China issues GPU vouchers for LLM training

A new measure to circumvent the limited access to high-end GPU in China.

China needs more compute power, particularly GPU power which is necessary for AI computations. It has become a significant bottleneck for China’s AI ambitions. Since the U.S. government has banned NVIDIA from exporting its high-end GPUs and accelerators to China, AI companies in the country were struggling to find a solution that would not put them in behind in development. A new solution for AI startups in China are computing vouchers, which can be used for inference and Large Language Model (LLM) training. This was a necessary step to overcome the rising costs of AI-related tasks.

While gamers still have access to plenty of options, or even special models that meet the export criteria, the AI accelerators and General Purpose GPUs equipped with large memory sets and offering substantial boost for inference workloads cannot be exported. This also includes AMD solutions, whose Instinct MI300 processor is banned from shipping to China as well, according to Bloomberg .

The Chinese authorities in major cities like Shanghai want their local AI startups to mitigate the problems associated with a lack of access to powerful computing power. They are offering special vouchers for computing that are worth around $140K to $280K and allow these startups to buy rent data-centers for their training.

Those vouchers can be used through major tech companies in China such as Alibaba, Tencent or ByteDance, who had to prioritize datacenter rentals to internal market. Such an idea will no doubt help smaller companies, but it will not change the problem, which is lack of access to AI accelerators.

In the end, in exchange for using such vouchers or government subsides, the applications and LLM weights that are computed must be first approved for public use.

Source: Financial Times via BBN