As the U.S. government scrambles to pull the semiconductor rug out from under China’s AI ecosystem, some AI researchers in the country are shrugging it off.
One reason: AI accelerators.
“Nvidia will lose a lot of [market share] in their high-end GPU graphic cards, but a lot of the startups in China making those AI acceleration cards will get orders,” according to an AI researcher and professor at a prestigious scientific university in Beijing who I spoke with via video chat this week. (The researcher asked not to be named for fear of political retribution.)
In China, where AI engineers and other developers are accustomed to technical workarounds to circumvent censors and other blockades, there may be some wiggle room to counteract U.S. export controls by building AI accelerator cards intended for more specific tasks, the researcher said.
Nvidia’s strengths have been in “very general-purpose GPUs that can handle [many] types of computations, like for gaming and computing.”
But building accelerators is not so difficult, the researcher said. “Just making accelerating cards is a lot easier because you just need to handle very few specific types of computations. So when the U.S. government shut Nvidia out of China, it actually [benefited] those startups in China.”
And as giants in China like Huawei open new chip fabs there, people familiar with the nuts and bolts of hardware for AI say there’s still lots of room for innovation in AI accelerators.
A version of this story first appeared in the Protocol Enterprise newsletter: sign up here.
Climate TRACE's data show emissions at the facility-level.Image: Climate TRACE