In late January, the launch of the DeepSeek R-1 model triggered discussions across the global tech industry. In this Q&A, Jimmy Goodrich provides a broad picture on this technological disruption and its implication for future U.S.-China AI competition.
Jimmy is a leading expert on technology, geopolitics, and national security with a focus on China and East Asia. He is a senior advisor for technology analysis at the RAND Corporation and a nonresident fellow at the University of California Institute on Global Conflict and Cooperation, where he conducts research on China, technology, and national competitiveness.
DeepSeek’s launch has been described as a shock to the world. What specific advancements set it apart, and why has it garnered such attention?
DeepSeek surprised the world with its ability to train frontier AI models with reasoning capability nearly as good as what global leaders like Anthropic, Google, and OpenAI have been able to do, with a much smaller number of chips. They did that with just under 2,000 Nvidia H800 chips, whereas for the global leaders like OpenAI, they’ve been using tens of thousands, if not over 100,000 chips. So, they showed a lot of technical ingenuity and capability in ways that were surprising to many people around the world. This led to a dramatic sell-off of stocks and global international reactions. And people are still debating today what DeepSeek means.
What’s the difference between the open-source AI model and closed-source AI model? Is being open source DeepSeek’s comparative advantage?
There are two main approaches in developing large language models. There are open-source models and proprietary models. Proprietary models are ChatGPT, Anthropic, and a few others that are training their models with their own data. And when they release them, they don’t release the model weights and other proprietary information, so that you can recreate the model on your own. And they want you to use their model and pay them for access to the model.
The open-source models—Meta, for example, released Llama, which is open source; and now DeepSeek, as well—are trying to create an ecosystem effect by getting more people to use their model because it’s free to use and they can recreate it on their own computer, on their own system, with their own information. Right now, particularly for a lot of applications like enterprises, universities, and researchers who want to do a lot of tinkering with the model, the open-source is very attractive, whereas for maybe your average user that wants to be able to generate answers quickly, the proprietary models are also still preferable.
So, both have drawbacks and benefits. Some would argue that open-weight, open-source models are more secure because you can know what the code is; others argue they’re less secure because they’re proliferating these powerful capabilities and they might be misused by bad actors.
Does DeepSeek signal a broader improvement in China’s ability to nurture talent and foster innovation, or is it more of an outlier in the industry?
I would say that DeepSeek represents the fact that China has improved its standing in innovation significantly over the last couple of decades. It has leading universities, top-tier talent, and different ecosystems around the country that are promoting innovation. DeepSeek is from Hangzhou, which is home to many competitive Chinese private enterprises, including Alibaba. And most of the staff that worked on developing the DeepSeek model were trained indigenously within China and local universities. So, I think this shows that when it comes to software and coding, Chinese companies and individuals are highly competitive, and there’s really not much of an advantage the United States has. The advantage the U.S. has in AI right now is in the hardware and the infrastructure.
How have U.S. semiconductor and chip sanctions influenced China’s indigenous technology development? Have they acted as a constraint or a catalyst for innovation?
I’d say it’s both a constraint and a catalyst. It is a constraint in that some of the leading-edge capabilities that China wants to develop are roadblocked—not permanently, but a very large speed bump—because of things like the extreme ultraviolet lithography from ASML. Some of the advanced Nvidia AI chips like the Blackwell and others are not yet available in China, and China has not yet developed their own alternative.
At the same time, the export controls and sanctions do provide an incentive for China to accelerate its own innovation, invest in its own companies and ecosystems. And over time, China will work to close the gaps. Export controls do not provide a permanent restriction; they only buy the United States and its allies some time to use that time to double down on its own strengths and innovation.
President Xi has emphasized the importance of technological advancement, aiming for China to become a global leader in science at the National People’s Congress. Given economic slowdown and ongoing trade wars, how can China realistically achieve this goal?
China wants to be a global leader in many different things: in technology, in foreign affairs, and defense. But many of these things take a long time. You cannot snap your fingers and become a leader in something overnight; they require decades of investment, cultivation of talent, promotion of business, deep connectivity to global supply chains and international cooperation. In many of these areas, China will have to make long-term investments and have long-term strategies.
But, China is now increasing its science and technology spend by over 10% while the U.S. is flatlining or decreasing its federal investments in science and technology. So, I think overall, the trajectory that China is on is the upward rise, whereas the United States is stagnating, particularly in public investment into research and development. There are plenty of dynamic U.S. companies that are making investments and creating startups; that’s still a strength of the United States. But we have to remember that investment in public research has historically been a comparative advantage for the US; the investments that its federal government has made seeded the growth for things like the internet, GPS, or even advanced semiconductors. The EUV lithography initial research was conducted by the Department of Energy, for example.
Moving forward, what should both the U.S. and China do to maintain competitiveness in this long-term AI race?
For both countries, I think the key here is that it’s a battle of ecosystems. Which country is going to have access to the most and the best talent? Who is going to be able to rapidly diffuse their technology, not just within their own country? Neither the U.S. nor the Chinese market, while both large, are large enough to help their companies become global leaders. So, it’s also a competition of who has access to the most amount of market share worldwide. And also who is in control of not just one part of the ecosystem but the widest or the largest number of nodes across the ecosystem?
It’s not enough just to be a leader in, say, AI data, but it’s about competing in semiconductors, algorithms, development, deployment, and energy production distribution, which is another key aspect. It requires leadership in many different areas, and I think the US and China are focused on trying to have alternative but competing ecosystems.
