The Export Ban Didn't Stop the GPU Trade. It Gave It a Supply Chain.
American export controls on high-end GPUs are written to stop the sale, not the purchase, and that asymmetry is the entire story. Between Hong Kong's retail counters and Shenzhen's wholesale floors, restricted cards move through hand-carried units, factory rejects and small trading companies, and the people moving them treat it as ordinary commerce. The policy's clearest output so far is a price structure, not a shortage.
That is not a comment on whether the rules are wise. It is an observation about where they can physically bite. A control that binds sellers subject to US law has to be enforced at export or in a cooperating transit country. It has no reach inside the market where the hardware ends up, and everyone in that market knows it.
A Rulebook That Regulates the Wrong End
The restriction applies to anyone under US jurisdiction who sells a covered part into China, Hong Kong or Macau without a license. Buying one inside China is not a violation of Chinese law, and no Chinese authority enforces an American export rule. So the entire compliance burden lands on the seller and on intermediaries in countries that choose to cooperate.
The threshold itself has been built from brochure numbers. Commerce's Total Processing Performance score multiplies a peak throughput figure — calculated at the theoretical maximum the manufacturer claims — against operation bit length. Memory capacity was left out of the formula, which is precisely the spec that decides whether a model fits in a machine at all. Vendors then engineered parts to sit under the line: A800, H800, L40S, and later the H20 with 96GB. Each time, the line moved again, after the compliant product already existed. That is a guess-and-check process dressed as a technical standard.
Six Roles, One Pipeline
What gets called a black market behaves like a distribution network with named stations. A source sells a card, sometimes an ordinary user on a classifieds listing with no idea where it is headed. A buyer-operator collects them regionally, often out of a car. A courier hand-carries or ships them across a border. A middleman warehouses them. A fence matches stock to buyers. An optional technician modifies a card before it reaches the user.
The vocabulary from the people doing the work is more revealing than any diagram. One phrase translates as ants moving house, describing the accumulation of many small individual carries. Another — open one eye, close one eye — came up independently from multiple sellers when they were asked whether NVIDIA knows. The geography explains the method: the gap between Hong Kong and Shenzhen is a single customs check, which is why so much of the volume is pedestrian and why nobody involved finds it remarkable.
Factory channels matter as much as couriers. Gaming-class cards are built at partner facilities in volume, and oversight of rejected boards is loose. A board written off as a defect still has salvage value in its GPU and memory packages, and a card that never officially ships leaves no serial number in any database that gets checked. Several unrelated people described the same mechanism, which is more convincing than any single account.
The Repair Shop Is the Real Insurgency
The most interesting operation in the whole chain is not a smuggler. It is a repair shop in Zhengzhou that desolders a GPU, reballs it, and mounts it on a third-party PCB built to carry twice the memory — turning a 24GB RTX 4090 into a 48GB card that NVIDIA does not sell. The shop claims a very high success rate and a turnaround measured in hours rather than weeks.
Read that as a supply response rather than a crime story. When memory capacity is what gates the workload, capacity is what the market builds for itself. Those shops are legitimate businesses operating openly inside China, and they are reacting to a constraint that policy created and cannot police from outside. The same logic explains why demand skews toward memory-heavy parts and away from whatever has the best floating-point figure on the box.
What I Think
I don't think the ban failed because it was porous. I think it failed because it was designed around a measurement that nobody on either side respected, then adjusted after vendors had already built around it. A rule that gets rewritten every time a company finds the gap is not a control. It is a negotiation with a lag, and the lag always favors the company with better engineers.
What bothers me more is the public posture. NVIDIA has dismissed documented smuggling cases as tall tales while separately arguing that the controls were a failure that cost it market share. Both of those can be sincere only if the company's position is that enforcement is simultaneously working and harming it. And the scale figures that circulate — billions per year, per-card spreads quoted by dealers in Shenzhen — come from participants with an interest in the number. Treat all of them as unverified until somebody independent audits the flow rather than interviewing the people profiting from it.
The structure, though, is not in dispute. Factories with loose reject handling, couriers who cross one border, warehouses that feed a wholesale floor, technicians who can double a card's memory in an afternoon, and an end market that treats a controlled part as inventory. You can ban a shipment. You cannot ban a supply chain that has already learned to route around the paperwork.