In 2017, Beijing unveiled an ambitious guide to managing the development of artificial intelligence, which aims to secure world leadership by 2030. By 2020, the program wanted “significant advances” in AI to demonstrate its progress. Then in late 2022, OpenAI’s release of ChatGPT took the world by surprise—and caught China off guard.
At the time, China’s leading tech companies were reeling from an 18-month government crackdown that drained an estimated $1 trillion from China’s tech sector. It was nearly a year before a handful of China’s AI chatbots received government approval for public release. Some have questioned whether China’s stance on research could hinder the country’s AI ambitions. Meanwhile, the Biden administration’s export controls, unveiled a month before ChatGPT began, are aimed at cutting off China from advanced semiconductors that are essential for training large-scale AI models. Without advanced chips, Beijing’s goal of AI supremacy by 2030 appeared increasingly unattainable.
But fast forward to today, and a slew of impressive Chinese releases suggest that the US AI lead has waned. In November, Alibaba and Chinese AI developer DeepSeek released conceptual models that, in some ways, rival OpenAI’s O1AI preview. That same month, Chinese videogame juggernaut Tencent unveiled Hunyuan-Large, an open-source model in which the company’s tests found the best-performing open-source models developed in the US across several benchmarks. Then in the last days of 2024, DeepSeek released DeepSeek-v3, which is now in the highest position among open source AI on the popular online leaderboard and can withstand the best closed programs from OpenAI and Anthropic.
Read more: How the Benefits—and Risks—of AI Will Grow in 2024
Before the release of DeepSeek-v3, this practice had already attracted the attention of Eric Schmidt, the former CEO of Google and one of the most influential voices in US AI policy. By May 2024, Schmidt confidently asserted that the US retained a two- to three-year lead in AI, “which is forever in my books.” But by November, in a speech at the Harvard Kennedy School, Schmidt had changed his tune. He cited developments from Alibaba, and Tencent as evidence that China is closing the gap. “This scares me,” he said. “I thought the restrictions we put on chips would keep them coming back.”
Beyond a source of national prestige, a lead in AI will likely have consequences for the world’s balance of power. If AI agents can automate large parts of the workforce, they may provide a boost to nations’ economies. And future systems, capable of targeting weapons or hacking enemies, could provide a major military advantage. As nations caught between two superpowers are forced to choose between Chinese or American AI systems, artificial intelligence could emerge as a powerful tool of global influence. China’s rapid development raises questions about whether US export controls on semiconductors will be enough to maintain America’s edge.
Read more: How Israel Is Using AI in Gaza—And What It Says About the Future of War
Building powerful AI relies on three key ingredients: data, innovative algorithms, and raw computing power, or computing. Training data for large language models such as GPT-4o is often dumped online, meaning it is available to developers all over the world. Similarly, algorithms, or new ideas on how to develop AI systems, cross borders easily, as new techniques are often shared in academic papers. Even if they weren’t, China has AI talent, producing far more AI researchers than the U.S. By contrast, advanced chips are more difficult to make, and unlike algorithms or data, they’re physically useful to stop at. border.
The supply chain for advanced semiconductors is controlled by America and its allies. US companies Nvidia and AMD have an active duopoly in datacenter-GPUs used for AI. Their designs are so complex—with transistors measured in single-digit nanometers—that currently, only the Taiwanese company TSMC makes these high-end chips. To do so, TSMC relies on multi-million dollar machines that can only be built by the Dutch company ASML.
The US wanted to use this to their advantage. In 2022, the Biden administration introduced export controls, rules that would restrict the sale of high-end chips to China. The move follows a series of measures that began under the Trump administration, which sought to curb China’s access to chip-making technology. These efforts have not only limited the flow of advanced chips to China, but also disrupted the country’s domestic chip industry. China’s chips are “years old,” U.S. Commerce Secretary Gina Raimondo told 60 Minutes in April.
Read more: Research Finds Stark Global Divide in Ownership of Powerful AI Chips
However, the 2022 export controls have encountered their first problem before the announcement, as Chinese developers have reportedly stockpiled chips that will soon be restricted. DeepSeek, a Chinese developer behind an AI thinking model called R1, which competes with OpenAI’s O1 preview, has assembled a batch of Nvidia A100 GPUs that will soon be closed to 10,000 a year before the introduction of export controls.
Smuggling can also undermine the effectiveness of export controls. In October, Reuters reported that TSMC’s chips were limited to a product made by Chinese company Huawei. Chinese companies have also reportedly acquired banned chips using shell companies outside of China. Some have circumvented export controls by renting GPU access to offshore cloud providers. In December, The Wall Street Journal reported that the US is preparing new measures that will limit China’s ability to access chips from other countries.
Read more: Has AI Development Really Slowed Down?
While US export controls limit China’s access to high-quality semiconductors, they still allow the sale of low-power chips. Deciding which chips should and shouldn’t be allowed can be a challenge. In 2022, Nvidia modified the design of its flagship chip to create a version for the Chinese market that falls within the limits of the restrictions. The chip was still useful for AI development, prompting the US to tighten restrictions in October 2023. “We had a year where [China] I can just buy chips as good as those,” said Lennart Heim, who leads AI and computing at the RAND Institute for Technology and Security Policy. This gap, he says, coincides with the time for new chips to find their way into the infrastructure of AI developers, which is why we are yet to see the full impact of export controls on China’s AI development.
It remains to be seen whether the current limit strikes the right balance. In November, Tencent released a language model called Hunyuan-Large that outperforms the powerful Meta variant of Llama 3.1 in several benchmarks. Although benchmarks are an imperfect measure to compare the overall intelligence of AI models, Hunyuan-Large’s performance is impressive because it was trained using low-power, unrestricted Nvidia H20 GPUs, according to a study by the Berkeley Risk and Security Lab. “It’s clear that they’re using hardware over better software,” said Ritwik Gupta, the study’s author, who also advises the Department of Defense. The rival Chinese lab’s DeepSeek-v3, believed to be the strongest open model available, was also trained using the incredibly small computer. While there is significant uncertainty about how President-elect Donald Trump will approach AI policy, several experts told TIME in November that they expect export controls to remain—and be expanded.
Before the new restrictions were introduced in December, Chinese companies also stocked up on chips that will soon be banned. “This whole strategy needs to be rethought,” Gupta said. “Stop playing whack-a-mole with these hardware chips.” He suggests that instead of trying to limit the development of large-scale languages by limiting access to chips, the US should focus on preventing the development of military AI systems, which he says often require little computing power to train. Although he acknowledges that restrictions in some parts of the chips supply chain—such as the ASML machines used to produce the chips—have been instrumental in slowing China’s domestic chip industry.
Heim says that over the past year, the US lead has shrunk, although he notes that while China may now match the best US open source models, these remain about one year behind the top closed source models. He added that closing the gap does not mean that export controls are failing. “Let’s get away from this binary of export controls that work or don’t work,” he said, adding that it may take a long time for China to feel the bite.
The last decade has seen a dramatic increase in computing used to train AI models. For example, OpenAI’s GPT-4, released in 2023, is estimated to have been trained using about 10,000 times more computing power than GPT-2, released in 2019. There are indications that the trend will continue, such as American companies such as X and Amazon. build supercomputers with hundreds of thousands of GPUs, far exceeding the computing power used to train today’s leading AI models. If it does, Heim predicts that restrictions on US chip exports will hamper China’s ability to keep up with AI developments. “Export controls hit you hard on quantity,” Heim said, adding that even if some limited chips find their way into the hands of Chinese engineers, by limiting the number, export controls make it difficult to train and deploy models at scale. “I expect export controls to hit harder over time, as long as computing remains relevant,” he said.
In Washington, “at the moment, there is a reluctance to bring China into the country [negotiating] at the table,” said Scott Singer, a visiting scholar in the Program on Technology and International Affairs at the Carnegie Endowment for International Peace. Complete sense: ‘[If the U.S. is ahead]why can’t we share something?’”
But he notes that there are compelling reasons to negotiate with China on AI. “China doesn’t have to lead to being a source of great danger,” he said, adding its progress despite computing limitations means it could one day produce AI with dangerous potential. “If China is so close, think about what kinds of conversations you want to have to make sure that the systems on both sides stay safe,” Singer said.
Source link
