The GPU Black Market That Washington Can't Shut Down

U.S. export controls on advanced AI GPUs have not stopped billions of dollars’ worth of chips from reaching China. Shell companies, Southeast Asian middlemen, neocloud operators, brokers and data‑center developers routinely route banned NVIDIA H100/H200-class devices through informal markets and regional warehouses, where prices sit near U.S. list levels - a sign that supply remains plentiful rather than scarce. Investigations and industry interviews show open tacit collusion: sellers, brokers and some logistics providers routinely look the other way. Multiple documented networks have moved transactions in the hundreds of millions, online listings ship from China, and prosecutors have filed major indictments alleging large‑scale diversions of servers to Chinese operators.

Those chips matter because frontier model training depends on thousands of high‑end GPUs running in parallel, and virtually all such devices and key fabrication tools trace back to a narrow supply chain - NVIDIA, TSMC and ASML - creating a chokepoint Washington has tried to exploit. Since October 2022 regulators have iteratively tightened controls (banning early chips, closing workarounds, restricting manufacturing tools and memory), expanded jurisdiction via the Foreign Direct Product Rule, and proposed multilevel frameworks that were later rescinded. Enforcement gaps, engineering workarounds, and China’s mix of domestic production and remote cloud access mean controls remain leaky, sustaining dual‑use risks and broad global spillovers.