Chinese engineers are hand-carrying massive datasets to Malaysia, bypassing U.S. restrictions on advanced AI chips
In a remarkable and highly coordinated effort to get around U.S. export controls, Chinese artificial intelligence companies are physically transporting massive volumes of data overseas by hand. According to a detailed investigation by The Wall Street Journal, engineers from Chinese AI firms are carrying suitcases filled with hard drives containing terabytes of sensitive training data to countries like Malaysia, where they can rent access to advanced American-made chips. This approach is becoming one of the most effective ways for Chinese companies to continue training large-scale AI models while technically staying within the bounds of current U.S. regulations.
In one striking example, four Chinese engineers flew from Beijing to Kuala Lumpur in March 2025, each with a suitcase packed with 15 hard drives. Altogether, they carried about 80 terabytes of data each—enough to train powerful AI models. Once they arrived, they headed to a data center in Malaysia where their company had rented access to around 300 Nvidia servers. After uploading the data and completing the training process, they returned to China with the final product: the trained AI model’s output and parameters.
Why Carry Data Instead of Transmitting It?
The decision to fly hard drives rather than send data over the internet was not a desperate last resort, but a carefully calculated move. Internet transfers of this scale—around 4.8 petabytes in total—could have taken weeks or even months and might have attracted unwanted attention. According to people involved, the company’s engineers spent more than eight weeks prepping and optimizing the datasets and training software before departure. Once the data was out of China, there would be little chance to make adjustments.
“This was something we were consistently concerned about,” said Thea Kendler, who oversaw export controls at the U.S. Commerce Department during the Biden administration. She noted that Chinese access to American AI chips—even indirectly—was a major worry. Yet because the chips remain outside China and are accessed through layers of foreign subsidiaries, the situation stays legally ambiguous.
How the System Works
The workaround hinges on exploiting a gray zone in export controls. Although U.S. companies like Nvidia are banned from directly selling their most advanced chips to Chinese entities, once those chips are sold to customers in countries such as Malaysia or Singapore, the trail becomes harder to follow.
Chinese companies often operate through a chain of foreign affiliates. In this case, the AI firm that trained its model in Malaysia first worked through a Singapore-based subsidiary. But as export scrutiny increased in Singapore, the Malaysian data center advised the company to set up a local front. According to corporate records reviewed by The Wall Street Journal, the Chinese firm registered a new entity in Kuala Lumpur, listing three Malaysian citizens as directors and an offshore holding company as the parent organization. This made the operation appear local, potentially helping it evade regulatory red flags.
To avoid detection by Malaysian customs, the Chinese engineers divided the hard drives into multiple bags. In the past, they had carried all the drives in one suitcase, but that method risked raising suspicion. By splitting the drives among four travelers, they lowered the chances of triggering inspections. After the model was trained, they flew back to China carrying several hundred gigabytes of AI model output—compressed and ready to be deployed.
Growing Pressure on Transshipment Routes
Although smuggling chips into China through third-party countries used to be common, that route is becoming increasingly difficult. U.S. pressure has caused Southeast Asian countries to tighten enforcement. In February 2025, Singaporean authorities charged three individuals with misrepresenting the destination of Nvidia servers worth millions of dollars. While Singapore’s officials said they are not required to enforce foreign export laws, they clarified that they would not allow their country to be used to dodge such restrictions.
These new obstacles are prompting companies to shift from smuggling hardware to exporting data. “They convinced me that it’s a business with promised returns and that we should move fast before the rules tighten up,” said Adam Ooi, a Malaysian investor who put $5 million into a fund leasing Nvidia servers. That fund now serves clients in China, Singapore, and Indonesia.
Southeast Asia and the Middle East: New AI Training Grounds
Southeast Asia has become a major AI hub. Real-estate firm Jones Lang LaSalle estimates that Singapore, Malaysia, Thailand, and Indonesia now have a combined data-center capacity of nearly 2,000 megawatts—comparable to the largest hubs in Europe like London and Frankfurt. Companies are rushing to install Nvidia and AMD chips across the region. Malaysia alone imported $3.4 billion worth of AI chips from Taiwan in just two months—more than in all of 2024.
The Middle East is also emerging as a key destination. During a recent visit by President Trump, Nvidia signed agreements to supply hundreds of thousands of AI chips to Saudi Arabia, Qatar, and the United Arab Emirates. These nations are building major data infrastructure projects, and Chinese developers are already starting to use their facilities for training AI systems.
In January, a data center in Malaysia originally leased 200 AI servers to an American cloud provider. That contract was quietly taken over by a Beijing-based firm, illustrating how companies from China are finding ways to indirectly access restricted computing resources.
Why the Data Is So Valuable
The data being carried overseas is not just bulk storage—it’s intellectual gold. Training an AI model requires huge amounts of information: documents, images, videos, and code. The training process adjusts billions of parameters that determine how the AI responds to inputs. Once trained, the model can be transferred back to China in a compact form, but the initial data and chip processing are what give it power.
According to the Journal, the Chinese engineers brought back “several hundred gigabytes of data, including model parameters that guide the AI system’s output.” This small bundle represents the final, usable product of a much larger and more intensive computation.
U.S. Policy Is Still Catching Up
The U.S. government has not yet found an effective way to close this loophole. The Biden administration proposed setting country-specific chip caps, which would have made it harder for places like Malaysia to serve as backdoors for Chinese firms. But the Trump administration later reversed that plan, saying it would be too burdensome for American companies like Nvidia. Instead, Washington issued guidance telling firms to prevent their chips from being used to train Chinese models—without offering a clear enforcement method.
An Nvidia spokesperson responded to the report by saying, “Ensuring that the world’s cloud services run on American infrastructure, and not on our foreign competition, promotes national and economic security and helps the U.S. set global AI standards.”
A Complex Challenge with No Simple Fix
These suitcase operations show just how hard it is to enforce modern export bans. While hardware shipments can be monitored and blocked, data is more flexible. And with the right setup, physical delivery of digital data can bypass even the most advanced trade restrictions.
For now, Chinese AI companies are using rented servers abroad, legal loopholes, and hand-carried hard drives to continue pushing forward. As enforcement tightens in one place, they adapt. And unless new rules are implemented that target not just the chips, but also how and where they’re used, these workarounds will likely continue.
ACZ Editor: This is an exceedingly difficult restriction to keep, and may well be useless until the next generation of equipment comes out.