Shortages of advanced AI chips are so acute that Beijing is intervening and tech companies are resorting to workarounds

Beijing is taking an aggressive approach to help its technology giants squeezed by America’s chip restrictions.

Shortages of advanced semiconductors are so acute that the government has begun intervening in how the output of China’s largest contract chip maker, Semiconductor Manufacturing International , is distributed, according to people familiar with the matter. 

Chinese authorities are trying to give priority to the needs of tech conglomerate and national champion Huawei Technologies, which uses SMIC technology to make artificial-intelligence chips, the people said.

Chinese tech companies are fighting to secure limited domestic capacity and, in some cases, labs are smuggling coveted supplies of high-performance Nvidia chips.

Buzzy AI upstart DeepSeek had to delay the release of its latest model earlier this year because of a shortage of chips, people familiar with its operations said. 

And companies such as Huawei are cobbling together workarounds, including by bundling thousands of chips into huge, power-hungry systems that can help train AI models, people familiar with their moves said.

The lengths to which Chinese companies and Beijing are going in the face of recent U.S. export restrictions are a sign of the stakes in the race for AI supremacy .

Top U.S. officials are divided on whether to continue limiting China’s access to chips and semiconductor manufacturing equipment, or allow more sales. 

Their goal is to prevent chips made by Huawei from becoming more advanced and in demand around the world. The White House’s decision has ramifications at home—for companies such as Nvidia—and abroad.

President Trump elected not to discuss the potential export of a new Nvidia chip to China in a recent meeting with Chinese leader Xi Jinping after top officials warned him about the security risks, which include boosting China’s military by providing better technology.

Jensen Huang , Nvidia’s chief executive, has argued that China is rapidly increasing production of semiconductors and is home to half of the world’s AI developers. 

He says the world’s most valuable company should be allowed to export a version of its Blackwell advanced-AI chip to China to compete with Huawei and counter the proliferation of its chips across the globe.

“China is nanoseconds behind America in AI,” he said recently.

Nvidia and allies such as White House AI Czar David Sacks say many experts underestimate how fast China is improving and that its state-led investments will eventually pay off.

“Our mature chip capacity accounts for about 28% of the global market, and the advanced chip sector has produced groundbreaking outcomes,” said Liu Pengyu, a spokesman for the Chinese embassy in Washington.

Representatives for the Ministry of Industry and Information Technology, which oversees the technology sector in China, SMIC, Huawei and DeepSeek didn’t respond to requests for comment.

China’s self-reliance
It is difficult to quantify China’s semiconductor production, but even the most aggressive forecasts fall short of the country’s demand, illustrating its challenge in becoming fully self-sufficient, many experts said.

“You could multiply the numbers by five and it might still not be close to satisfying the domestic market,” said Saif Khan, a distinguished technology fellow at the think tank Institute for Progress, who worked on export restrictions in the Biden administration.

Chinese officials acknowledge their chipmaking capacity isn’t as advanced as that of the U.S., but say that it has developed quickly and is underpinned by a political system capable of accelerating industrial production.

“Maybe now, it’s not as good as in the U.S., but the gap will become smaller and smaller,” Ma Xiaoxiao, China’s deputy consul general in New York, said at a news conference last week.

As China tries to become more self-sufficient, authorities have instructed state-owned data centers to stop using Nvidia’s chips, with some even removing existing Nvidia products from service, people with knowledge of the moves said. 

Authorities have also told companies not to use an older Nvidia product, a move some analysts view as posturing to negotiate access to a better chip with the U.S.

The shift has been painful. Engineers are accustomed to Nvidia’s proprietary software ecosystem, and some say they struggle with domestic alternatives, citing problems such as overheating, system crashes and a lack of software support.

Workarounds
Up against restrictions, some semiconductor companies such as Shanghai-based MetaX are designing chips on older, more available technology , bundling two or more smaller chips together to compensate for more limited computing power. 

Bundling strategies at Chinese companies have resulted in electricity-guzzling data centers, prompting multiple local governments to start subsidizing their power bills, people familiar with the matter said.

Companies in China also continue to smuggle Nvidia chips or access them remotely in other countries using cloud computing.

According to contracts reviewed by The Wall Street Journal, at least 16 refrigerator-sized Blackwell racks—shipped in smaller parts and reassembled in China—were set to be delivered by November. 

That amount of chips isn’t sufficient to train advanced, large-scale AI models, but it can still be useful for conducting research and developing powerful applications.

Nvidia has said smuggling doesn’t occur on a significant scale and disputes figures showing a capacity crunch or need for the company’s products in China. 

“China’s own industry has more than enough domestic AI chips and servers for any undesired or military use, with millions left to spare,” a spokesman said.

The chips that China is able to produce on its own have limitations. The U.S. has banned the sale of tools known as extreme ultraviolet lithography machines , which inscribe silicon wafers with microscopic patterns filled by billions of transistors.

Chinese fabrication plants have been forced to rely on older manufacturing techniques that are less efficient, opening the door to more errors and reducing the percentage of chips manufactured that are usable. 

In a September report, Morgan Stanley analysts estimated that making Huawei’s advanced 910C chip using SMIC’s technology could mean 95 out of every 100 pieces of silicon produced are unusable.

The fight in Washington

The Institute for Progress recently found U.S. production of chips comparable to Nvidia’s top-tier B300 will be nearly 25 times China’s output this year. That gap is forecast to widen to roughly 40 times by next year.

To Khan and many experts, China’s chip struggles and shortages show that U.S. export controls are working. They say that allowing China to access even downgraded chips would help close the AI gap.

Opponents of the export restrictions say recent advancements by Chinese model developers show that opaque estimates of chip production are part of a broader AI puzzle that the nation is solving on its own.

Huawei’s plans to export chips to other regions like the Middle East and release new products to compete with Nvidia such as an AI supercomputer unveiled this year show the company’s ambitions and rapidly evolving capacity, they say.

Michael Frank, CEO of a geopolitical risk platform and a longtime tech analyst, recently visited China and said policy experts seemed less concerned about chip-export restrictions than they were previously. AI researchers felt comfortable they could get roughly 80% of the performance they could otherwise, he said.

“It’s much much more complicated than any one aspect and there are tons of workarounds,” Frank said.

Nvidia has spent nearly $3.5 million on lobbying in the first three quarters of this year, up from $640,000 in all of 2024, according to data provider OpenSecrets. That includes spending on policies beyond exports.

In recent weeks, Jack Mallery, a former Central Intelligence Agency analyst focused on China who now works for Nvidia, told administration officials, Congressional staffers and think tanks that estimates of Chinese production are too low, according to people familiar with the discussions. 

Mallery at times contradicted government figures, in some cases citing industry estimates that are publicly available.

The potential discrepancy in data is at the heart of the dispute over export controls. And performance differences make apples-to-apples comparisons of chips difficult.

In June, Commerce Secretary Howard Lutnick testified to Congress that Huawei was capable of making about 200,000 AI chips a year. However, many analysts and people close to Huawei say that its output is much higher.

Research shop SemiAnalysis estimates Huawei will have produced 805,000 Ascend processors this year, though not all of those are comparable to Nvidia’s top products. Huawei plans to more than double its total capacity by next year, according to people close to the company.

The U.S. also expects Huawei’s output to jump in 2026, an administration official said. Treasury Secretary Scott Bessent recently signaled that the U.S. could allow some exports in one to two years—as long as the chips in question were generations behind Nvidia’s best.

Write to Lingling Wei at Lingling.Wei@wsj.com , Amrith Ramkumar at amrith.ramkumar@wsj.com and Robbie Whelan at robbie.whelan@wsj.com