The $300 Million Nvidia Server Case Exposes Gaps in AI Chip Controls

The US Department of Justice has arrested Greg Lui, chief executive of Earthmade Computer, in a case involving the alleged diversion of Nvidia-powered servers to China. Prosecutors say the shipments were worth more than $300 million and were disguised with false paperwork that listed other Asian countries as their final destinations.

Key points

• Lui allegedly worked with freight-forwarding companies in countries including Malaysia and Singapore to reroute restricted equipment.
• The servers contained Nvidia A100 and H100 GPUs. They are not Nvidia’s newest products, but they remain capable of processing large datasets and training large language models.
• Investigators say emails, bank records, purchase orders, and shipping documents helped expose the alleged scheme.
• Lui faces charges involving export-control violations, smuggling, and money laundering. Some counts carry a potential maximum sentence of 20 years, while the smuggling charge carries up to 10 years.

According to the indictment, one transaction involved 27 servers that were formally described as destined for Malaysia. Communications cited by prosecutors allegedly indicated that all of them were instead shipped to China. Another shipment was routed through Singapore, Malaysia, and Hong Kong before allegedly reaching a Chinese company in Hangzhou, a city widely associated with China’s technology industry.

The indictment also describes an effort to disguise a Malaysian buyer in a shipment of 100 servers. Prosecutors allege that the paperwork used a fake buyer and that Lui had previously obtained another person’s identifying documents to support transactions connected to the alleged evasion scheme. In 2024, his company allegedly received more than $176 million linked to the activity.

The case illustrates a structural problem with AI-chip controls. Export restrictions are generally written around products, destinations, and end users, but high-end GPUs often move inside servers or complete data-center systems. A shipment can therefore pass through several distributors and logistics hubs before the hardware reaches its real operator. False declarations, shell buyers, and transshipment through friendly countries can make a product’s final destination difficult to verify.

Nvidia has said that less than one-half of one percent of its products were allegedly diverted to China. The company also argues that such quantities are small compared with the computing capacity China already has domestically. Yet the size of the black market is inherently difficult to measure. Publicly identified cases may represent only the transactions that investigators have managed to trace.

The broader scrutiny concerns Nvidia’s due diligence. Investigators and industry sources have questioned whether certain orders should have triggered additional checks, particularly when a customer’s facility appeared too small to house the number of servers ordered, or when a buyer relied on questionable leases and incomplete end-user information. Nvidia has disputed the idea that a fast-growing startup in a friendly country should automatically be treated as suspicious. It has also said that American companies should follow the law as written rather than anticipate rules that have not yet been imposed.

That position leaves regulators with a difficult choice. They can expand official restrictions and require manufacturers to report more suspicious orders, or they can place greater responsibility on companies to investigate transactions that are legal on paper but implausible in practice. Either approach would increase compliance costs and could slow legitimate AI infrastructure projects.

The likely policy shift is from controlling individual GPU models to examining the entire computing supply chain: the buyer, financing, physical facility, logistics provider, server integrator, and final operator. Without that broader view, restrictions on advanced chips may continue to be undermined by changes in packaging, paperwork, and shipping routes.

Source: Ars Technica AI