Apparently, Alibaba has created an AI accelerator in response to mounting pressure from Beijing to reduce the country’s need on Nvidia GPUs. The latest chip from the e-commerce behemoth, which was first revealed by the Wall Street Journal on Friday, is expressly designed to serve models rather than train them—a process known as AI inference.
For some time now, T-Heat subsidiary of Alibaba has focused its attention on researching AI silicon. Alibaba announced the launch of Hanguang 800 in 2019. This hardware part was developed by Alibaba mainly for conventional machine learning models like ResNet, rather than the huge language as well as diffusion models used by AI chatbots along with picture producers today, as compared to more advanced CPUs like Nvidia and AMD.
According to reports, the new chip will be able to manage a wider range of workloads. In April, Alibaba unveiled its Qwen3 family, establishing itself as one of the most prominent open model developers. Therefore, its primary attention to inference does not come as surprise. Serving models provides a perfect place for people to start making the transition to homegrown hardware considering that it usually calls for fewer resources than instructing them. For the upcoming several years, Alibaba will probably continue to keep on training its models utilizing Nvidia accelerators.
According to the Journal, AI chip of Alibaba will work with software platform of Nvidia, enabling engineers to reuse pre-existing code, in contrast to Huawei’s Ascend line of NPUs. This is uncommon and not needed for inference, although it may appear like CUDA, Nvidia’s low-level programming language for GPUs.
Alibaba is most likely aiming for higher-level abstraction layers that offer a programming interface that is independent of hardware, such as PyTorch or TensorFlow. Despite the fact that many of these edge cases have been addressed by projects like Triton, there is still a lot of PyTorch code that utilizes libraries created specifically for Nvidia hardware.
Unfortunately, since there are US export ban on semiconductor technology, the chips might need to be manufactured domestically. These US export bans hinder many Chinese businesses from doing business with TSMC or Samsung Electronics,
If we had to guess, China’s Semiconductor Manufacturing International Co. (SMIC) would be the company assigned to fabbing the chip, although the paper is silent on the subject. In addition, SMIC is the manufacturing facility which manufactures Huawei’s Ascend series of NPUs.
Local chip programs of China suffer many more challenges than just manufacturing problems. Large amounts of rapid memory are necessary for AI accelerators. This usually refers to high bandwidth memory (HBM), which is likewise constrained as a result of the US-China trade war. The distribution of HBM2e along with newer models in banned by China, except when they already have access to a processor.
Therefore, unless Chinese memory vendors are prepared to bridge the gap, Alibaba’s chips will either rely on slower GDDR or LPDDR memory, current inventories of HBM3 and HBM2e, or older HBM2, which isn’t constrained.
The introduction of the local silicon overlaps with the ambitions of Chinese government to discourage the IT giants of the region from implementing H20 accelerators of Nvidia and to minimize the worries regarding remote kill switches as well as backdoors. Nvidia, that recently obtained the green light to continue exporting H20s to China yet again, has denied the existence of any such features.
Even though Nvidia was officially provided the go ahead to resume chip shipments, as we reported earlier this week, the company is not expecting to generate any revenues in the region this quarter as it waits for Uncle Sam to get through the complex bureaucracy needed to impose a 15 percent export tax on AI chips headed for China.
Many of AI executives of China continue to search for replacements regardless of Nvidia’s anticipated return to the Middle Kingdom. DeepSeek revised its groundbreaking models last month so they could operate on a new generation of locally produced silicon.
Even though, the supplier of chips was not made public by DeepSeek, it has presumably failed to change its model training to Huawei’s Ascend accelerators.
There are other companies attempting to reduce China’s dependency on Western silicon besides Alibaba and Huawei. Tencent-backed startup Enflame was working on a new AI processor, the L600, last month, according to an article in EE Times China. The chip would have 144GB of on-chip memory and 3.6Tb/s of bandwidth.
In the meantime, MetaX has shown the C600, which will have 144GB of HBM3e. On the other hand, there is indication that at the moment HBM3e stockpiles may be restricting chip production.
The Siyuan 690, a homegrown accelerator being developed by Cambricon, another AI chip contender perceived by some as China’s Nvidia, is anticipated to outperform Nvidia’s now three-year-old H100 accelerators.
