After the presentation of DeepSeek-R1 on January 20 caused chipmaker Nvidia's stock price to plummet and other tech companies to plummet, some observers immediately declared the event a "Sputnik moment" in the Sino-American race for supremacy in artificial intelligence (AI). While the American AI industry clearly needed a jolt, the event raises some tough questions.
American technology companies are investing huge sums of money in artificial intelligence. According to Goldman Sachs, "megatech firms, corporations and energy companies will spend about $1 trillion on AI-related capital investments in the coming years." However, many observers (including myself) have long questioned the soundness of the direction of AI investment and development in the United States.
All the leading companies have essentially the same strategy (although Meta has somewhat distinguished itself by choosing a partially open source model), which has led the entire industry to rely on a single approach. American technology companies, without exception, are obsessed with scale. Referring to the as-yet-unproven "law of scaling", they believe that increasing the amount of data and computing power in their models is the key to unlocking new, vast possibilities. They even claim that "scaling is all you need".
Until January 20, US companies were reluctant to consider alternatives to their core models, which had been trained on huge datasets to predict the next word in a sequence. Once they had set their priorities, they focused almost exclusively on diffusion models and chatbots, designed to perform human (or human-like) tasks. DeepSeek’s approaches are largely similar, but it appears to have been more active in using methods such as “reinforcement learning,” “expert fusion” (using multiple smaller but more efficient models), refinement (“distillation”), and fine-tuning in the “reasoning pipeline.” This strategy, it is said, allowed it to develop a competing model at a significantly lower cost.
While there is debate about whether DeepSeek told us everything, the event has exposed the problem of “groupthink” in the US AI industry. The blind disregard for alternative, cheaper, and more promising approaches, combined with media hype, is exactly what Simon Johnson and I predicted in our book Power and Progress, written just before the dawn of the era of generative AI. Now the question arises - are there even more dangerous “blind spots” in the US AI industry? For example, are leading US technology companies missing the opportunity to steer their models in a more “humane” direction? I suspect the answer is “yes,” but time will tell.
There's a second, even bigger question: Has China begun to overtake America? If so, does that mean that authoritative top-down governance structures (what James Robinson and I have called "extractive institutions") can be as good, or even better, than bottom-up arrangements at driving innovation?
I tend to believe that top-down control slows down innovation (as James Robinson and I argued in Why Nations Fail ). While DeepSeek’s success seems to challenge this claim, it is not convincing evidence that innovation under extractive institutions can be as powerful or long-lasting as in systems with inclusive institutions. Moreover, DeepSeek draws on years of progress in the US and Europe. All of the underlying methods were originally developed in the US. “Expert Blending” and “Reinforcement Learning” originated in research institutes decades ago, while large US tech companies were the first to implement “model transformers,” “chain reasoning,” and “distillation.”
What DeepSeek has demonstrated is engineering excellence - it has managed to combine the same methods more effectively than American companies have done. However, it remains to be seen whether Chinese firms and research institutes will be able to take the next step and develop their own revolutionary methods, products and approaches.
In addition, DeepSeek is clearly not your typical Chinese AI company. Most other Chinese AI firms develop technologies for the government or with state funding. DeepSeek, however, comes from a hedge fund and has successfully operated under the radar. But will it be able to maintain its creativity and dynamism now that it is in the spotlight? Whatever happens, the success of one company cannot be considered conclusive evidence that China can beat more open societies in innovation.
There’s also the geopolitical question: Does the DeepSeek case mean that US export control measures and other strategies to limit Chinese AI projects have failed? The answer here is not entirely clear. DeepSeek trained its new models (V3 and R1) on older, less powerful chips, but further progress and scaling may require the most powerful chips.
It is clear, however, that the American approach based on a zero-sum game does not work. Such a strategy would make sense only if we were sure that we were getting closer to creating artificial general intelligence - models that could rival humans in any cognitive task - and that whoever developed it first would have a huge geopolitical advantage. By adhering to these assumptions (none of which have been confirmed), the United States has abandoned fruitful cooperation with China in many areas. For example, if one country develops models that improve human productivity or help regulate the energy sector, such innovations would be beneficial to both countries, especially if they were widely applied.
Like its American counterparts, DeepSeek aims to create artificial general intelligence (AGI), so developing a model that is significantly cheaper to train seems like a milestone. However, reducing development costs using already known methods will not magically lead us to the emergence of AGI in the next few years. The question is whether it is possible to create AGI in the near future. Even more controversial is whether it is even desirable.
While we don't yet know all the details of how DeepSeek developed its models, nor do we understand the significance of its apparent success for the future of the AI industry, one thing is clear: because of the Chinese newcomer, the tech sector's obsession with scale is now being challenged, and may even force them to step out of their comfort zones.
The author is a 2024 Nobel Prize laureate in economics; he is a professor of economics at the Massachusetts Institute of Technology (MIT)
Copyright: Project Syndicate, 2025. (translation: NR)
Bonus video:
