Sources: ByteDance has been renting Nvidia's GPUs from Oracle to bypass US sanctions; China Telecom and other Chinese companies are using the same loophole
The U.S. government forbids Nvidia from selling some of its most advanced artificial intelligence chips to customers in China. X: @amir X: Amir Efrati / @amir : America: you can't buy our Nvidia chips, China. China: Ok, we will come get them where you live. https://www.theinformation.com/ ... @anissagardizy8 $nvda [image]
Context & Ripple Effects
This report sits after evidence that Chinese organizations were still obtaining limited quantities of restricted Nvidia hardware through other channels, including small-batch chip purchases by Chinese entities. It identifies remote cloud access as a distinct route: the relevant control point is not only the chip sale, but also who can use deployed compute.
The route matters because later coverage found that renting A100 and H100 cloud capacity could be cheaper in China than in the U.S. Export restrictions on shipments therefore do not necessarily translate into equivalent limits on access to advanced AI infrastructure.
First-order effects
- ByteDance, China Telecom, and other reported customers can access advanced Nvidia GPU capacity without directly purchasing the restricted hardware.
- Oracle becomes a critical intermediary whose cloud-rental practices determine whether the reported access route remains available; Nvidia hardware remains usable for Chinese AI workloads through remote access.
Second-order effects
- The disclosure raises the compliance burden on cloud providers: customer screening and controls over remote GPU use become as consequential as restrictions on physical chip shipments.
- Chinese demand can shift toward offshore or third-party cloud capacity when direct procurement is blocked, preserving demand for Nvidia-based compute while changing where and how it is delivered.
Third-order effects
- If remote rental remains outside or harder to enforce under export controls, policy effectiveness will increasingly depend on governing compute access rather than only hardware transactions.
- The longer contest may move toward an access-and-enforcement regime in which compliant chip variants, cloud controls, and domestic alternatives compete; later reports of Nvidia adapting chips for China-compliant sales show the pressure to find permitted channels.
The trend: This is one data point in the shift from controlling AI chips at the border to controlling access to the compute those chips provide.