Chinese AI Use Accelerating as US Export Controls Ignore ‘Real-World Deployment’: Congressional Commission

The low-cost and open-source nature of Chinese artificial intelligence (AI) products is accelerating their adoption, a key factor in the AI race that U.S. export controls currently do not address, according to a congressional advisory panel.

The U.S.-China Economic and Security Review Commission released a paper on March 23 showing how China has gone “all in” with its open source strategy and how that has reinforced its industrial dominance in the sector despite U.S. leads in technological benchmarks.

“This has resulted in the acceleration of global uptake of Chinese AI and created a feedback loop where widespread adoption drives iteration, then further adoption,” the report reads.

It states that U.S. companies are shifting away from open-source models, in part driven by concerns that Chinese labs may take advantage of them, and that U.S. investments in AI “may be misaligned with the competitive dynamics” at issue. Namely, U.S. policy and grants prioritize the development of frontier models over “real-world deployment,” according to the paper.

China lags behind the United States in total compute power—the focal point of debate about U.S. chip export controls to China—but the innovation afforded by widespread adoption has helped narrow the gap in large language model performance, according to the paper.

As a result, these Chinese companies have “developed key architectural and training advances that are now industry standards,” the report reads.

‘Divergent National Strategies’

Each nation is playing to its strengths, according to the report, and right now, U.S. companies are leading on key benchmarks, while Chinese competitors are seeking an edge with low-cost models.

Major U.S. AI companies “have concentrated investment in compute-intensive frontier models, betting that superior hardware resources will yield transformative capabilities,” the report said.

U.S. export controls follow this logic. Restricting advanced semiconductors and the equipment to manufacture them from being sold to China drastically slowed the regime’s ability to catch up in computing power.

The Chinese approach “is a fundamentally different theory of how AI leadership is won,” according to the report. As with other industries that the Chinese regime considers a strategic priority, the Chinese AI industry is highly subsidized and aims for “ubiquitous adoption and iteration,” which, in effect, spreads out development costs.

Yet this approach “has reshaped the competitive landscape,” and U.S. labs now “risk losing” not only the global market but also the ability to set technical standards and norms, the report said.

“Between 2022—the year ChatGPT was released—and 2025, the number of Chinese open models expanded more than tenfold, from 32 to 337,” the paper reads.

Meanwhile, the United States started with more models and reported a threefold growth in the same period, from 213 to 622.

“Bigger is better, until it isn’t,” the report states.

U.S. models dedicate millions of dollars to training runs meant to “extract as much learning as possible from each unit of compute” under the premise that bigger models trained on more data have better performance, it said.

According to the report, this bet on “scaling” will start to see “diminishing returns” if it is true, as estimated, that Chinese spending on AI will stay flat while major U.S. tech companies spend more than $400 billion, or 10 times what Chinese companies spend, in AI capital expenditure this year.

The Chinese regime steers both the supply and demand of Chinese AI. It also slashed costs for Chinese AI users to accelerate adoption, and it is subsidizing electricity costs, energy infrastructure, and cloud providers’ power bills, according to the report.

The aggressive pricing of Chinese models means many U.S. companies are using them over U.S. models. As one Andreessen Horowitz partner said, up to 80 percent of startups pitching to the firm are using a Chinese model as opposed to a U.S. model.

“U.S. researchers and companies increasingly rely on Chinese base models, creating long-term dependency on infrastructure with embedded censorship and potential security vulnerabilities,” the paper reads.

And because China has such a massive manufacturing industry, where AI is being adopted and creating large data feedback loops, it may have an edge in deployed, specialized small models, the report said. This is significant because industry leaders, such as Nvidia, suggest that this is where the “center of gravity in AI value creation” will shift, according to the report.

“The United States’ AI advantages are concentrated in frontier scale: the largest training runs, the most advanced chips, and the most capital-intensive models. These remain important for research and for select consumer applications,” it reads.

“But they may not determine which country captures value from AI in manufacturing, logistics, scientific research, and robotics—domains where deployment scale, iteration speed, and access to operational data matter more than benchmark performance.”

The paper’s author recommends that Congress consider expanding U.S. President Donald Trump’s Genesis Mission, which directs 17 national labs to accelerate AI deployment in energy, national security, and scientific development, to also cover industrial AI, logistics, and manufacturing.