In a bold move last month, Jensen Huang, the CEO of Nvidia, arrived in Beijing with a pledge to "unswervingly serve the Chinese market." This commitment came just days after the Trump administration introduced fresh controls effectively banning the sale of Nvidia's H20 AI processor to China. This clash highlights the ongoing technological cold war, where America aims to stifle its primary AI rival by cutting off access to advanced semiconductors. Yet, a shadowy global supply chain has emerged, proving that these chips are still finding their way into Chinese hands.
The Sanctions Tighten, But Loopholes Remain
The US strategy to curb China's AI ambitions hinges on restricting the sale of powerful chips. An AI processor's performance depends on computing power and memory bandwidth. In October 2022, the Biden administration barred sales of American chips exceeding specific thresholds in both areas. Nvidia's response was the H800, a China-specific model designed to stay just under those limits. When America tightened rules again a year later, focusing solely on computing power, Nvidia created the H20.
However, the reality is that restricted Nvidia chips continue to fuel China's AI development. A complex ecosystem has evolved to bypass sanctions, involving offshore data center leases and murky intermediaries. Further attempts to stem this flow face significant, entrenched challenges.
Malaysia: The New Data Center Backdoor to China
To understand the difficulty of enforcement, look at Johor in southern Malaysia. Once known for palm-oil plantations, this region bordering Singapore has transformed into a major data center hub. With cheap land, affordable electricity, and easier permits, it attracts all major US cloud providers—Amazon, Google, Microsoft, and Oracle.
According to property consultancy Knight Frank, Johor's total data-center capacity skyrocketed from 10 megawatts in early 2021 to over 1,500 megawatts by 2024. This infrastructure provides a convenient backdoor for Chinese firms. Companies like ByteDance, TikTok's owner, rent capacity there, gaining indirect access to chips banned for import into mainland China.
Consultancy SemiAnalysis estimates that nearly half of Johor's projected 2027 data-center capacity will feature AI processors like Nvidia's. While operators claim compliance with US rules, workarounds are simple. A lawyer in the region notes that Chinese companies easily acquire restricted chips by establishing local subsidiaries. Trade data supports this: Taiwan exported $3.6 billion worth of graphics-processing units (GPUs) to Malaysia in Q1 2025, nearly matching the total for all of 2024.
The Smugglers' Network and Global Re-export Hubs
Beyond cloud leasing, a direct smuggling network traffics chips into China through third countries not covered by US restrictions. Goods often pass through multiple jurisdictions and front companies, with export papers doctored and products mislabeled. Erich Grunewald of the Institute for AI Policy and Strategy estimates that smuggled American chips constituted between 10% and 50% of China's AI-model-training capacity last year.
The shifting patterns of Nvidia's revenue tell a revealing story. Before the 2022 controls, China accounted for about 22% of Nvidia's revenue, a share that has since fallen to 13%. Meanwhile, sales to Singapore—a city with few end-users—more than doubled, now making up nearly 18% of the total and becoming Nvidia's second-largest market after the US. Nvidia attributes this to routine invoicing practices, but a February case saw Singaporean police arrest three men over the sale of $390 million worth of servers containing Nvidia chips, allegedly re-exported to Malaysia.
The financial incentive is clear: banned Nvidia chips now sell at a 30-50% mark-up through intermediaries. China isn't the only final destination; in October, America sanctioned several Indian firms, including Mumbai-based Shreya Life Sciences, for re-exporting restricted chips to Russia.
Nvidia's Dilemma and the Enforcement Challenge
Nvidia finds itself in a precarious position. The company insists on compliance with US export rules, but its vast operations—expecting to sell over 6 million AI chips this year—make end-to-end oversight nearly impossible. Chips flow from Nvidia to cloud giants and server makers like Dell, where responsibility for vetting end-users becomes diffuse. An executive at a server manufacturer admitted that properly verifying all customers is "practically impossible."
Enforcement is severely hampered by a lack of resources. The Bureau of Industry and Security (BIS), the US agency enforcing tech export controls, has just one export-control officer for all of Southeast Asia and Australasia—a region central to the shadow trade. While technical solutions like location-disabling chips have been proposed, Nvidia argues they are unworkable and could create security vulnerabilities.
The Biden administration had drafted a complex plan restricting chip access to over 120 countries to choke off China. The Trump administration has promised a simpler rule, which relieves Nvidia as those countries represented about a quarter of its sales. Future measures may focus on tightening restrictions on intermediary countries like Malaysia and using chip access as leverage in trade talks.
Ultimately, even improved enforcement has limits. Nvidia cannot trace every chip, and BIS cannot inspect every server. Smugglers will persistently find new loopholes. The article concludes that if America wants to maintain its lead in the AI race, it may need to focus more on accelerating its own innovation rather than solely relying on increasingly stringent and complex export controls.
