Smugglers are reportedly soldering top-tier HBM onto deliberately useless processors to sneak it into China, and the trick isn't new - TechRadar

This HBM smuggling scheme exposes a structural shift most analysts missed: the AI compute bottleneck has migrated from the logic die to the memory stack. Soldering HBM3E onto a worthless ASIC is, in effect, treating memory as a standalone commodity — a move that physically deconstructs the "integrated circuit" classification underpinning BIS export controls. The supply-chain impact is immediate. SK Hynix and Samsung already run HBM3E fabs near capacity; a 10-15% gray-market leakage directly elongates lead times for NVIDIA H200 and AMD MI300X boards. More critically, once domestic Chinese AI chip designers secure HBM supply, the performance ceiling of their in-house accelerators rises materially, compressing the "de-NVIDIA" timeline by six to nine months. On compliance, the ECCN framework targets finished devices. A dummy-die-plus-HBM assembly likely falls below the control threshold, a gap BIS will close with HBM-specific regulations within two quarters. SK Hynix and Samsung should brace for DOJ scrutiny; quarterly compliance overhead could climb $200-300M. Over the next 12-24 months: CXMT and YMTC will accelerate HBM yield ramp-up; black-market premiums will hold at 30-50%; and the "dead-die packaging" trick will force a new generation of AI-assisted X-ray inspection standards. The control-versus-evasion arms race has now descended from the chip level to the solder-ball level.