Beyond the Headlines on DeepSeek’s Sputnik Moment: A Conversation with Jimmy Goodrich

The release of DeepSeek-R1 rattled markets at the end of January. What’s the real story here—and what did the headlines miss?

DeepSeek is a private Chinese technology company based in Hangzhou that started out in quantitative trading before shifting to developing artificial intelligence (AI) models. For the last couple years, DeepSeek has been pursuing artificial general intelligence (AGI), which seeks to mimic or surpass human abilities on cognitive tasks.

Released on January 20, DeepSeek-R1 is a reasoning model, much like OpenAI’s o1 model. It took nearly a week before it exploded in Western media. What grabbed people’s attention was that DeepSeek appeared to develop the new model with far less computing power than competitors like OpenAI and Anthropic. Nvidia, which designs the most advanced chips that U.S. AI companies use but which DeepSeek does not have access to, saw its stock price fall significantly as a result.

The wider market panic was, frankly, mostly a misunderstanding. DeepSeek actually has quite a lot of computing. It bought 10,000 Nvidia A100 chips in 2021 and has its own supercomputer. DeepSeek may have used a smaller number of chips to train this algorithm than its U.S. competitors, but it’s been experimenting on a large supercomputer for a long time. These models benefit from the scaling law, meaning that as computing power increases, the power of these models also increases. Therefore, even though DeepSeek managed to train this current algorithm with a lot less computing power, to move forward it will have to scale up with a larger number of powerful chips.

It appears that the competitive moat that OpenAI, Anthropic, Meta, and other American AI companies thought would keep China out was an illusion.

Even so, it appears that the competitive moat that OpenAI, Anthropic, Meta, and other American AI companies thought would keep China out was an illusion. The Americans thought they had a two-year lead over their Chinese rivals, but now that seems more like three-to-six months. The United States and China could even be at break-even at this moment, particularly on the algorithms that power AI.

What DeepSeek did is really impressive. It employed emerging techniques known as “test-time compute” and “reinforced learning,” allowing it to train the model much more efficiently. DeepSeek claimed to spend $5.5 million, although some third-party estimates say it’s probably closer to $100 million if you include engineering and other accumulated costs. Still, DeepSeek is a highly innovative company, run by local Chinese, most of whom have not been educated overseas. Clearly, the United States doesn’t have a monopoly on AI talent either.

The U.S. government has tried to restrict China’s access to advanced chips and semiconductors to halt its progress on emerging technologies. Does the release of DeepSeek-R1 show that these policies aren’t working?

First, export controls have clearly had a very large impact on China’s domestic semiconductor industry. But they have a lag time. You can’t control what’s already been exported. There are signs, however, that China is having trouble producing enough of its own chips to train AI systems on indigenous technology.

It’s worth remembering that DeepSeek still trained R1 on U.S. chips. When the Biden administration’s October 2022 export controls on advanced chips were implemented, DeepSeek already had over 10,000 Nvidia chips. From October 2022–October 2023, DeepSeek was legally allowed to acquire Nvidia’s less-advanced workaround chips built for the China market—the A800 and the H800—which are still very good for training AI systems.

Had the Biden administration moved more quickly to close the loophole and block the sale of Nvidia’s workaround chips to China, DeepSeek might have been forced to use its older A100s, and training R1 would have taken longer and been more difficult than it was. However, current export controls will probably bite in the future as Chinese companies like DeepSeek seek to scale up their systems, which will require access to more chips.

In the wake of the R1 release, bipartisan groups of lawmakers have called for restrictions on Nvidia’s H20 chip, the company’s most advanced chip that can be sold without a license to China. Although it is below the current thresholds for export restrictions, a number of experts have noted that the H20 is still very good for inference, which happens when an AI chatbot takes a question and provides an answer. In order for DeepSeek to deploy R1 to customers, it needs inference, which requires a lot of memory bandwidth to access data as users query the system.

American AI companies have announced hundreds of billions in investments to advance the technology. When DeepSeek can achieve this sort of a breakthrough at a fraction of the cost, are those investments a mistake?

Computing power still matters for the future of large language models and frontier AI. The highly efficient techniques that R1 used can be deployed by OpenAI, Anthropic, Meta, and Google on a much larger scale, theoretically giving them an even stronger advantage.

Investment in AI data centers is going to remain strong. If you want to competitively offer your model to hundreds of millions—maybe even billions—of users, you need a lot of chips to enable inference, even if you might be able to use a smaller amount to train the system.

It appears DeepSeek actually ran into constraints due to its limited computing power. After its model was released, it had to cancel new account registrations. DeepSeek claimed this was because of cyberattacks, which may have been the case. But more likely, DeepSeek just didn’t have enough inference computing power available to support all its new users.

After the DeepSeek-R1 release, OpenAI, Nvidia, and Anthropic all put out statements saying that the prevailing AI compute paradigm still holds. While those statements may be self-interested and biased, their capital expenditures remain unchanged. Microsoft and Meta announced that their capital expenditures for cloud infrastructure are going to remain the same, another signal that U.S. tech isn’t reversing course.

What insight can DeepSeek offer into China’s innovation system?

There’s a lot of focus on how the Chinese government is driving innovation by setting metrics, drafting five-year plans, and releasing state funds. But DeepSeek is an outlier. It has no strong connectivity to the government. It shows that within China, some of the best innovation can come where the government least expects it.

After all the international headlines, the Chinese Communist Party (CCP) is going to keep a close eye on the company.

The question is: will this new attention on DeepSeek be good for the company? DeepSeek’s founder, Liang Wenfeng, was invited on January 20—the same day as the model release—to attend a State Council meeting. He met with Chinese Premier Li Qiang, the second-ranking Politburo Standing Committee member after President Xi Jinping. After all the international headlines, the Chinese Communist Party (CCP) is going to keep a close eye on the company.

DeepSeek publicly and openly talks about pursuing the goal of AGI. Even before we get to AGI, this means creating agentic AI systems with the potential to complete tasks autonomously on our behalf at the same level or better than humans. Think about telling a chatbot to book your travel for a week-long trip—hotel, flights, Uber—and it doing so without any mistake. Once systems are able to do those type of tasks, automating large swathes of the economy and putting people potentially out of work, that will attract high-level attention by governments.

The CCP prioritizes stability above all else. Economic stability and guaranteed employment are important priorities for the party. If AI threatens to displace workers or creates security threats, the party could get involved. Now that DeepSeek has become famous, it is possible that the CCP could seek to slow or install more control over it to make sure its development is not done in a “disorderly” way, as Xi Jinping says.

DeepSeek was released January 20, the day of the U.S. presidential inauguration. Was that coincidental, or was a message being sent?

DeepSeek-R1’s release on January 20 may have been a coincidence. Certainly, Chinese media used the release to hype up the country’s technological prowess. China has a lot of incentive to prove that its technology, even under sanctions, is as good or better than Western technology. Beijing’s perspective is this narrative could act as a deterrent against further export controls and decoupling measures. The message from Beijing is that the West can do whatever it wants, it doesn’t matter because China is strong and powerful. That message was clearly amplified by state media and by Chinese nationalist online trolls.

The R1 release also came amid other high-profile technological accomplishments. Also on January 20, China announced achieving a record 1,000-second burn rate in its fusion reactor. China test flew its new sixth-generation fighter jet back on December 26. China is eager for the West to recognize it as a technological superpower that cannot be constrained. Indeed, China has made significant progress and is certainly more self-sufficient than it was a decade ago. But China still has a number of weaknesses—and no country is fully self-sufficient.

In that vein, is it fair to say that DeepSeek-R1 represents a real Sputnik moment?

One thing is certain: we cannot deny that China is an increasingly formidable scientific and technology superpower.

It’s still too early to say. One thing is certain: we cannot deny that China is an increasingly formidable scientific and technology superpower.

Science and technology is a top priority of the CCP. Wherever politburo members and provincial governors go, they talk about leading in global innovation. They are directing massive investments into new science and technology projects ranging from fusion reactors to deep-sea exploration projects.

Most recently, at the 20th Party Congress, China established a new Central Science and Technology Commission to rally its scientists toward national scientific greatness. Government investment is being unleashed into universities, national labs, tech startups, venture capital funds. Coinciding with the DeepSeek release, the government also announced a National AI Industry Investment Fund, much like the national “Big Fund” for semiconductors. The Bank of China also announced that it will provide up to $134 billion in loans to support the domestic AI industry.

China is not backing down under American pressure. Instead, it’s doubling down on strengthening its innovation ecosystem, investing in basic science, its national labs, cultivating its workforce. That is something that America needs to pay attention to.

If this is a Sputnik moment that wakes the West up to a resurgent China that is a force to reckon with in terms of science and technology, that’s a good thing. That should encourage the United States to focus on reinvigorating its own innovation ecosystem, ensuring that our universities remain open to talent around the world, that we have investments from the federal government into basic science and our national labs, and that companies can create startups and grow. If this triggers a response in terms of focusing on our own strengths and making them better, that’s a very positive thing.

Jimmy Goodrich is an IGCC nonresident fellow. The views expressed by the author are his own and do not reflect those of any organization with which he is affiliated.

Thumbnail credit: Matheus Bertelli (Pexels)