Issue #031 — 2026-09-21

C4ADS traced three routes moving export-restricted Nvidia AI chips into China; reports tie a $4.6 billion Nvidia importer to the network. Export rules fail at the border when the buyer chain stays legal on paper. Selling GPUs or compute? Know your end buyer.

Reuters reports China is slowing humanoid robot IPO approvals, judging valuations have outrun revenue, after Unitree's volatile debut. IPOs priced robots that customers have not yet paid for. Read order books, not IPO filings, for robot demand.

India's MeitY Secretary urged Microsoft to build future data centres closer to power generation hubs, Moneycontrol reports. Siting near generation skips the transmission queue that kills projects. Pricing Indian compute? Map it against generation hubs.

Samsung began trial production of Tesla's AI5 chip on a 2nm process at its Taylor, Texas fab, per Seoul Economic Daily. Tesla now has two advanced fabs for one chip, weakening TSMC's pricing power. Watch Samsung yields before counting Tesla chips in 2027.

Alibaba released Qwen-Image-2.1, a 7B image model with 2048x2048 output; AI Weekly reports the license moved from Apache to research-only. Builders need licenses they can ship under; research-only ends that. Check the license before swapping it into a product.