How a Chinese AI firm tapped Nvidia’s top chips via Indonesia

Representative image

In a quiet corner of South Jakarta, inside a new, windowless data centre squeezed between a private school and luxury apartments, sit around 2,300 of Nvidia’s most advanced AI chips. Officially, they are part of an Indonesian telecom project. In practice, they will power artificial intelligence models for a Shanghai startup, highlighting how Chinese firms can still reach US technology without importing a single chip, the Wall Street Journal reported.

The chain that put those processors in Jakarta has spanned California, Shanghai and multiple regulatory grey zones. It has also underscored the tension between Washington’s desire to keep China away from cutting edge AI hardware and the tech industry’s push to keep selling its most valuable products.

A supply chain built around the rules

President Trump has made it clear that he does not want Nvidia’s leading AI chips sold into China. Since 2022, direct exports of the most powerful US semiconductors to the Chinese market have been restricted on national security grounds. At the same time, Nvidia remains the dominant player in AI chips and wants as much of the world as possible, short of blacklisted entities, running on its hardware.

To square that circle, the company leans on partners that assemble servers and sell computing systems. One of those partners is Aivres, based in Silicon Valley. Its parent company is one third owned by Inspur, a big Chinese tech firm placed on a US trade blacklist for military supercomputing work. Nvidia cannot sell to Inspur and some of its affiliates, but US based subsidiaries such as Aivres are not covered by those specific restrictions. Aivres remains a close partner and even co-sponsored Nvidia’s flagship corporate event this year.

From California chips to Jakarta racks

In mid 2024, Aivres began negotiating with Indosat Ooredoo Hutchison, a major Indonesian telecom operator that is pushing into cloud computing. Indosat agreed to buy 32 server racks equipped with Nvidia’s new GB200 systems, at a price of about 100 million dollars. Each rack holds 72 Blackwell chips, giving the Jakarta site roughly 2,300 high end processors in total. That is small compared with the vast clusters used by the very largest AI models, but it is still serious computing power.

Indosat, a joint venture between Qatar’s Ooredoo and Hong Kong based CK Hutchison, only moved ahead after it had a customer lined up. That customer did not sit in Jakarta, but in Shanghai.

Shanghai startup rents power, not hardware

With Aivres acting as a bridge, Indosat struck a deal with INF Tech, a Shanghai AI startup founded in 2021 by Qi Yuan, a Chinese born American citizen with a PhD from MIT and early machine learning experience at Alibaba. INF focuses on financial and healthcare AI products and is looking to expand beyond China, working with data centre operators across Southeast Asia.

In Jakarta, INF is not buying the chips directly or moving them into China. Instead, it is renting computing time on Indosat’s servers, sending data in, training AI models remotely and pulling the results back out. INF says it does no work with military applications and complies with US export controls. Indosat stresses that the Chinese company has no physical access to the hardware and that the computing power will also support AI services tailored for Indonesia and the wider region.

A grey zone in export controls

Lawyers who follow export rules say that, under current Trump administration policy, this setup fits within the law. The chips are installed in Indonesia, a country not under the strictest controls, and there is no evidence they are being used for military intelligence or weapons design. The Chinese customer is paying for cloud services, not for the chips themselves.

Even so, some former and current US national security officials worry that arrangements like this undercut the intent of the rules. They argue that, under Beijing’s policy of civil military fusion, Chinese firms that build commercial AI systems today could be tapped to support the People’s Liberation Army tomorrow. In their view, US regulators should scrutinise these kinds of transactions more closely, especially when they involve companies connected to the trade blacklist.

In its final days, the Biden administration drafted a rule that would have tightened controls on advanced chip sales to countries such as Indonesia, giving Washington more power to review who was ultimately using the technology. The Trump administration later said it would not enforce that rule, effectively leaving more of the risk assessment to companies themselves.

A preview of future battles over AI access

Nvidia says its compliance teams vet partners before they receive products and that it supports the administration’s broader goal of securing US AI leadership while keeping innovation and jobs at home. Critics counter that, without tougher, clearer rules, powerful chips will keep finding their way to Chinese users through cloud deals and overseas data centres.

For Indosat, the Jakarta project is a bet on AI as a new growth engine beyond traditional telecom services. For INF, it is a way to train models on top tier hardware as it builds a regional footprint. For Washington and Beijing, it is another small front in a larger struggle over who controls the computing power that will shape the next generation of technology.