China blocks ByteDance from Nvidia chips in data centers: report

Chinese regulators have blocked ByteDance from using Nvidia chips in new data centers, according to The Information.

ByteDance, the Beijing-based owner of TikTok, reportedly bought more Nvidia chips than any other Chinese company in 2025 to boost computing power for its large user base.

The ban highlights China’s efforts to reduce dependence on US technology amid tighter US export controls on advanced semiconductors.

In August, Chinese authorities told local firms to stop new orders of Nvidia AI chips and encouraged them to use domestic processors, Bloomberg reported.

A Nvidia spokesperson said the regulatory landscape does not allow the company to offer a competitive data center GPU in China, leaving the market to foreign competitors, according to Reuters.

Reuters also reported that the Chinese government has issued guidance requiring new data center projects receiving state funds to use only domestically made AI chips.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

Domestic chip substitution faces a wide software ecosystem gap

Huawei’s Ascend 910D (a Chinese AI accelerator chip) uses 7nm while Nvidia builds on 4N, Taiwan Semiconductor Manufacturing Company’s 5nm-class Extreme Ultraviolet lithography 1. Nvidia wins on developer support with tooling, libraries, and drivers that Ascend lacks 1.
ByteDance plus peers face policy pressure to move new or China-based AI to chips like Ascend 910B (another Ascend variant), which lands between Nvidia’s A100 and H100 GPUs 2. Software compatibility creates the bottleneck.
Ascend’s PyTorch (an open-source AI framework) adapter needs Compute Architecture for Neural Networks (CANN) plus architecture-specific settings. Maintainers run daily tasks to fix issues within ~48 hours 3, so production at ByteDance scale can carry higher engineering risk and maintenance than Nvidia’s Compute Unified Device Architecture (CUDA) ecosystem.

Migration tooling vendors can serve China’s transition

Non-Chinese developer-tool firms can build PyTorch-to-CANN migration frameworks that automate the porting process 4 5. Teams write Dockerfiles (container build files), set up proxy configuration (network access settings), and change device abstractions (code that targets different accelerators).
Cloud providers plus enterprise AI platform teams that manage AI infrastructure can offer dual-stack Machine Learning Operations (MLOps) platforms that abstract hardware differences. Chinese clients keep model portability while meeting domestic chip mandates.
Systems integrators (firms that assemble and operate complex IT systems) can ship pre-configured training environments that combine Ascend Neural Processing Unit (NPU) support with model versioning and testing pipelines 5. This setup fits compatibility and maintenance needs raised by the repository’s ~48-hour issue-resolution target 3.