Alibaba unveils homegrown chips

Alibaba and China Telecom said on April 8 they had opened an artificial intelligence data center in Shaoguan, Guangdong, built around 10,000 Alibaba-designed Zhenwu chips. (cnbc.com) The companies said the site is designed for both training and inferencing, the two main jobs behind modern artificial intelligence systems: building models from data and then running them for users. Alibaba said the cluster can support models with hundreds of billions of parameters and could later expand to 100,000 chips. (cnbc.com) Zhenwu is part of Alibaba’s chip effort inside its T-Head unit, which has previously designed processors for cloud and edge computing. This Shaoguan project is the first large-scale data-center deployment of the Zhenwu line, moving Alibaba’s chip work from design into shared commercial infrastructure. (finance.yahoo.com) The location matters because Shaoguan sits in Guangdong, the province at the center of southern China’s manufacturing base and the Greater Bay Area technology corridor. Alibaba’s cloud unit said this is the first Zhenwu project at that scale in the Greater Bay Area. (finance.yahoo.com) The timing also matters because Chinese technology groups have spent the past two years trying to replace foreign artificial intelligence hardware with domestic alternatives. Washington tightened export controls on advanced artificial intelligence chips in October 2023 after earlier restrictions in 2022, cutting Chinese buyers off from many Nvidia products. (cnbc.com) Even before this launch, Nvidia’s China-specific H20 chip had become the main legal option for many Chinese customers under those rules. Reuters reported last year that the H20 was introduced after the October 2023 controls and was the primary Nvidia chip still permitted for sale in China at the time. (straitstimes.com) Alibaba is not the only Chinese group building around local silicon. CNBC reported that Huawei and other domestic suppliers have been pushed into a larger role as Chinese companies assemble homegrown computing clusters for large language models and other artificial intelligence services. (cnbc.com) Alibaba has tied that hardware push to its cloud business, which is central to its current artificial intelligence strategy. The company has been rolling out its Qwen models and selling cloud computing to business customers, so owning more of the chip stack gives it another way to control supply and costs. (cnbc.com) What Alibaba has not published are the full chip-level specifications that would let outsiders compare Zhenwu directly with Nvidia’s top data-center processors. For now, the clearest fact is scale: Alibaba has put 10,000 of its own chips into one live facility and says the next target is ten times bigger. (cnbc.com)