China’s biggest technology company has launched a long and difficult process of switching artificial intelligence development to homemade chips to tighten the declining stockpile of Nvidia processors and US export controls.
According to industry executives, Alibaba, Tencent and Baidu are among those who are beginning to test alternative semiconductors to surge AI-related internal demand and client needs.
They were forced to bolster their accidental plans last month as Donald Trump’s administration came to close on sales of their favourite chips: a watered-down product tailored to complying with the Joe Biden era curb.
According to insiders with knowledge of the issue, tightening controls added urgency to the Chinese tech group’s movements.
Typically, new chip orders take 3-6 months to ship. It is also unclear when Nvidia will be able to compete for new processors aimed at China that comply with Trump’s stricter export rules and compete well against local rivals.
Shen Dou, head of Baidu’s AI Cloud Group, told analysts last week that they have a variety of chip options to replace Nvidia’s, particularly for problem-solving inference processing.
“Over time, we believe that the self-sufficient chips developed in the country, together with an increasingly efficient homemade software stack, will jointly form a strong foundation for long-term innovation in China’s AI ecosystem,” Shen said.
“We are actively investigating diversified solutions to meet the growing demand of our customers,” Alibaba chief Eddie Wu said in a revenue call earlier this month.
In another revenue call, Tencent President Martin Lau said his company is trying to make the chips more efficient while considering alternative products.
“We need sufficient high-end chips to continue our training model for generations to come,” Lau told analysts, adding that Tencent can meet the growing need for inference by “potentially utilizing other chips.”
This month, a think tank belonging to China’s National Security Department said Washington’s export controls were painful, but “spurred a surge in independent innovations in domestic high-end AI chips in Huawei’s Ascend Chip series The Prime Emplement.”
“Chinese domestic companies have already begun to procure large scale and use ascend chips,” the Institute for Modern International Relations China said in a social media post.
So far, Huawei’s biggest chip buyers have been state-owned companies such as China Mobile and companies in sensitive industries such as defense, healthcare and finance. Currently, a much wider range of domestic tech companies is expected to compete for chips for China’s national champions.
Recommended
After Washington issued guidance on export restrictions this month, those looking for Huawei as a possible alternative remain largely quiet about testing Ascend Chips.
Analysts at GF Securities estimate that Nvidia could begin creating the next chip in early July for exports to China, which comply with US export regulations.
The new processor is based on Nvidia’s Advanced Blackwell products, but reports show that there is no high-bandwidth memory (HBM), a key component for quickly processing large amounts of data.
Certain important details, such as whether the new processor has NVIDIA’s high-speed interconnect nvlinks, remain unknown.
In a call to analyst revenues on Wednesday, Nvidia chief Jensen Huang said the options are limited as he thought about new China products. “There’s nothing at this point,” he said.
The High Tech Group faces considerable costs if it decides to move its system from Nvidia chips to a domestic alternative. Porting training code developed using NVIDIA’s CUDA software framework for Huawei Cann is very time-consuming and often requires important support from Huawei engineers.
Executives at a major Chinese tech company estimated that the switch to Huawei will cause roughly three months of disruption in AI-related developments.
Most companies are considering a hybrid approach, with AI training continuing to run on existing NVIDIA chips, and local processors are used to infer that demand is growing as AI is being adopted more widely in China.
Huawei is trying to increase the production capacity of its partners and has set up its own manufacturing plant, but supply does not meet current demand.
Chips from other Chinese chip makers, such as Cambricon and Hygon, have also been tested by technology giants, but Baidu and Alibaba are developing their own processors to meet the rising demand.
Additional Reports by Eleanor Orcott of Taipei
