Chinese tech giants reveal how they’re dealing with U.S. chip curbs to stay in the AI race

Tencent and Baidu, two of China’s largest technology companies, revealed how they’re keeping in the global artificial intelligence race even as the U.S. tightens some curbs on key semiconductors.

The business’ methods include stockpiling chips, making AI models more efficient and even using homegrown semiconductors.

While the administration of U.S. President Donald Trump scrapped one controversial Biden-era chip rule, it still tightened exports of some semiconductors from companies including Nvidia and AMD in April.

Big names in the sector addressed the issue during their latest earnings conference calls.

Martin Lau, president of Tencent — the operator of China’s biggest messaging app WeChat — said his company has a “pretty strong stockpile” of chips that it has previously purchased. He was referring to graphics processing units (GPUs), a type of semiconductor that has become the gold standard for training huge AI models.

These models require powerful computing power supplied by GPUs to process high volumes of data.

But, Lau said, contrary to American companies’ belief that GPU clusters need to expand to create more advanced AI, Tencent is able to achieve good training results with a smaller group of such chips.

“That actually sort of helped us to look at our existing inventory of high-end chips and say, we should have enough high-end chips to continue our training of models for a few more generations going forward,” Lau said.

Regarding inferencing — the process of actually carrying out an AI task rather than just training — Lau said Tencent is using “software optimization” to improve efficiency, in order to deploy the same amount of GPUs to execute a particular function.

Lau added the company is also looking into using smaller models that don’t require such large computing power. Tencent also said it can make use of custom-designed chips and semiconductors currently available in China.

“I think there are a lot of ways [in] which we can fulfill the expanding and growing inference needs, and we just need to sort of keep exploring these venues and spend probably more time on the software side, rather than just brute force buying GPUs,” Lau said.