Thanks to glaring loop holes in US export controls, about 2,300 of Nvidia’s most advanced Blackwell AI chips — which are barred from being sold to China — inside a tall, windowless building in Indonesia’s capital are ready to do work for a Chinese AI company. This data center got its chips through a chain of deals across several countries. A company that arranged the transaction is a subsidiary of a Chinese business on a US trade blacklist.
There’s no evidence to suggest the deals violated US law despite American rules intended to stop China from accessing the tech industry’s most coveted hardware. So Chinese companies and organizations are still using some of Nvidia’s most advanced products.
Some bring the chips physically into China using middlemen. Another increasingly popular workaround, which has been employed in Australia and Malaysia, is renting computing power abroad and bringing data out of China and back — sometimes by packing suitcases with hard drives.
In the Indonesia case, US technology is being made available to a 
company through these four steps.
Nvidia sells chips to a partly -owned business partner
Nvidia sells many chips to partners that build servers for AI computing. One partner that makes these servers is Silicon Valley-based Aivres. Its parent is one-third-owned by Inspur, which has been added to a US national-security trade blacklist since 2023 for its involvement in military supercomputing.
Nvidia can’t do business with Inspur or some subsidiaries added to the blacklist this year. But the rules don’t cover US-based subsidiaries such as Aivres, which has close ties to Nvidia and co-sponsored Nvidia’s most important corporate event this year.
The partner finds a buyer in Indonesia for servers with Nvidia chips
In mid-2024, Aivres began negotiating a deal with the cloud-computing business of Indonesian telecom services provider Indosat Ooredoo Hutchison, which bought 32 Nvidia GB200 server racks from Aivres for about $100 million. Each server rack contains 72 of Nvidia’s leading-edge Blackwell chips.
At roughly 2,300 chips, the processing power is small compared with the tens or hundreds of thousands chips typically required to develop the most advanced AI systems.
Indosat is a joint venture between Qatari telecom company Ooredoo and Hong Kong-based conglomerate CK Hutchison.
The Indonesian cloud-computing business lands a Chinese customer
Indosat bought the servers only after securing a client with Aivres’s help. That customer is Shanghai-based AI startup INF Tech (无限光年).
INF was founded by Qi Yuan (漆远) aka Alan Qi, a professor who heads an AI institute at Fudan University in Shanghai. Fudan representatives were part of the deal negotiations, but ultimately INF signed the contract for the computing power.
(Qi was previously a VP at Alibaba’s ANT Group. Before that, he was a tenured professor at Purdue University’s Department of Computer Science and Department of Statistics. Qi received his PhD from MIT in 2004 and worked as a research fellow at MIT CSAIL before joining Purdue in 2007. He received an MS from the University of Maryland, College Park, and the Chinese Academy of Sciences. He received his BS from Huazhong University of Science and Technology in Wuhan.)
The Chinese customer aims to use the chips for finance and health AI
As of October, the servers had been delivered and were being installed. When operational, they will focus on helping INF train AI for financial applications and science research, such as drug discovery.
wsj.com/tech/ai/china-
Quote
Byron Wan
@Byron_Wan
On a June night in 2024, Nvidia CEO Jensen Huang held court with several of his company’s major Asian customers at a bar in Taipei. Next to him was Huang Xiaole (黄小乐) aka Alice Huang, an executive of Megaspeed, a shady Singapore-based data center company.
Nvidia’s compliance
Show more
