Loading page…
Leo Wang says alternatives to Nvidia have matured, giving the self-driving company more flexibility—but not yet the same cost-effectiveness.
Listen to this story
Pony.ai is building more flexibility into its AI training as compute supply falls short: it is using more Chinese-made chips while directing engineers to make workloads fit available hardware. CFO Leo Wang says these chips are a more mature alternative than they were a few years ago, but are not yet as cost-effective. The shift gives Pony.ai a contingency if Nvidia export restrictions tighten, though Wang described that scenario as a risk to prepare for, not a change that has already happened.
Wang said engineers use midtier computing resources instead of high-end ones where possible.
He said Chinese companies have been practicing a backup plan for several years, so alternatives are already available.
Wang discussed the shortage and export-restriction scenario in an interview published by TechCrunch on November 15.
Pony.ai is adjusting how it trains AI and using more Chinese-made chips because computing demand already exceeds supply, CFO Leo Wang told TechCrunch. His account describes a backup strategy taking shape under an existing shortage, with one important drawback: alternatives to Nvidia are more mature, but still less cost-effective.
In an interview published November 15, TechCrunch transportation editor Kirsten Korosec asked Wang about the challenges facing China’s self-driving industry and the possibility of tighter Nvidia export restrictions. Wang’s starting point was a constraint Pony.ai already faces: getting enough computing power. He described supply as one of the company’s pain points.
The first part of Pony.ai’s response is to make its AI training workloads more efficient. Wang said engineers are directed to use computing resources carefully, including choosing midtier resources rather than high-end ones where possible. He presented that as a way to work within the supply constraint, not as a blanket replacement of high-end hardware.
The second part changes the source of that computing power. Pony.ai is turning increasingly to domestically produced Chinese chips for AI training, Wang said. He pointed to the attention and development those chips have received in recent years as another option for the company.
Asked what would happen if Nvidia chip export restrictions tightened, Wang said Chinese companies had been practicing a “Plan B” for several years. He described alternatives as already available, rather than something the industry would have to develop from scratch after a policy change. That was his assessment of contingency planning, not an announcement that new restrictions had taken effect.
For Pony.ai, the domestic-chip option offers flexibility now, according to Wang. The remaining tradeoff is economic: having another route to AI training does not yet mean having an equally cost-effective one.
Loading discussion...
Join the conversation
Explain how much you would prioritize supply security over cost.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.