WSJ just dropped another strong investigation on how China is exploiting loopholes in U.S. export controls. To summarize what is happening and the loopholes that China is exploiting:

1️⃣ Nvidia ships Blackwell AI chips to Aivres Systems, a U.S.-based company that is 100% owned by China-based Inspur Electronic, which is on the U.S. Entity List. (This is 100% legal, because the only place that Chinese companies face no restrictions on buying AI chips is if they buy them inside the United States)

2️⃣ Avires ships these AI chips to an Indonesian cloud provider, Indosat, which is majority-owned by a joint-venture between a Qatari company and a Chinese company. (This is 100% legal, because there are no restrictions on AI chip shipments to Indonesia or most other countries unless the company is majority owned by a China-based company, which Indosat technically is not)

3️⃣ Avires identifies a China-based client for Indosat, INF Tech, which is banned from purchasing AI chips but is interested in renting access to these chips from Indosat. INF Tech rents access to the Indosat-owned chips that Avires originally purchased, and uses them to develop and run advanced AI systems. (This is 100% legal, because there are virtually no restrictions on China-based companies renting access to banned AI chips from companies located outside China, including from U.S. cloud providers, even though the export of the chips themselves to China-based companies is banned)

All of these loopholes can and should be closed. Commerce should: (1) ban the sale of AI chips to Chinese companies located in the United States by issuing an ICTS regulation prohibiting these transactions; (2) require a US export license for any sale of large numbers of AI chips to untrusted companies in Southeast Asia at a minimum, and potentially to a larger set of countries; and (3) close the remote access loophole that allows Chinese companies to circumvent export controls by renting access to AI chips instead of buying them, by prohibiting the use of AI chips by China-based companies to train or provide inference for any advanced AI models.

If we’re serious about controlling compute as a strategic resource, all of these steps are necessary.

https://t.co/VxQNN2uCeT