Does China's DeepSeek Mean U.S. AI Is Sunk? | Opinion

Last week, the Chinese startup DeepSeek sent shockwaves through the global technology community when it unveiled a powerful new open-source AI system. The model, known as R1, reportedly matched the performance of leading models from U.S. companies like OpenAI, even though it was built at a fraction of the cost. In the days since its release, R1 has shaken global financial markets and left AI experts scrambling to understand how a relatively unknown Chinese startup could achieve such a breakthrough.

While new AI models often generate buzz, R1's release is especially notable in light of the U.S. government's recent efforts to curtail China's AI ambitions. Since 2022, Washington has sought to slow China's AI development by restricting the country's access to the advanced computer chips needed to build and run large AI systems. With R1, DeepSeek showed that, at least for now, those measures have not entirely suppressed China's capacity for AI innovation.

Crucially, the R1 release also underscored a major difference in how China and the United States are approaching AI development—and their strategies for dominating the global AI market in the years ahead.

One of the key ingredients for building AI systems is "computing power," or the hardware, energy, and other computational resources that developers use to train models. In recent years, developers have generally improved their models by increasing the amount of computing power they use.

This dynamic has driven U.S. tech companies to spend hundreds of billions of dollars on new computing infrastructure, with the hopes of increasing the power and performance of their models. Just last week, President Donald Trump announced a plan to build up to $500 billion of new computing infrastructure, led by OpenAI, Softbank, and Oracle.

The federal government's strategy for knee-capping China's AI sector is also based on the importance of computing infrastructure. By cutting off their access to advanced semiconductors, the thinking goes, it will be harder for Chinese developers to achieve AI breakthroughs.

DeepSeek's achievement revealed the potential shortcomings of this strategy—though U.S. policymakers can potentially restrict China's access to computing hardware, they cannot control the inventiveness of Chinese AI developers. Rather than pouring more resources into the training process, as has generally been the strategy of U.S. companies, DeepSeek pivoted toward developing a more efficient model. In the end, they succeeded in creating an AI system that delivers more bang-for-the-buck than its leading American counterparts. While U.S. firms have themselves made progress on building more efficient AI models, the relative scarcity of advanced chips gives Chinese developers like DeepSeek a greater incentive to pursue such approaches.

While much remains unclear about DeepSeek's long-term commercial prospects, we can draw three key takeaways from the company's initial success.

First, it reveals that massive investments in AI infrastructure might not be the only, or even most viable, strategy for achieving AI dominance. An approach that combines compute buildout with a greater focus on algorithmic innovation may be the more cost effective and efficient path forward, especially for second movers. That said, compute will remain a vital component of the AI development process going forward—DeepSeek CEO Liang Wenfeng has noted his company's lack of high-performance chips has been its most significant challenge.

Second, China appears to be closer to the United States in AI capabilities than some experts thought, despite extensive U.S. efforts to constrain the Chinese AI sector's access to cutting-edge semiconductors. The U.S. lead appears to be more on the order of months than years.

Finally, R1's release underscores China's potential advantages in global AI competition: the model is free to use, cheap to run, and open-source, all of which could make it more appealing to resource-constrained users, particularly in the Global South. If Chinese semiconductor manufacturers succeed in building out its inference chip offerings, Chinese models may become more widely used in other parts of the world.

Going forward, U.S.-China AI competition could evolve in various ways. China could keep pushing the envelope in algorithms, potentially decreasing its reliance on compute-intensive AI paradigms. Alternatively, U.S. controls on cutting-edge chips may ultimately constrain China's ability to scale AI systems, even if theirs are more efficient. Some analysts argue that the Biden administration's chip controls will start to bite more in the coming years, which could help U.S. AI developers once again open up a commanding lead on their Chinese rivals.

Either way, it is clear that DeepSeek has ruffled both Wall Street and Silicon Valley, who are now beginning to recognize that going all-in on computing infrastructure may not provide the unassailable advantage they crave.

Jack Corrigan is a senior research analyst at Georgetown University's Center for Security and Emerging Technology (CSET).

Sam Bresnick is a research fellow at Georgetown University's Center for Security and Emerging Technology (CSET).

The views expressed in this article are the writers' own.