The United States has worked steadily over the past three years to limit China’s access to the cutting edge computer chips that power advanced artificial intelligence systems. Its aim has been to slow China’s progress in developing sophisticated A.I. models.
Now a Chinese firm, DeepSeek, has created that very technology. In recent weeks, DeepSeek released multiple A.I. models and a chatbot whose performance rivals that of the best products made by American firms, all while using far fewer of the high-cost A.I. chips that companies typically need. Over the weekend, DeepSeek’s chatbot shot to the top of Apple’s App Store charts as people downloaded it around the world.
The development has raised big questions about export controls built by the United States in recent years. The Biden administration set up a system of global rules and steadily expanded them to try to keep advanced A.I. technology — particularly chips made by Nvidia — out of Chinese hands. They were concerned that technology would give China an edge not just economically, but also militarily.
DeepSeek’s development has provoked a fierce debate over whether U.S. technology controls have failed. Here’s what to know.
DeepSeek’s innovations suggest the Biden administration may have acted too slowly to keep up with private companies sidestepping its controls.
DeepSeek has said that its most recent model was trained on Nvidia H800s. This is an A.I. chip that Nvidia developed specifically for the Chinese market after export controls were first imposed, and that caused a fair amount of drama in Washington.
When the United States put restrictions on Nvidia’s most advanced chips in 2022, Nvidia quickly adapted by creating slightly downgraded chips that fell just under the threshold the government had set. These chips were technically legal for Chinese companies to use, but allowed them to achieve practically the same results.
This angered Biden officials, and they moved to restrict the new chips as well. But the government moved slowly, and it took them about a year to ban the H800 and other downgraded chips. In the meantime, Chinese companies stockpiled a lot of them.
It’s not clear how DeepSeek obtained its Nvidia H800s, but it would have been legal for the company to buy them in late 2022 or 2023. Now, however, such purchases would not be.
“You can’t control what’s already there,” said Jimmy Goodrich, a senior adviser for technology analysis at the RAND Corporation. “Had the Biden administration more quickly responded and limited the H800 to China, there’s no doubt DeepSeek would have been more challenged in putting this model out.”
DeepSeek also spent years building up its chip supply before Washington’s controls took effect. By 2021, DeepSeek was one of just a handful of Chinese companies that had acquired at least 10,000 Nvidia A100s, the advanced chip Nvidia released in 2020, according to an interview with Liang Wenfeng, the founder of DeepSeek, in the Chinese media outlet 36Kr.
The U.S. has also struggled to stamp out chip smuggling.
There’s no evidence that DeepSeek has used smuggled chips. But many Chinese A.I. companies have. Alexandr Wang, the chief executive of the A.I. training giant Scale AI, told The New York Times that Chinese companies had far more high-end chips than U.S. restrictions allowed, and that DeepSeek probably had about 50,000 Nvidia advanced H100 processors, “which they obviously can’t talk about.”
Both Nvidia and the U.S. government have argued that the scale of smuggling was limited. But The Times last year reported an active trade in China in restricted A.I. technology. In a bustling market in Shenzhen, in southern China, chip vendors reported engaging in sales involving hundreds or thousands of restricted chips.
Representatives of 11 companies said they sold or transported banned Nvidia chips — including A100s and H100s, the company’s most advanced at the time — and The Times found dozens more businesses offering them online. One vendor in Shenzhen showed a reporter messages arranging deliveries of servers containing more than 2,000 of Nvidia’s most advanced chips, a transaction totaling $103 million.
Since then, more reports have emerged documenting large-scale smuggling, particularly through other countries in Asia.
The Biden administration released a sweeping regulation this month that aims to deal with the smuggling issue, by setting caps on the number of chips that Nvidia can sell to every country worldwide.
It remains to be seen what the Trump administration will do about it. In a trade executive order President Trump signed on his first day in office, however, he ordered his officials to review the U.S. export control system, including “how to identify and eliminate loopholes in existing export controls.”
U.S. controls appear to have encouraged Chinese ingenuity — but they have also clearly held back China’s A.I. development.
American technology restrictions appear to have accelerated the efforts of Chinese researchers to try to do more with less.
The most notable thing about DeepSeek’s model is that, according to the company, it was developed with just a fraction of the high-priced chips that Western companies have used to make similar technology. DeepSeek’s engineers said they used only about 2,000 Nvidia chips, whereas most top companies have trained chatbots using 16,000 chips or more. Nvidia’s shares plunged sharply on Monday on fears that technology companies will be able to do cutting-edge A.I. in the future while paying Nvidia far less.
Jeffrey Ding, a professor at George Washington University who studies emerging technologies, said that most global companies have been using ever-larger amounts of computing power and data to improve A.I. performance. But DeepSeek and other Chinese firms had been “forced to go down this other pathway to find out whether we can get good enough performance with lower training costs and less compute,” he said.
The implications of cheaper models like DeepSeek’s could be profound. With DeepSeek openly sharing details about how it built its model, companies in China and around the world will be able to replicate its low-cost approach.
That means “it will be much cheaper and could be far less energy intensive for anyone to build and run A.I., from U.S. hyperscalers to Midwestern small businesses, North Korean hackers and Russia’s military,” said Martin Chorzempa, a senior fellow at the Peterson Institute for International Economics.
Still, China would likely be much further ahead in A.I. without the export controls. In interviews, DeepSeek’s founder has acknowledged that the lack of access to computing power was a limitation for the company.
Unlike American A.I. companies, DeepSeek will not be able to legally purchase the newest generation of A.I. chips that Nvidia is rolling out right now, which multiplies the speed and performance of the previous chips.
“Anyone worried about what DeepSeek can do today would be more worried if they had done it with access to the far superior computing resources their U.S. competitors have,” Mr. Chorzempa said.
DeepSeek’s success suggests that Silicon Valley’s lead on A.I. has shrunk, despite efforts by Washington to limit Chinese access to the advanced chips. But it’s notable that DeepSeek is still building its models on Nvidia chips — not on the rival A.I. chips that the Chinese technology firm Huawei is trying to develop.
Some Chinese computer engineers have suggested it would be possible to run the latest DeepSeek model on a larger number of less advanced chips, including those made by Huawei, even though Huawei’s A.I. chips are much lower performing.
But no Chinese company is yet able to make advanced A.I. chips that rival Nvidia’s, or the type of complex machinery needed to make those chips. “The only advantage the United States still has over China at this moment is in hardware,” Mr. Goodrich said.
