$NVDA Zephyr's reading of China's draft is right, and the annex backs it almost word for word. The AI chip entry covers "chip architecture, self-developed instruction sets, AI IP cores and the supporting software stack," the same list he gives. China keeps selling the products and puts a license on the know-how behind them. The model entry is where DeepSeek comes in. Model weights, checkpoints, source code, and training and inference techniques need a license if any one of three tests is met: 10^26 FLOPs of training compute, 2 trillion parameters, or a top-10 place in public global rankings, which the draft says include usage. DeepSeek's V4.1-Flash-1017, released the day after the draft, has 763 billion parameters and no published compute figure. The V4.1 Flash it replaces was No. 1 by tokens on OpenRouter last week. On open source, the draft leaves the door open for now. Technology already public through open source or papers is out of scope, and open-sourcing these items gets an "automatic license." DeepSeek put 1017's weights on Hugging Face under an MIT license on Saturday. Cheryl Wu's framing fits: the state now holds a switch, and today it's on. Taiwan is tightening from the other side. According to Bloomberg, its economy minister said AI servers and advanced chips will be added to export controls for China, with unlicensed shipments becoming a crime. There's no start date yet. DeepSeek's own Ascend tests in September show why that matters. Its matrix kernels reach 99.8% of a single Ascend 950DT's limit, but all-to-all bandwidth drops from 374 GB/s at 8 cards to 317 GB/s at 128. The single chip is close to done, and linking thousands of them is still the gap. China's draft lists cluster interconnect at 800 GB/s or more as restricted technology. What China has is price and reach. 1017 lists at $1.20 per million output tokens at peak, against $20 for Claude Opus 5.5. Chinese models held 7 of OpenRouter's top 10 last week and 84% of the tokens those ten processed, and CXMT took 9.5% of global DRAM revenue in 2Q26. Our read on the way out: ship open models that stay under the size lines, keep the frontier know-how at home, and build compute where chips can still be bought. The draft's automatic-license list already reads that way, since open-sourcing, Chinese-owned compute centers abroad and taping out chips at overseas fabs are all fast-tracked. The cost lands on Zephyr's ecosystem point. If the usage test survives into the final version, China's most-used open model would sit above a control line, and each open release would depend on the state keeping the switch on. Comments close Nov. 15. The clause we'd watch is the usage ranking in the third test.