On 27th January 2025, the stock price of one of the world’s highest-flying companies, Nvidia, fell 17 per cent. The stock ultimately lost almost $600bn in value, a record loss for any publicly traded company in the US.
Nvidia makes powerful chips that power generative AI tools, such as ChatGPT, which have captured public imagination and many billions in investment. The unfathomably complex “models” that allow chatbots to carry on humanlike conversations require thousands of chips and months of training, in which billions of text documents are turned into predictions of what word should follow another. The way to build better generative AIs was to buy more chips from Nvidia and keep coal-fired power plants burning to provide the immense amounts of electricity these systems need.
And then a Chinese company, apparently started as a side project to hedge fund High-Flyer, released a new AI chatbot that appeared to challenge conventional wisdom. DeepSeek’s chatbot performed as well as the top US models on many tasks and quickly became the top downloaded app for iPhone. But what crashed Nvidia’s share price was DeepSeek’s claim that it had trained its model using lower-performance chips than the newest ones Nvidia had produced, with surprisingly little time and money: two months and under $7m, between one-tenth and one-twentieth of what comparable systems have cost to train. Suddenly, it looked like the mathematics of AI had been turned on its head, and that building powerful AI might not require the resources of a small nation.
DeepSeek is a tech pundit’s Rorschach test: your interpretation of it is less a reflection of the ground truth than it is of your own worldview. AI sceptic Ed Zitron sees DeepSeek as proof positive that generative AI rivals Anthropic and OpenAI are run by snake-oil salesmen who are throwing billions of dollars at expensive hardware to make up for the fact that they have no viable business models. The CEO of OpenAI, in turn, raised questions about whether DeepSeek’s numbers can be believed, or whether DeepSeek might have “cheated” by using outputs of ChatGPT to train the new system. (Given that OpenAI is being sued by hundreds of creators for stealing their copyrighted content, these accusations seem a bit rich.) Anthropic’s CEO claims that DeepSeek was able to achieve its performance using chips illegally smuggled into China. DeepSeek denies wrongdoing.
Here’s my take, which is in no way immune from the aforementioned Rorschach effects: the rise of DeepSeek shows the weaknesses of nationalism and isolationism as an economic policy at a moment of high technological complexity.
Whether or not DeepSeek was quite as frugal as initially reported, clearly it built a novel system based on the constraints it faced. Attempting to protect the US’s lead in AI, the Biden administration issued rules preventing the export of advanced AI chips, including the most powerful ones made by Nvidia. DeepSeek said it was using chips it had acquired before the restrictions were put in place, as well as less advanced ones that Nvidia had developed for export to China. These constraints forced DeepSeek to make some clever and efficient engineering decisions.
DeepSeek uses a “mixture of experts” architecture. It is very good at moving computing resources from one part of a system to another, using precise and expensive systems when necessary and cheaper ones when possible. It leverages a technique that’s become popular with the newest US systems, called “chain of thought”. In essence, it asks an AI model to break down the steps in its reasoning, which makes AI systems slow down and deliver more accurate answers.
Notably, DeepSeek is “open”, though that term is fraught in the world of AI. Virtually no AI systems are open in the sense that we can review what data they were trained on, and DeepSeek is no exception. Furthermore, it’s clear that DeepSeek has compliance with Chinese law built into its models—you’re unlikely to get a neutral opinion on Taiwanese independence from the tool. But unlike the most advanced models from Anthropic and OpenAI, you can download DeepSeek to your own (sufficiently powerful) computer and run it yourself, or build new tools based on its capabilities. That’s likely to make DeepSeek a favourite—alongside Meta’s open Llama model—for AI developers to experiment with. As one Nvidia research manager, Jim Fan, declared: “We are living in a timeline where a non-US company is keeping the original mission of OpenAI alive—truly open, frontier research that empowers all.”
The progress that’s been made in AI thus far is built on profound globe-spanning cooperation. The Nvidia chips the US government has been keeping from China are designed primarily in the US, but manufactured by Taiwan Semiconductor Making Company, the world’s leading chipmaker. Manufacturing those chips requires a panoply of raw materials sourced from around the world, including China, which manufactures much of the germanium and gallium needed for chip production. As much as Donald Trump may hope to bring the world back to the late 1800s, when early automobiles could be built within a single city, AI, like most advanced technologies, requires a global supply chain.
And then there’s the software that runs atop these chips. While Google, Meta, OpenAI and Anthropic can claim credit for tech’s latest developments, the foundation these innovations rest on was developed within the world’s universities, a space that’s remarkably transnational and cosmopolitan. The technical development that made generative AIs like ChatGPT possible—reinforcement learning—was recognised with this year’s Turing Prize, sometimes called the Nobel Prize for Computing. The winners were Andrew Barto of the University of Massachusetts Amherst (where I teach) and Richard Sutton of the University of Alberta. Such transnational cooperations are the norm in academic science, where advanced students travel from all over the world to labs where they can learn the latest techniques, while scientific conferences bring participants from all over the globe.
The US has held a special place within the academy for many years. Many of the world’s most prestigious universities in science and technology are here and, until 2018, the US outpaced the rest of the world in scientific publications. China now produces more scholarly output than the US, though the US retains a lead in the most influential and most cited publications. However, that too may be changing.
The Trump administration has recently declared war on the American academy, pulling hundreds of millions of dollars in grants from universities that have earned his ire for alleged antisemitism, or because the grants might advance the disfavoured causes of diversity and equity. A change in how the National Institutes of Health compensates universities for the expenses of running their labs may withdraw additional billions in funding. Already, universities like mine are reducing the number of new students we are admitting and slowing our hiring in fear of further cuts. At the same time, cruel immigration policies are discouraging international students from enrolling at US universities, fearful that their education will be interrupted by a travel ban or by deportation should they decide to speak out about international politics.
It requires one level of ignorance and arrogance to believe that a nation can go it alone in an era of globalised advanced manufacturing. But it requires a truly Trumpian level of stupidity to blow up the US university system and the economic benefits connected to it. DeepSeek, following on the heels of the consumer-friendly AI embedded within TikTok, is an unmistakeable sign that the US does not have a monopoly on technical creativity, and that Chinese AI is profoundly on the rise.
Despite its impressive recent performance, China has a significant barrier to becoming a central node in the world’s academic network. Scholars tend to flee closed societies for open ones. The US’s scientific leadership after the Second World War was due, in part, to an influx of European scholars escaping the Nazis. The most interesting battle over AI may not be between the US and China, but between closed and closing societies, and nations that remain open to free speech and scholarship.
