Nobody Owns the Demand Side

The H200 story isn't an export control failure. It's a demand-side enforcement failure, and the difference matters because one is structurally hard and the other is a job no agency has been assigned.

Bloomberg's reporting this week, based on a review of procurement records, found at least seven Chinese universities with military and defense-industry ties actively seeking H200 chips, the most powerful AI processors the U.S. has approved for sale into China. The records aren't leaked. They aren't hidden. They're institutional purchasing documents sitting in public-facing university portals.

The standard read is that this proves the whole export control architecture is theater. China gets what it wants through gray markets, third countries, and shell resellers, so the licensing regime is kabuki. There's real evidence for that view. The Singapore and Malaysia routing lanes are real. The Gulf reseller fronts are real. The H800-to-H100 substitution math is real. If you believe controls have failed structurally, you have a lot of company and a lot of citations.

But this story isn't that story. The H200 sale into China is legal. Nvidia has the license. The chips ship through approved channels. The military-linked universities aren't smuggling. They're filling out purchase orders. The paper trail isn't broken. The paper trail is sitting in plain view in databases that researchers and reporters can read, which is how Bloomberg wrote the piece.

The U.S. export control regime is supply-side architecture. It governs what can be sold, to whom, at what specification. Once a chip is cleared for sale into the country, the enforcement question shifts to the buyer. That's where the system goes dark. Commerce licenses the sale. BIS lists denied entities. Treasury sanctions individuals. None of them is reading the procurement portal at a Chinese aerospace university to check whether the buyer on the PO matches the entity that's allowed to buy.

Who should be? That's the question the Bloomberg piece raises and politely declines to answer. It names the universities. It names the chips. It shows the records, and it stops short of saying which U.S. agency owns the procurement-records gap. That omission is generous to everyone in Washington who'd otherwise have to claim the work. Commerce points at the licensing process. State points at allied coordination. Treasury points at sanctions designations. The demand-side verification gap sits in the middle of the room and nobody picks it up.

This isn't even a hard problem. The records exist. They're searchable. They're the same data a Bloomberg reporter pulled together on a deadline. A real demand-side enforcement regime would assign one agency the task of cross-referencing buyer entities against end-use restrictions, with the authority to claw back licenses when a discrepancy surfaces. It would treat the procurement portal as the enforcement surface it already is. None of this requires a new statute. It requires somebody being told the work is theirs.

Prediction: nobody builds that regime in 2026. Commerce will issue another rule tightening supply-side specifications. H200 shipments will keep landing on campuses with PLA-affiliated labs. And when the next Bloomberg story runs in eighteen months, the framing will still be "export controls aren't working," because naming the actual gap would require naming the agency that was supposed to close it.

The procurement records are right there. Go read them.