DeepSeek: Making Sense of the Reaction—and Overreaction

Michael Froman is president of the Council on Foreign Relations.

This week, Silicon Valley, Wall Street, and Washington were all fixated on one thing: DeepSeek. Earlier this month, the Chinese artificial intelligence (AI) company debuted a free chatbot app that stunned many researchers and investors. While there is a lot of uncertainty around some of DeepSeek’s assertions, its latest model’s performance rivals that of ChatGPT, and yet it appears to have been developed for a fraction of the cost.

Using creative methods to increase efficiency, DeepSeek’s developers seemingly figured out how to train their models with far less computing power than other large language models. As a result, they say, they were able to rely more on less sophisticated chips in lieu of more advanced ones made by Nvidia and subject to export controls.

On Monday, American tech stocks tumbled as investors reacted to the breakthrough. (Prices recovered partially later in the week.) If a Chinese upstart mostly using less advanced semiconductors was able to mimic the capabilities of the Silicon Valley giants, the markets feared, then not only was Nvidia overvalued, but so was the entire American AI industry.

Some also argued that DeepSeek’s ability to train its model without access to the best American chips suggests that U.S. export controls are ineffective, or even counterproductive. Many have called the DeepSeek shock a “Sputnik moment” for AI—a wake-up call that should sow doubt about U.S. competitiveness and spur renewed investment to secure America’s lead. Others view this as an overreaction, arguing that DeepSeek’s claims should not be taken at face value; it may have used more computing power and spent more money than it has professed.

Why was there such a profound reaction to DeepSeek? And what does it mean for U.S.-Chinese competition? To make sense of this week’s commotion, I asked several of CFR’s fellows to weigh in. 

Sebastian Elbaum, Technologist-in-Residence, explained how DeepSeek was able to match the performance of other AI models while incurring far lower training costs: 

Paradoxically, some of DeepSeek’s impressive gains were likely driven by the limited resources available to the Chinese engineers, who did not have access to the most powerful Nvidia hardware for training. This constraint led them to develop a series of clever optimizations in model architecture, training procedures, and hardware management.

Two optimizations stand out. First is the low-level programming of hardware to address bandwidth limitations. (Using the latest Nvidia hardware would have been easier, but they did not have access to it.) Second is the use of “reinforcement learning,” but without human intervention, allowing the model to improve itself.

Kat Duffy, Senior Fellow for Digital and Cyberspace Policy, argued that these innovative methods highlight a downside to the United States’ approach to AI: 

The focus in the American innovation environment on developing artificial general intelligence and building larger and larger models is not aligned with the needs of most countries around the world. For them, the greatest interest is in seizing the potential of functional AI as quickly as possible. The existing chips and open models can go a long way to achieving that.

The more the United States pushes Chinese developers to build within a highly constrained environment, the more it risks positioning China as the global leader in developing cost-effective, energy-saving approaches to AI. These will be far more compelling to many governments and entrepreneurs than the “compute or bust” mindset that has been driving AI investments and innovation priorities in the United States.  

Sebastian Mallaby, Senior Fellow for International Economics, noted the enthusiasm among AI researchers: 

At a dinner on Monday with machine learning scientists, most of whom were either in academia or at AI startups, the DeepSeek model elicited excitement. Academics hoped that the efficiency of DeepSeek's model would put them back in the game: for the past couple of years, they have had plenty of ideas about new approaches to AI models, but no money with which to test them. 

There was also excitement about the way that DeepSeek’s model trained on reasoning problems that were themselves model-generated. Hitherto, a lack of good training material has been a perceived bottleneck to progress. That constraint now may have been solved.

Adam Segal, the Chair in Emerging Technologies and National Security and Director of the Digital and Cyberspace Policy Program, explained that the Chinese are even more excited:

Nationalist pride about DeepSeek is quite high in China. The China Daily, for example, trumpeted, “For a large Chinese model, being able to surpass the U.S. ChatGPT is a historic moment.” A number of prominent tech executives have also praised the company as a symbol of Chinese creativity and innovation in the face of U.S. export controls. But no one is saying the competition is anywhere finished, and there remain long-term concerns about what access to chips and computing power will mean for China’s tech trajectory.

Is Beijing’s newfound confidence justified? Michael C. Horowitz, Senior Fellow for Technology and Innovation, argued that DeepSeek’s accomplishment shows that America’s advantage in AI is hardly guaranteed: 

While U.S. companies remain in the lead compared to their Chinese counterparts, based on what we know now, DeepSeek’s ability to build on existing models, including open-source models and outputs from closed models like those of OpenAI, illustrates that first-mover advantages for this generation of AI models may be limited.

As a general-purpose technology with strong economic incentives for development around the world, it’s not surprising that there is intense competition over leadership in AI, or that Chinese AI companies are attempting to innovate to get around limits to their access to chips. After all, export controls are not a panacea; they generally just buy you time to extend technology leadership through investment. 

How vulnerable are U.S. export controls? Rush Doshi, Senior Fellow for Asia Studies and Director of the China Strategy Initiative, laid out the competing answers to this question: 

There are two schools of thought. The first is the downplayers, those who say DeepSeek relied on a covert supply of advanced graphics processing units (GPUs) that it cannot publicly acknowledge. This camp argues that export controls had, and will continue to have, an impact because future applications will need more computing power.

The second group is the hypers, who argue DeepSeek’s model was technically innovative and that its accomplishment shows the ability to cope with scarce computing power. They argue that U.S. export controls are less effective than it seems because DeepSeek either acquired GPUs despite those controls or innovated around them (or likely both). In this view, AI is a commodity without a moat, so export controls are a mistake.

All that said, there’s a lot we still don’t know. DeepSeek’s innovations are important, but they almost certainly benefited from loopholes in enforcement that in theory could be closed. It's premature to say that U.S. export controls should be reversed or weakened. Once all the facts are in, one might instead conclude that they should be strengthened.

The DeepSeek challenge is not a zero-sum race but a test of systemic resilience. By weaponizing openness responsibly, hardening IP moats, and aligning global AI adoption with democratic values, the U.S. can transform this moment into a durable advantage. Critically, this strategy avoids knee-jerk protectionism; instead, it combines market-driven innovation with targeted safeguards to ensure America remains the architect of the AI age.

How good is the company’s latest model? Judge for yourself. The paragraph above wasn’t my writing; it was DeepSeek’s. The future is coming fast.