Alibaba, ByteDance train AI models to access Nvidia chips: report

Chinese tech firms are reportedly training AI models overseas to gain access to Nvidia chips, according to the Financial Times.

Alibaba and ByteDance are among the companies said to be using data centers in Southeast Asia for large language model development.

This move comes as US restrictions limit Chinese firms’ access to advanced technology, including high-end AI chips from Nvidia.

The Financial Times cited unnamed sources with direct knowledge of the situation.

Reuters could not independently verify the report.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

The U.S. Export Administration Regulations (EAR), the main U.S. export-control rules, do not clearly cover remote access to controlled items in the cloud such as advanced chips 1.
From 2009 to 2014, the Commerce Department’s Bureau of Industry and Security (BIS) said cloud services fell outside EAR because no physical goods or software or technology leaves the country 1.
This gap lets Chinese firms train AI models by tapping cloud servers that use controlled advanced semiconductor chips without triggering current export limits 1.
Bipartisan Remote Access Security Act passed the House in September 2024 and would let BIS regulate remote access when it risks national security 1.
Until it becomes law and rules are clear, some Chinese companies are using Southeast Asian data centers to reach Nvidia chips for training 1.

In Singapore and across Southeast Asia, cloud operators see near-term demand from Chinese AI teams that need offshore GPU capacity.
BIS now requires licenses for advanced computing items worldwide, with compliance due by May 15, 2025 2.
Providers should vet customers and end use on remotely accessible controlled items, then lock down access to get ready for the Remote Access Security Act 1.
Investors can favor Southeast Asian hosts that offer Nvidia H100 capacity or, where available, H200, since they can catch spillover from firms under U.S. curbs. (Nvidia’s H100 and H200 are high-end AI accelerator chips used to train large language models.)
The current gap gives a short window before rules tighten, so early deals with operators that run strong compliance programs can be useful.