how to beat Silicon Valley at its own game.
The model can “reason” to solve complex scientific problems and performs at par with leading-edge software from US tech giants, but was apparently developed at a fraction of the price of those models.
It quickly dislodged OpenAI’s ChatGPT as the most-downloaded free app on the US iOS App store.
Alongside the geopolitical challenge, DeepSeek’s breakthrough has dual implications for the tech industry. Firstly, it is likely to accelerate the commercial development and uptake of AI, much as ChatGPT did in 2022.
At the same time, it threatens to demolish investment assumptions that have underpinned the entire US stock market, by seeming to show that developing advanced AI models does not require vast amounts of infrastructure and thus capital.
The question being asked with sudden urgency from California to Wall Street: Has China caught up in AI at just the moment that many working in the field claim they are on the brink of a historic breakthrough that will put machines on a par with human-level intelligence — a threshold known as artificial general intelligence?
“[DeepSeek’s] algorithmic innovations remind us that China and the US are neck and neck and that our technological edge isn’t guaranteed, pushing our industry to make AI more efficient,” says Eric Schmidt, the former chief executive and chair of Google.
“To get to AGI first, we’ll need to continue to invest in talent, support our vibrant open-source ecosystem, and ensure we out-innovate, not just outspend, our competitors.”
DeepSeek was founded on Silicon Valley-style levels of ambition. It started out in 2023 as a side project for the eccentric hedge fund billionaire Liang Wenfeng, just as the race to replicate ChatGPT was heating up. It has since turned into one of China’s leading AI labs.
“Why is Silicon Valley so innovative? Because they dare to do things,” Liang said in an interview last year. “When ChatGPT came out, the tech community in China lacked confidence in frontier innovation.”
He added: “From investors to [Chinese] big tech, they all thought that the gap was too big and opted to focus on applications instead. But innovation starts with confidence.”
As state-owned funds in China have taken on a larger role funding start-ups in the past few years, the entrepreneurial ecosystem has felt pressured to guarantee returns for fear of losing the country’s assets.
DeepSeek is distinctive among Chinese generative AI start-ups in that it has not raised any external financing and has therefore been free from these constraints.
A pure research lab, echoing the early days of DeepMind in the UK and OpenAI in the US, DeepSeek has focused all its effort on pushing the field of AI forward, rather than trying to make money. And even though it prides itself on being entirely founded on homegrown talent, it has adopted a culture often found in the US tech heartland.
“It’s unique among Chinese AI companies,” says an AI investor in China. “There is no politics or management friction like at the other big tech companies or larger start-ups. People don’t have specific titles or reporting lines.”
DeepSeek’s origins as a quantitative hedge fund meant it had engineering talent with a deep understanding of chips. Its breakthrough turned on its apparent success at training advanced AI models without spending the hundreds of millions of dollars its US rivals have.
It claimed that the final training step for R1 cost only $5.6mn. The figure, however, doesn’t include many other costs involved in developing its models, including computing infrastructure and previous training runs, making it hard to draw precise comparisons.
It may also have cut corners to save costs: OpenAI claims to have evidence that DeepSeek trained on the output from OpenAI’s own models — not something allowed under its terms of use, though an underhand practice thought to be widely used by US companies as well.
Ironically, Washington’s attempt to hamper China’s AI sector by imposing export controls on high-end US chips from 2022 onwards may have contributed to DeepSeek’s breakthrough.
Without access to leading-edge silicon, the company was forced to find innovative ways to squeeze higher performance out of the less sophisticated chips it was able to buy.
The company’s claims about the low cost and advanced capabilities of its models have touched off a heated debate about how disruptive the company will turn out to be. Silicon Valley’s leaders have paid tribute to its innovations, while also playing down their significance.
OpenAI CEO Sam Altman called the R1 model “impressive”, while Mark Zuckerberg, chief executive of Meta, credited the company with making “advances that we will hope to implement in our systems”.
Yet Zuckerberg also cast DeepSeek’s breakthrough as just one among many in a field that is moving at breakneck speed, making it hard to tell how deeply its low-cost approach would change the dynamics of the industry, he said.
According to some China tech watchers, DeepSeek’s advances aren’t significant enough to change the fact that the country’s AI companies have been fast followers, largely devoted to emulating their US counterparts rather than setting the direction themselves.
“DeepSeek’s work falls into that category. What would really turn the tables in US-China competition is if they built something that actually pushed the frontier. We’ll see if they get there,” says Helen Toner, an AI policy analyst at Georgetown’s Center for Security and Emerging Technology, and a former board member at OpenAI.
Yet others say that the fact that a Chinese tech start-up has been behind the latest head-turning advance represents a watershed moment, changing the dynamic in the AI race between the two countries.
“DeepSeek’s latest models may not mean that China is pulling ahead of the US in the AI race, but it does prove that Chinese companies are making remarkable strides in software innovation that mitigate the constraints imposed by US export controls,” wrote Tilly Zhang, a China tech analyst at Gavekal Dragonomics, a China-focused research firm, in a note published this week.
“The race for AI leadership is no longer just about who has access to the best chips, but about who puts them to best use.”
While the significance of DeepSeek’s technical breakthroughs is a matter of debate, there is no mistaking the shockwave that passed through the stock market as investors digested the implications of its main innovation: dramatically cutting the cost of training for the most advanced AI models.
DeepSeek’s models appear to undermine the argument that US AI companies have made vociferously over the past year: that AI advances require vast amounts of capital and infrastructure to develop and deploy their technology at scale.
Instead, they suggest that far more expensive, US-developed models would have little to differentiate them, raising fears among investors of a sharp deflationary shock.
“There was a sense of American exceptionalism — that only America had this technology, and only Americans had the money to do this,” says Jim Tierney, a US growth stock investor at AllianceBernstein. “The commoditisation of these models is happening much faster than we thought.”
Much of Silicon Valley fell back on the argument that others in the industry will quickly copy DeepSeek’s innovations, bringing down the cost of training AI models across the board. Executives like Microsoft’s Satya Nadella claimed it will make the technology more affordable for customers and boost its use — something that would benefit the entire industry.
In a sign of the company’s confidence in its status, DeepSeek has published its research and released its models in “open-weights” form, a more limited version of open-source software that allows anyone to download, use and modify the technology.
The move will attract a wide international following among software developers looking for “open” models to build applications on. Most models developed by Silicon Valley’s leading AI companies remain closed, though there are exceptions — notably Meta, whose open models have surged in popularity.
But Deepseek’s model is accessible at a far lower cost. The Chinese company says it charges only 1.4 cents for each 1mn tokens it generates — roughly equivalent to 700,000 words. By contrast, Meta charges $2.80 for the same output from its largest models.
“A whole number of developers are experimenting with what’s now a Chinese open source AI-based solution,” says Keegan McBride, a researcher at the Oxford Internet Institute who focuses on the geopolitics of AI. “It really shows that in the AI space, the US isn’t the only option on the table.”
While Yann LeCun, chief AI scientist at Meta, described DeepSeek in glowing terms as proof that “open-source models are surpassing proprietary ones”, the start-up nonetheless poses a direct challenge to Meta.
The company’s “claim to fame has been creating open-weight models that aren’t too far behind the bleeding edge, and DeepSeek just beat them at their own game,” says Toner.
Beyond Deepseek’s impact on the market for AI products, its breakthrough also promises to have geopolitical repercussions, coming at what many believe is a pivotal moment in the competition between the US and China for AI supremacy.
If R1 and its successors become the global standard for “open” AI models, it would handicap the US, warned Meta’s Zuckerberg. “For our national advantage, it’s important that it’s an American standard,” he said. “We want to build the AI system that people around the world are using.”
Some content could not load. Check your internet connection or browser settings.
DeepSeek has “accelerated the urgency for people in every country to assess . . . the technological balance of power emerging between various countries,” says Craig Mundie, a Microsoft veteran and former White House adviser, who counsels OpenAI’s Sam Altman on tech policy and strategy.
If China has managed to get on to an equal footing with the US on AI, it has implications for everything that the technology could eventually be used for, warned Dario Amodei, CEO of US AI start-up Anthropic.
“It seems likely that China could direct more talent, capital and focus to military applications of the technology,” Amodei wrote of DeepSeek’s advances. “Combined with its large industrial base and military-strategic advantages, this could help China take a commanding lead on the global stage, not just for AI but for everything.”
Mundie, who also chairs the US-China AI Dialogue diplomatic forum established by the late US secretary of state Henry Kissinger, pointed out that AI was the “ultimate dual use technology”, meaning it has both positive and dangerous purposes.
The emergence of DeepSeek is likely to hang over discussions when the diplomatic group convenes within the next 90 days to discuss a multilateral but common safety structure for AI software, which Mundie writes about in his latest book Genesis, co-authored with Kissinger and Schmidt.
“It doesn’t mean everyone will have the same laws or rules, but [building] the architecture by which these machines grow up understanding human values and comport with societal choices everywhere in the world, I think is an urgent task,” Mundie says.
Meanwhile, aspiring young entrepreneurs in China are looking towards DeepSeek and its founder as inspiration to build a new generation of powerful technology.
A teenager who came to pay respects at Liang’s house in the village of Mililing this week says: “He is a pragmatic technologist. He put together a team that . . . surpassed those of companies like OpenAI that we couldn’t compete with before. He is a great person who has made contributions to China.”
Additional reporting by Melissa Heikkila in London