Alibaba fires up AI engine on homemade chips | Computer hardware components and their functions | Is gpu a graphics card | Hardware definition and examples | Turtles AI

Alibaba is testing a new AI chip dedicated to inference, manufactured in China and compatible with PyTorch/TensorFlow, to reduce its dependence on Nvidia. This is part of a broader effort toward domestic silicon and technological self-determination.

Key points:

• New AI accelerator for inference developed by Alibaba, intended for the domestic cloud
• Compatibility with Nvidia software (PyTorch/TensorFlow), simplifying code porting
• Chip produced domestically in China, likely by SMIC, to circumvent US restrictions
• It joins other domestic projects (Huawei Ascend, Enflame L600, Cambricon Siyuan 690)

Alibaba has launched an AI chip aimed exclusively at inference, no longer based on Western processors like Nvidia, but developed internally and intended to run pre-trained models like Qwen 3, launched in April. The T-Heat unit, which created the Hanguang 800 in 2019, now offers a more flexible architecture, capable of handling modern workloads such as language models or image generators, while maintaining compatibility with PyTorch and TensorFlow to avoid disrupting the existing software ecosystem.

Production, which was not entrusted to TSMC or Samsung due to US controls, has been entrusted to a Chinese factory, presumably SMIC, to ensure technological autonomy and circumvent the ban on importing advanced US chips. The project envisions the use of alternative memory: GDDR, LPDDR, or legacy HBM2 stockpiles, since HBM2e and subsequent versions are subject to import restrictions not associated with processors.

The new chip is being tested just as the Chinese government is pushing major tech companies to reduce or eliminate the use of Nvidia’s H20 accelerators, which have been accused, without official evidence, of possible backdoors or remote kill switches. Nvidia, which has been given the green light to resume shipments of the H20 model, has firmly denied the existence of hidden control capabilities.

Alibaba’s cloud posted 26% year-over-year growth in the second quarter of 2025, significantly contributing to the group’s overall results and reinforcing its billion-dollar investment strategy in AI infrastructure through 2028. Meanwhile, other Chinese companies such as Enflame (with the integrated 144GB L600 chip), MetaX (C600 with HBM3e), and Cambricon (Siyuan 690) are also developing local accelerators with performance that aims to surpass (or approach) that of the latest Nvidia GPUs.

This dual approach—continuing to use Nvidia accelerators for training while leveraging domestic chips for inference—is in line with the strategy of many Chinese companies operating in a complex geopolitical environment, balancing business continuity and national security.

Alibaba is working to deliver AI models on domestic silicon, bridging international software ecosystems and domestic hardware.