Chip export controls fraying

The United States is trying to use artificial intelligence chips the way older powers used oil pipelines: approve the flow to friends, block it to rivals, and keep the chokepoint in Washington. This week, that strategy looked shaky at both ends — the approvals are slowing down, and some banned hardware appears to be getting through anyway. (bloomberg.com, bloomberglaw.com, bloomberg.com) The slowdown is inside the Bureau of Industry and Security, the Commerce Department office that signs off on sensitive exports. Bloomberg reported that the bureau has lost dozens of experienced employees in the past year, with nearly 20% turnover among rulemaking and licensing staff. (bloomberg.com, ttnews.com) That matters because Nvidia and other companies cannot just load top-end artificial intelligence chips onto a plane and send them abroad. Individual export licenses for foreign buyers are now taking months, and Bloomberg said the pileup has created billions of dollars in backlogged shipments, including sales meant for United States allies. (bloomberg.com, ttnews.com) The policy itself has also been moving around. The Trump administration scrapped the Biden-era “artificial intelligence diffusion” rule in 2025, then floated a replacement that would give Washington a hand in a much wider range of overseas chip sales, including some deals in the Middle East. (techcrunch.com, ttnews.com) So companies are dealing with two kinds of uncertainty at once. The written rules are still being reworked, and the office enforcing the rules is short on people and pulling more decisions up to senior officials, which Bloomberg said has slowed the system further. (bloomberg.com) Then came the leak on the other side of the wall. On April 10, Bloomberg reported that a Shenzhen company called Sharetronic disclosed invoices showing sales of 276 Super Micro systems loaded with Nvidia H100 or H200 processors, worth 632 million yuan, or about $92 million. (bloomberg.com, theedgesingapore.com) Those Nvidia H100 and H200 chips have been restricted for China without Washington’s permission since 2022. The invoices Bloomberg reviewed were dated May and June 2025, which means the hardware appears to have reached a Chinese buyer well after the restrictions were already in place. (bloomberg.com, theedgesingapore.com) The timing made the disclosure hit harder. It surfaced hours after federal prosecutors charged Super Micro co-founder Yih-Shyan “Wally” Liaw with helping divert Nvidia-powered servers to China, and Bloomberg Law reported that he pleaded not guilty in Manhattan. (bloomberglaw.com, bloomberg.com, cnbc.com) Sharetronic said it follows the rules and has no business relationship with Super Micro. But the invoices Bloomberg described listed Super Micro systems, which is exactly the kind of paper trail export controls are supposed to stop before the machines arrive, not after they show up in procurement records. (bloomberg.com) Washington’s problem is that chip controls only work if the gate is both narrow and fast. A slow gate frustrates allies and sellers, and a leaky gate still lets restricted computing power reach China. (bloomberg.com, bloomberg.com) The United States still controls the most valuable artificial intelligence chips and the legal permissions around them. But this week’s picture was a bottleneck in Washington, a backlog worth billions, and 276 restricted servers on Chinese invoices — which is a much messier version of control than the strategy was supposed to deliver. (bloomberg.com, ttnews.com, bloomberg.com)