Four years ago, the US threw a one-two punch at China’s AI ambitions. The White House rolled out sweeping rules blocking the sale of the highest-end chips to China, and a month later, OpenAI introduced generative AI chatbot ChatGPT, a groundbreaking development that ushered in the present AI boom.

Defending the US decision to impose semiconductor export controls as necessary, former National Security Advisor Jake Sullivan said it would allow America to maintain “as large a lead as possible” over Chinese competitors in autonomous and military systems. ChatGPT left Chinese competitors scrambling for answers, and a botched chatbot launch by Chinese AI giant Baidu the following March revealed the Chinese were still far behind.

Fast forward to today, not only have Chinese AI companies refused to be hamstrung by the restrictions, but their large language models are also more innovative, efficient, and cheaper to run than Western peers. Industry experts say Chinese open-weight models are topping certain benchmarks without the deep pockets and limitless compute available to US rivals, and are only six months behind American frontier AI labs.

The US and China are in the race to develop the world’s most cutting-edge technologies, as advanced AI is expected to upend militaries, global economies, and trade. The commercial stakes are huge: industry researcher Gartner forecasts the AI market to hit US$2.59 trillion in global spending this year.¹

To be sure, US frontier labs call out their Chinese peers for distillation and say the scale of Chinese distillation presents AI safety and economic and national security risks. Distillation is a process where smaller AI models are trained on the outputs of a much larger and more powerful model in order to mimic its capabilities. There have also been glaring loopholes in US export control policy, which have allowed the Chinese access to high-end semiconductors to train and run their algorithms.

Still, not only has the US-China gap closed from four years ago, but Chinese models are also proving hugely popular.

Training a model from scratch can cost millions of dollars, making Chinese open-weight systems an attractive option for many US businesses and researchers at top universities eager to avoid such hefty costs. Models from Chinese tech giant Alibaba have become the most downloaded globally on the popular AI code repository Hugging Face this year², far outstripping Google and Meta’s open models. Within the tech community, China’s AI developers have garnered widespread respect for making advanced computing accessible.

This was a scenario the Biden White House unlikely foresaw in 2022, when they imposed export restrictions on chips, and this is what the Trump administration is mulling as they reassess these policies.

China’s AI advances

Since 2022, Chinese companies such as DeepSeek, Z.Ai, and Alibaba have pushed the boundaries of innovation, making changes to algorithmic structures to improve efficiency in areas where Silicon Valley had long taken for granted.

Unlike US companies, which focused on frontier innovation and pushing toward artificial general intelligence, Chinese innovation was more marginal. Faced with a dearth of computing power to train their algorithms, they focused on intermediate steps such as training innovation and product innovation to improve productivity and efficiency.

The Chinese open-sourced their models to benefit from collective expertise and turned to mixture-of-experts models and context caching to lower inference costs.³ Aggressive competition in China also drove down the prices of application programming interfaces, critical to inference and AI use, even further.

China’s openness and cost have attracted users and contrasted with the closed systems developed by OpenAI and Anthropic. Large US tech giants DoorDash and Airbnb have deployed open-weight AI models from Alibaba and DeepSeek onto their tech infrastructure for nonsensitive, high-volume tasks.

By Spring 2026, the AI platform Hugging Face said in a report that “the geographic composition of the open source ecosystem has fundamentally changed.” Over the 12 months to February 2026, Chinese models accounted for the bulk of the downloads from the site, or about 41%.⁴

Analysts say cheaper and compressed Chinese AI models may win the global race for deployment and adoption in the longer term, as users in Asia, Latin America, and Africa are more cost-conscious and compute-constrained.

In July, open-source AI platform Hugging Face was breached by a rogue cyber attacker. After strict safety guardrails on Western frontier AI models refused to help Hugging Face analyze the attack, the platform turned to Zhipu’s GLM-5.2 model to contain it. It later turned out that autonomous agents from OpenAI had broken out of their sandbox and attacked Hugging Face.

More broadly, Chinese AI research has become more influential. China’s share of the top 100 most-cited AI papers grew from 33 in 2021 to 41 in 2024, according to the Stanford Institute for Human-Centered Artificial Intelligence in its 2026 AI report. China already overtook the US as the top publisher of AI papers in 2017, and Stanford said Chinese institutions accounted for 17.8% of global AI publications in 2024, versus 7.3% from the US. It is also catching up in quality.

With AI hardware and chips, where Nvidia once had an undisputed monopoly on the tools needed for AI training and inference in China, it no longer does. Chinese alternatives have sprung up.

DeepSeek’s latest algorithms are optimized for domestic chip use, and according to Reuters News, Nvidia’s share of China’s market for data-center AI chips fell to 34% in the first half of 2026, as export controls limited supplies and domestic rivals gained ground. Nvidia shipped 678,000 AI accelerators to China in the six months, down 44% from a year ago, Reuters wrote, citing a report by industry researcher IDC.

Necessity is the mother of innovation and faced with the prospect of growing irrelevance, the Chinese have thrived.

The state’s role

Export controls were designed to keep China’s AI developments behind the US. On that count, they succeeded. Even so, there have been unintended consequences.

US export controls laid bare how reliant the country was on US imports and galvanized the country’s leaders to create China’s own self-sufficient semiconductor ecosystem and, more broadly, technology ecosystem.

China has always harbored ambitions to be a world leader in technology. Even so, despite strong government rhetoric in the past about chasing global dominance in high-end technologies, the reality was that in areas that required extreme complexity in manufacturing such as aircraft engines and advanced integrated circuits, Chinese companies were still weak.

To get ahead in AI, Beijing adopted a familiar playbook of industrial policies and state financing, throwing its weight and resources behind the sector.

Since June 2024, Chinese President Xi Jinping has called for the country to take what he termed a “new-style whole of nation approach” to nurture AI breakthroughs. The term comes across as jargon typical of Chinese leaders, but carries strong intent. This involves devoting China’s Party attention, government resources, private companies, and the markets toward achieving a high-priority national objective.

China previously used such language to describe developing an atomic bomb under Chairman Mao Zedong between the mid-1950s and 1960s. Analysts say current references seem to indicate a return to a policy environment in which national objectives, such as overcoming technological bottlenecks, drive research goals.⁵

In addition, Beijing utilized government guidance funds: public-private venture capital vehicles designed to support industrial policy while investing alongside the private sector. In January 2025, the Ministry of Industry and Information Technology and the Ministry of Finance set up a 60 billion yuan (US$8.2 billion) National AI Fund, and later that year, started another fund to support early-stage companies⁶ in "hard tech" like AI and semiconductors.⁷

Similar support was offered to the semiconductor industry in China’s quest for self-sufficiency in the field. According to the Organisation for Economic Co-operation and Development’s Manufacturing Groups and Industrial Corporations (MAGIC) database of industrial subsidies, subsidies to semiconductor manufacturers in China was US$1.74 billion in 2024, representing about 8% of their revenue that year — the highest percentage globally.⁸

US export control loopholes

US export restrictions also succeeded in cutting Chinese technology companies off from advanced semiconductors and the specialized machinery used to make them. Without a breakthrough, the gap between China and the US in cutting-edge chip production and quality is likely to widen as such technologies advance. However, this gave the Chinese a reason to innovate and develop their own alternatives. DeepSeek’s Chief Executive Officer Liang Wenfeng has said publicly that its biggest bottleneck was never money but export restrictions on high-end chips. The Information reported this week that Huawei engineers were embedded in DeepSeek, working closely to design future AI chips.

Export controls worked well when targeting specialized tech areas that are difficult to reproduce, such as the making of high-bandwidth memory chips and advanced lithography machines. Yet, failed miserably when deployed in areas where customer due diligence was critical, giving the US a false sense of security.

High-bandwidth memory chips are critical to the efficiency and function of AI chips, and ChangXin Memory Technologies, China’s leading maker of these chips, is still generations away from similar chips churned out by South Korean market leaders SK Hynix and Samsung. An inability to manufacture high-bandwidth memory chips has restrained China’s ability to mass-produce AI accelerators.

US export controls also limited Holland’s ASML — the sole maker of extreme ultraviolet (EUV) lithography machines — from selling such machinery to China. Lithography, a vital process in the manufacturing of high-end silicon, remains a huge bottleneck for China. Without access to EUV machines, Chinese chipmakers have been using more mature technology to produce advanced logic chips, with much lower efficiency and higher waste.

The controls also resulted in diversionary tactics by Chinese companies and universities, such as the creation of a cottage industry in which middlemen companies bought advanced chips from Nvidia or signed long-term contracts to rent the compute power to Chinese entities. ByteDance, the Beijing-based Chinese tech giant, is working with a Southeast Asian company called Aolani Cloud on plans to use some 500 Nvidia Blackwell computing systems in Malaysia totaling around 36,000 of its premium B200 chips, The Wall Street Journal reported in March.⁹

That same month, the US government charged three people, including Supermicro co-founder Yih-Shyan "Wally" Liaw and a sales manager, for orchestrating a massive US$2.5 billion smuggling operation. The three allegedly used a front company in Southeast Asia to route advanced Nvidia-powered AI servers directly to Chinese buyers. They falsified documents and staged warehouses with nonworking "dummy servers" to deceive compliance auditors.

Huawei also used proxy companies and intermediaries to source chips from Taiwan Semiconductor Manufacturing Company, the world’s largest foundry. Such activity was halted in October 2024.¹⁰

More importantly, such export rules have come at a cost: straining US relationships with its allies and other nations. US allies such as Holland, Japan, and South Korea hold the keys to advanced technologies relevant to AI and are a vital advantage that China doesn’t have.

To be sure, critics argue that China’s industrial policy is wasteful and inefficient and point to what they call “industrial-scale distillation attacks” by the Chinese, and say such behavior represents intellectual property infringement and national security risks.

Still, as the fourth anniversary of the Biden administration’s sweeping semiconductor controls approaches, it is becoming clear that export barriers are no silver bullet. China continues to be able to innovate. And if export controls are not a silver bullet, what must the US do in order to stay ahead?

As Jeffrey Ding, an assistant professor of political science at George Washington University, wrote in a recent blog: the AI race is many marathons, not just one sprint. In order to win the AI race, it is key for the US to focus on self-improvement and plan smart, long-term investments in infrastructure, education, and workforce development, rather than be fixated on fears that China will catch up.

The US-China AI race will likely prove to be one of the most consequential rivalries of this century, and neither side can afford to underestimate the other.