Sections

Commentary

DeepSeek shows the limits of US export controls on AI chips

January 29, 2025


  • U.S. AI export control rules are designed to impede China’s AI progress, but they may actually be accelerating it.
  • AI engineers in China are innovating on ways to use limited computing resources more efficiently.
  • Overly broad export controls on advanced computing chips can harm U.S. companies by reducing global sales opportunities while doing little to enhance U.S. AI leadership.
Binary code displayed on a laptop screen and DeepSeek logo displayed on a phone screen are seen in this illustration photo taken on January 28, 2025.
Binary code displayed on a laptop screen and DeepSeek logo displayed on a phone screen are seen in this illustration photo taken on January 28, 2025. Jakub Porzycki/NurPhoto

In recent weeks, Chinese artificial intelligence (AI) startup DeepSeek has released a set of open-source large language models (LLMs) that it claims were trained using only a fraction of the computing power needed to train some of the top U.S.-made LLMs. (There has been some suggestion, as yet unsubstantiated, that DeepSeek actually had access to more computing power than it has disclosed.)

Despite the much lower reported development costs, DeepSeek’s LLMs, including DeepSeek-V3 and DeepSeek-R1, appear to exhibit extraordinary performance. Venture capitalist Marc Andreesen posted on X that “Deepseek R1 is one of the most amazing and impressive breakthroughs I’ve ever seen — and as open source, a profound gift to the world,” and “Deepseek R1 is AI’s Sputnik moment.” The news about DeepSeek jolted stock markets as well, with the NASDAQ index dropping over 3% on January 27.

DeepSeek’s models are a stark illustration of why U.S. export controls on advanced computing chips, instead of impeding China’s AI progress, may actually be accelerating it.

Limits of US AI export controls

In the past several years, the Biden administration issued a series of increasingly strict export control rules on advanced computing chips, including a particularly onerous new rule published in the final week before the Trump administration took office. A central goal of these rules is to impede China’s progress on AI.

If DeepSeek’s claims regarding training costs prove to be accurate, the company’s achievements underscore how U.S. chip export control rules can in some respects have the opposite of their intended effect. Forced to operate under a far more constrained computing environment than their U.S. counterparts, AI engineers in China are innovating in ways that their computing-rich American counterparts are not. And in doing so, they are upending the view that has underpinned both the U.S. technology industry’s approach to AI as well as the thinking of U.S. policymakers.

Rethinking the role of massive computing power

The animating assumption in much of the U.S. AI industry has been that creating highly advanced AI models requires access to truly massive amounts of computing power. This is reflected in the investments by companies including Amazon and Meta in multibillion dollar AI computing facilities. Under this paradigm, more computing power is always better. Spending lavishly on computing is viewed as just as important as hiring good engineers.

But having access to extraordinary amounts of computing power has a key downside: It means less pressure to use those resources efficiently. By contrast, faced with relative computing scarcity, engineers at DeepSeek and other Chinese companies know that they won’t be able to simply brute-force their way to top-level AI performance by filling more and more buildings with the most advanced computing chips.

With easy access to unlimited computing power off the table, engineers at DeepSeek directed their energies to new ways to train AI models efficiently, a process they describe in a technical paper posted to arXiv in late December 2024. While DeepSeek is the most visible exponent of this approach, there are sure to be other Chinese AI companies, operating under the same restrictions on access to advanced computing chips, that are also developing novel methods to train high-performance models.

A silver lining

The silver lining to the consternation caused by DeepSeek lies in the opportunity for a more rational approach to export control of advanced computing chips. Even before DeepSeek, attempts by the U.S. to limit access to high-performance computing chips suffered from two key weaknesses. First, there is a robust black market in the trade of controlled computing chips. Second, as it isn’t necessary to physically possess a chip in order to use it for computations, companies in export-restricted jurisdictions can often find ways to access computing resources located elsewhere in the world.

DeepSeek illustrates a third and arguably more fundamental shortcoming in the current U.S. computing chip export control framework: Scarcity fosters innovation. As a direct result of U.S. controls on advanced chips, companies in China are creating new AI training approaches that use computing power very efficiently. When, as will inevitably occur, China also develops the ability to produce its own leading-edge advanced computing chips, it will have a powerful combination of both computing capacity and efficient algorithms for AI training.

All of this illustrates that the best way for the U.S. to maintain AI leadership is to outrun the competition through the combination of domestic investment and an innovation-friendly AI regulatory climate. Trying to stay ahead by tripping up rivals can have the opposite of its intended effect.

  • Acknowledgements and disclosures

    Google, Meta, Amazon, and Microsoft are general, unrestricted donors to the Brookings Institution. The findings, interpretations, and conclusions posted in this piece are solely those of the authors and are not influenced by any donation.

The Brookings Institution is committed to quality, independence, and impact.
We are supported by a diverse array of funders. In line with our values and policies, each Brookings publication represents the sole views of its author(s).