Chinese AI Leaders Now Admit They Can't Surpass US Without Better Chips

Deepseek and other Chinese tech companies were perceived until recently as major threats to AI hegemony. But inability to access state-of-the-art chips from Nvidia is turning out to be a significant challenge because China's domestically manufactured chips are not competitive. The result is that China's ability to surpass the US in AI has been seriously compromised. 

And just as China is using its control of rare earths as a strategic weapon, so the US is using access to chips. Making the situation more complex, China's government is leery of its industry becoming dependent on US chips, preferring it use Chinese-made tech, even though chips from Huawei and other Chinese companies are insufficiently capable and contributing to the Chinese AI fall-off. It would appear that this US-China double-bind will become the focus of further negotiations. JL

Rafaelle Huang and Tracy Qu report in the Wall Street Journal:

After a year of China’s gains in AI, Chinese AI researchers are becoming pessimistic. The country’s chances of catching up to the U.S. are slim because of a bottleneck in chips. Deepseek tried using chips from Huawei and other domestic vendors, but the results were unacceptable. Justin Lin, who heads Alibaba’s AI, asked about China's leapfrogging OpenAI and Anthropic over the next three to five years: his guess was 20% or less. China is barred by the US from acquiring top-of-the-line chip-making technology. The decision to allow Nvidia to sell its H200 chip in China isn’t a game-changer because the chip is two generations behind the Rubin series, so insufficient for state-of-the-art AI.  “The primary bottleneck is chip-manufacturing capacity. The gap is actually widening.” 

After a year of gung-ho news about China’s gains in artificial intelligence, some elite Chinese AI researchers are coming to a more pessimistic conclusion. The country’s chances of catching up to the U.S. are slim in the short run, they say, because of a bottleneck in chips.

“The truth may be that the gap is actually widening,” Tang Jie, founder of the Chinese AI startup Zhipu, said at a conference last weekend in Beijing. “While we’re doing well in certain areas, we must still acknowledge the challenges and the disparities we face.”

One illustration: When the AI chip leader, Nvidia, introduced its next-generation Rubin hardware in January, it named a number of American companies as customers, but no Chinese AI developer was named because U.S. rules block direct sales to China.

Chinese companies have started discussions about renting computing power at data centers in Southeast Asia and the Middle East to get access to Rubin chips, according to people involved in the talks. That follows companies’ efforts last year to access chips in Nvidia’s Blackwell series.

Deals by Chinese companies to use Nvidia chips in third countries are generally viewed as legal. But they require circuitous arrangements and typically leave Chinese AI developers with fewer chips and more inconvenience compared with well-funded American competitors. 

Justin Lin, who heads development of Alibaba’s AI model Qwen, was asked at the Beijing conference about the chance of any Chinese company leapfrogging the likes of OpenAI and Anthropic over the next three to five years. His on-the-spot guess was 20% or less.

Washington’s export controls limiting China’s access to the world’s most advanced AI chips have dissuaded many Chinese companies from pursuing state-of-the-art AI, which consumes a huge amount of computing power. Instead, they apply the technology for everyday uses, while American companies invest in the latest chips to push the boundaries.

“A massive amount of compute at OpenAI and other American companies is dedicated to next-generation research, whereas we are stretched thin,” said Alibaba’s Lin. “Just meeting delivery demands consumes most of our resources.”

UBS analysts estimate that the combined capital spending of China’s internet leaders—much of it for AI—was equivalent to around $57 billion last year. That is roughly one-tenth of U.S. peers.

 

No one is counting China out yet, because developers such as DeepSeek have shown skill at adapting to limited resources. Two other AI developers—Zhipu, formally known as Knowledge Atlas Technology, and MiniMax—raised more than $1 billion combined in Hong Kong this month in initial public offerings. MiniMax’s share price more than doubled from its IPO price.

“Despite a more challenging operating environment, investors continue to price in the possibility of technological catch-up or breakthrough,” said Alyssa Lee, a veteran tech investor who now works at an AI startup. “That optimism itself speaks to the level of innovation Chinese companies have demonstrated.”

Since it turned heads in the U.S. with a high-quality AI model a year ago, DeepSeek has been publishing techniques to improve the efficiency of AI development, some of which have been adopted by Western researchers. DeepSeek published two research papers this month discussing a new training architecture that allows for development of larger models using fewer chips, and a memory design that makes models run more efficiently.

This year, models developed by DeepSeek and Alibaba have narrowed the gap with the best U.S. models to as little as four months, compared with an average gap of seven months in recent years, according to the nonprofit researcher Epoch AI. Many of China’s leading models are open-source—meaning they are free for users to download and modify. That is helping Chinese companies raise their global profile while top American models remain closed-source.

Chinese companies such as Alibaba are making progress in the field of AI. Gilles Sabrié for WSJ

Yet DeepSeek has run into snafus. Last year, when it was developing its new flagship model, it tried using less-advanced chips from Huawei and other domestic vendors, but the results were unacceptable, so it turned to Nvidia chips for some training workloads, said people familiar with the development. DeepSeek then made progress, and it is preparing to introduce the model in coming weeks, they said.

In recent years, Huawei and many Chinese chip startups have made headway with their products—Zhipu said Wednesday it created an open-source image-generation model using only Huawei chips—but the performance gap with the best American chips remains wide.

China is barred by Washington from acquiring top-of-the-line chip-making technology. Companies can’t tap leading chip manufacturers in Asia such as Samsung and Taiwan Semiconductor Manufacturing Co. to produce many advanced chips, but have to rely on less-advanced imported and domestic machines to increase capacity.

 

“The primary bottleneck is chip-manufacturing capacity,” said Yao Shunyu of the Chinese internet company Tencent at the Beijing event. Yao recently left OpenAI to lead Tencent’s AI push.

Washington’s recent decision to allow Nvidia to sell its H200 chip in China isn’t likely to be a game-changer in helping Chinese companies catch up, said people in the industry. Nvidia Chief Executive Jensen Huang has said Chinese demand for the chip is high, but people at Chinese tech companies said the chip, which is two generations behind the Rubin series, has become insufficient for training state-of-the-art AI. 

Companies are still waiting for Beijing’s approval to purchase the H200s. Chinese officials have recently told some companies that any purchases should be for “necessary” uses such as advanced AI research, people familiar with the guidance said. They said China would continue to push adoption of domestic chips.