A new Chinese AI model, created by the Hangzhou-based startup DeepSeek, has stunned the American AI industry by outperforming some of OpenAI’s leading models, displacing ChatGPT at the top of the iOS app store, and usurping Meta as the leading purveyor of so-called open source AI tools. All of which has raised a critical question: despite American sanctions on Beijing’s ability to access advanced semiconductors, is China catching up with the U.S. in the global AI race?
At a supposed cost of just $6 million to train, DeepSeek’s new R1 model, released last week, was able to match the performance on several math and reasoning metrics by OpenAI’s o1 model – the outcome of tens of billions of dollars in investment by OpenAI and its patron Microsoft.
The Chinese model is also cheaper for users. Access to its most powerful versions costs some 95% less than OpenAI and its competitors. The upshot: the U.S. tech industry is suddenly faced with a potentially cheaper and more powerful challenger, unnerving investors, who sold off American tech stocks on Monday morning.
Yet not everyone is convinced. Some American AI researchers have cast doubt on DeepSeek’s claims about how much it spent, and how many advanced chips it deployed to create its model.
Few, however, dispute DeepSeek’s stunning capabilities. “Deepseek R1 is AI’s Sputnik moment,” wrote prominent American venture capitalist Marc Andreessen on X, referring to the moment in the Cold War when the Soviet Union managed to put a satellite in orbit ahead of the United States.
So, what is DeepSeek and what could it mean for U.S. tech supremacy?
What is DeepSeek?
DeepSeek was founded less than two years ago by the Chinese hedge fund High Flyer as a research lab dedicated to pursuing Artificial General Intelligence, or AGI. A spate of open source releases in late 2024 put the startup on the map, including the large language model “v3”, which outperformed all of Meta’s open-source LLMs and rivaled OpenAI’s closed-source GPT4-o.
At the time, Liang Wenfeng, the CEO, reportedly said that he had hired young computer science researchers with a pitch to “solve the hardest questions in the world”—critically, without aiming for profits. Early signs were promising: his products were so efficient that DeepSeek’s 2024 releases sparked a price war within the Chinese AI industry, forcing competitors to slash prices.
This year, that price war looks set to reach across the Pacific Ocean.
Yet DeepSeek’s AI looks different from its U.S. competitors in one important way. Despite their high performance on reasoning tests, Deepseek’s models are constrained by China’s restrictive policies regarding criticism of the ruling Chinese Communist Party (CCP). DeepSeek R1 refuses to answer questions about the massacre at Tiananmen Square, Beijing, in 1989, for example. “Sorry, that’s beyond my current scope. Let’s talk about something else,” the model said when queried by TIME.
What DeepSeek’s success could mean for American tech giants
At a moment when Google, Meta, Microsoft, Amazon and dozens of their competitors are preparing to spend further tens of billions of dollars on new AI infrastructure, DeepSeek’s success has raised a troubling question: Could Chinese tech firms potentially match, or even surpass, their technical prowess while spending significantly less?
Meta, which plans to spend $65 billion on AI infrastructure this year, has already set up four “war rooms” to analyze DeepSeek’s models, seeking to find out how the Chinese firm had managed to train a model so cheaply and use the insights to improve its own open source Llama models, tech news site The Information reported over the weekend.
In the financial markets, Nvidia’s stock price dipped more than 15% on Monday morning on fears that fewer AI chips may be necessary to train powerful AI than previously thought. Other American tech stocks were also trading lower.
“While [DeepSeek R1] is good news for users and the global economy, it is bad news for U.S. tech stocks,” says Luca Paolini, chief strategist at Pictet Asset Management. “It may result in a nominal downsizing of capital investment in AI and pressure on margins, at a time when valuation and growth expectations are very stretched.”
But American tech hasn’t lost—at least not yet.
For now, OpenAI’s “o1 Pro” model is still considered the most advanced in the world. The performance of DeepSeek R1, however, does suggest that China is much closer to the frontier of AI than previously thought, and that open-source models have just about caught up to their closed-source counterparts.
Perhaps even more worrying for companies like OpenAI and Google, whose models are closed source, is how much—or rather, how little—DeepSeek is charging consumers to access its most advanced models. OpenAI charges $60 per million “tokens”, or segments of words, outputted by its most advanced model, o1. By contrast DeepSeek charges $2.19 for the same number of tokens from R1—nearly 30 times less.
“It erodes the industrial base, it erodes the margin, it erodes the incentive for further capital investment into western [AI] scaling from private sources,” says Edouard Harris, the chief technology officer of Gladstone AI, an AI firm that works closely with the U.S. government.
… but is Deepseek being transparent?
DeepSeek’s success was all the more explosive because it seemed to call into question the effectiveness of the U.S. government’s strategy to constrain China’s AI ecosystem by restricting the export of powerful chips, or GPUs, to Beijing. If DeepSeek’s claims are accurate, it means China has the ability to create powerful AI models despite those restrictions, underlining the limits of the U.S. strategy.
DeepSeek has claimed it is constrained by access to chips, not cash or talent, saying it trained its models v3 and R1 using just 2,000 second-tier Nvidia chips. “Money has never been the problem for us,” DeepSeek’s CEO, Liang Wenfeng, said in 2024. “Bans on shipments of advanced chips are the problem.” (Current U.S. policy makes it illegal to export to China the most advanced types of AI chips, the likes of which populate U.S. datacenters used by OpenAI and Microsoft.)
But are those claims true? “My understanding is DeepSeek has 50,000 H100s,” Scale AI CEO Alexandr Wang recently told CNBC in Davos, referring to the highest-powered Nvidia GPU chips currently on the market. “They can’t talk about [them], because it is against the export controls that the U.S. has put in place.” (An H100 cluster of that size would cost in the region of billions of dollars.)
In a sign of how seriously the CCP is taking the technology, Liang, Deepseek’s CEO, met with China’s premier Li Qiang in Beijing last Monday. In that meeting, Liang reportedly told Li that DeepSeek needs more chips. “DeepSeek only has access to a few thousand GPUs, and yet they’re pulling this off,” says Jeremie Harris, CEO of Gladstone AI.
“So this raises the obvious question: what happens when they get an allocation from the Chinese Communist Party to proceed at full speed?”
Even though China might have achieved a startling level of AI capability with fewer chips, experts say more computing power will always remain a strategic advantage. On that front, the U.S. remains far ahead. “It’s never a bad thing to have more of it,” says Dean Ball, a research fellow at George Mason University. “No matter how much you have of it, you will always use it.”
Where does this leave America’s tech rivalry with China?
The short answer: from Washington’s perspective, in uncertain waters.
In the closing days of the Biden Administration, outgoing National Security Adviser Jake Sullivan warned that the speed of AI advancement was “the most consequential thing happening in the world right now.” And just days into his new job, President Trump announced a new $500 billion venture, backed by OpenAI and others, to build the infrastructure vital for the creation of “artificial general intelligence”— the next leap forward in AI, with systems advanced enough to make new scientific breakthroughs and reason in ways that have so far remained in the realm of science fiction.
And although questions remain about the future of U.S. chip restrictions on China, Washington’s priorities were apparent in President Trump’s AI executive order, also signed during his first week in office, which declared that “it is the policy of the United States to sustain and enhance America’s global AI dominance in order to promote human flourishing, economic competitiveness, and national security.”
Maintaining this dominance will mean, at least in part, understanding exactly what Chinese tech firms are doing—as well as protecting U.S. intellectual property, experts say.
“There’s a good chance that DeepSeek and many of the other big Chinese companies are being supported by the [Chinese] government, in more than just a monetary way,” says Edouard Harris of Gladstone AI, who also recommended that U.S. AI companies harden their security measures.
Where does AI go from here?
Since December, OpenAI’s new o1 and o3 models have smashed records on advanced reasoning tests designed to be difficult for AI models to pass.
DeepSeek R1 does something similar, and in the process exemplifies what many researchers say is a paradigm shift: instead of scaling the amount of computing power used to train the model, researchers scale the amount of time (and thus, computing power and electricity) the model uses to think about a response to a query before answering. It is this scaling of what researchers call “test-time compute” that distinguishes the new class of “reasoning models,” such as DeepSeek R1 and OpenAI’s o1, from their less sophisticated predecessors. Many AI researchers believe there’s plenty of headroom left before this paradigm hits its limit.
Some AI researchers hailed DeepSeek’s R1 as a breakthrough on the same level as DeepMind’s AlphaZero, a 2017 model that became superhuman at the board games Chess and Go by purely playing against itself and improving, rather than observing any human games.
That’s because R1 wasn’t “pretrained” on human-labeled data in the same way as other leading LLMs.
Instead, DeepSeek’s researchers found a way to allow the model to bootstrap its own reasoning capabilities essentially from scratch.
“Rather than explicitly teaching the model on how to solve a problem, we simply provide it with the right incentives, and it autonomously develops advanced problem-solving strategies,” they claim.
The finding is significant because it suggests that powerful AI capabilities might emerge more rapidly and with less human effort than previously thought, with just the application of more computing power. “DeepSeek R1 is like GPT-1 of this scaling paradigm,” says Ball.
Ultimately, China’s recent AI progress, instead of usurping U.S. strength, might in fact be the beginning of a reordering—a step, in other words, toward a future where, instead of a hegemonic power, there are many competing centers of AI power.
“China will still have their own superintelligence(s) no more than a year later than the US, absent [for example] a war,” wrote Miles Brundage, a former OpenAI policy staffer, on X. “So unless you want (literal) war, you need to have a vision for navigating multipolar AI outcomes.”
Discussion about this post