Smugglers are reportedly pairing premium HBM from Samsung and SK hynix with deliberately underperforming processors, allowing the combined package to clear export regulations as a low-end chip. Once the shipments reach Shenzhen, workshops heat the modules to detach the memory stacks, which are then resold for roughly $4,000 each—primarily to Huawei affiliates.
The scheme exploits a December 2024 exclusion from the U.S. Bureau of Industry and Security (BIS) for co-packaged chips, provided the “dominant function” of the package is processing. By mating top-tier HBM3E with a processor so weak no data center would buy it, exporters classify the bundle as a logic device, sidestepping license requirements that apply to standalone high-bandwidth memory.
Chinese tech commentator AB Kuai.Dong amplified the findings from The Information on X, describing the decoy processors as “clunky” and the arrangement as a “compliant export.” Further reporting from Wccftech identifies a key figure: a man in his 40s, fluent in Chinese and Korean with a background in the Korean semiconductor industry. He ostensibly runs a chip design firm whose processors are intentionally nonviable commercially, creating a pretext to procure high-end HBM in volume.
Packaging houses in Taiwan, South Korea, and Southeast Asia reportedly bond the memory and logic die onto a single substrate. The finished packages ship to Hong Kong declared as processors, then cross into mainland China via luggage, vehicle trunks, and daily commuters. In Shenzhen workshops, technicians apply heat to separate the HBM stacks—a risky step; Seoul Economic Daily notes that excessive or rapid heating can warp or crack the memory, yielding high failure rates. The margins, however, remain substantial enough to absorb the losses.
A Regulatory Gap at $4,000 a Stack
The BIS rules published in December 2024 require a license to export standalone HBM with bandwidth density above 2 GB/s/mm² to China. A technical note, however, excludes co-packaged chips combining HBM and logic, judging them instead by the package’s processing performance. Pair a fast memory stack with a slow processor, and the bundle qualifies for export even if the memory is the sole object of value.
The regulator anticipated some evasion: the same note retains controls on HBM permanently affixed to a logic die serving only as a physical-layer controller. It did not anticipate a processor engineered to be useless. In the rule’s own language, the exclusion applies only “where the dominant function of the co-packaged integrated circuit is processing”—a condition difficult to argue when the package’s value resides almost entirely in its memory.
A Recurring Pattern
This tactic is not new. In April 2025, SemiAnalysis reported that CoAsia Electronics, Samsung’s HBM distributor in Greater China, shipped HBM2E to Taiwanese design house Faraday Technology, which had SPIL package it with a low-cost 16 nm logic die before sending it to China for desoldering. “The HBM export ban is specifically for raw HBM packages,” the analysts wrote. Faraday denied violating export rules and denied ties to Huawei; its president, Steve Wang, told the Taipei Times that “it is impossible to recover an HBM unit through desoldering.” Eighteen months later, Shenzhen workshops are doing precisely that.
SemiAnalysis estimated that by September 2025 the CoAsia-Faraday flow had ceased, but not before China had amassed roughly 13 million HBM stacks—11.4 million directly from Samsung, with 7 million arriving in the single month between the rule’s announcement and enforcement. Teardowns have since found Samsung and SK hynix memory inside Huawei’s Ascend 910C. As recently as February 2026, AI Frontiers described the packaged-HBM breach as “apparently contained,” while urging BIS to issue red-flag warnings on co-packaged HBM deployed in ways that allow easy memory removal post-export.
If The Information‘s sources are accurate, the technique was never contained—only quieted. It has since migrated to HBM3E and spread. A Shenzhen broker in his early 20s told the outlet that a single HBM3E stack sold for a few hundred dollars in China a year ago; today it commands about $4,000, a near tenfold increase driven by acute scarcity. With six to eight stacks required per AI accelerator, memory alone accounts for $24,000–$32,000 per chip before factoring in the processor, server, or power costs.
TrendForce, citing the same report, notes that Huawei has raised the suggested price of its Ascend 950DT above RMB 250,000 (roughly $37,000), attributing the increase to memory costs and tight supply. At current black-market rates, the HBM for a single accelerator would exceed half the device’s sticker price, though the 950DT’s exact stack count is unspecified.
SK hynix stated via SBS News that it complies with U.S. and South Korean export controls and has not supplied Huawei since U.S. restrictions took effect in 2020. Samsung has not commented publicly. Neither BIS nor the South Korean government has issued a public statement. Seoul Economic Daily highlights the precedent: TSMC faced a U.S. investigation in 2024 after one of its chips appeared in a Huawei product.
Enforcement is accelerating. In March, prosecutors charged Super Micro co-founder Yih-Shyan Liaw and two others over an alleged $2.5 billion scheme to route servers built on restricted Nvidia chips to China. On October 1, federal agents arrested Greg Lui, CEO of California-based Earthmade Computer, over more than $300 million in export-controlled servers allegedly shipped via Malaysia and Singapore.
The Information‘s report relies heavily on anonymous dealers, with no disclosed shipment volumes, seizures, company names, or failure-rate data. Yet the incentives it describes are real, and the method was documented over 18 months ago by SemiAnalysis. Until BIS either counts the memory itself or rigorously applies the “dominant function” test to co-packaged chips, the market for processors nobody wants is likely to remain surprisingly healthy—even as China’s appetite for AI memory stays voracious.
