Alibaba aims to lessen its dependence on Nvidia for artificial intelligence processing
Alibaba and the Pursuit of Domestic AI Accelerators
In a move to bolster its AI capabilities, Alibaba has returned its market-rattling models to run on a new generation of domestic silicon, but has encountered challenges in transitioning model training to Huawei's Ascend accelerators.
The tech giant's efforts to develop homegrown AI hardware started in 2019 with the introduction of the Hanguang 800, primarily aimed at conventional machine learning models. Alibaba's AI accelerator, manufactured by Semiconductor Manufacturing International Corporation (SMIC), is produced by Alibaba itself.
However, the new chip is expected to handle a more diverse set of workloads compared to the Hanguang 800, as Alibaba is also reportedly developing an AI accelerator aimed at AI inference. The details of this new accelerator are not yet publicly available.
China's Semiconductor Manufacturing International Co. (SMIC) is a possible candidate for manufacturing Alibaba's AI accelerator, given the US export controls on semiconductor tech that have pushed Alibaba to build its AI accelerator domestically.
Meanwhile, the Chinese government is reportedly pressuring tech companies not to use Nvidia's H20 accelerators, citing concerns about backdoors and remote kill switches. Tencent-backed startup Enflame is developing a new AI chip called the L600, featuring 144GB of on-chip memory capable of 3.6Tb/s of bandwidth. However, the L600 is not compatible with Nvidia's low-level programming language CUDA.
Cambricon is another Chinese company working on a homegrown accelerator called the Siyuan 690, widely expected to outperform Nvidia's now three-year-old H100 accelerators. AI accelerators rely on large quantities of fast memory, but due to US-China trade restrictions, high bandwidth memory (HBM) is either restricted or not available in China. As a result, Alibaba's new chip will likely use slower GDDR or LPDDR memory, existing stockpiles of HBM3 and HBM2e, or older HBM2, until Chinese memory vendors are ready to fill the void.
Despite these developments, Alibaba is likely to continue using Nvidia accelerators for model training for the foreseeable future. Nvidia has been cleared to resume shipments of H20s to China, but the company doesn't expect to realize revenues in the region this quarter due to a 15 percent export tax on AI chips. MetaX has unveiled its C600, featuring 144GB of HBM3e, but chip production may be limited by existing stockpiles of HBM3e.
In a significant stride towards open models, Alibaba has become one of the leading developers of open models with its Qwen3 family launched in April. The race for domestic AI accelerators continues, with Alibaba, Cambricon, Enflame, and MetaX leading the charge in China.
