How Alibaba and Baidu Are Building Homegrown AI Chips to Challenge Nvidia

For the past decade, China’s artificial‑intelligence industry has been tightly coupled with Nvidia, with H100 and A100 GPUs serving as the de‑facto standard for large‑model training and everyday research, but escalating U.S. export restrictions are forcing a shift away from this dependence.

According to The Information , Alibaba and Baidu have begun using self‑developed chips for AI model training to partially replace Nvidia products. Alibaba started deploying its own chips for small‑scale models earlier this year, while Baidu is testing the Kunlun P800 chip on the new version of its Wenxin model.

Alibaba’s chip ambition dates back to 2018 when it acquired Zhongtian Microelectronics and created the “Pingtouge” semiconductor unit under the Damo Academy, subsequently launching processors such as Hanguang 800, XuanTie, and Yitian 710, and integrating them into cloud and inference scenarios. Earlier this year the company reportedly began internal testing of a new AI inference chip fabricated domestically, aiming to close gaps in large‑model inference and cloud computing while maintaining compatibility with Nvidia’s ecosystem.

Financially, Alibaba has poured more than 100 billion CNY into AI infrastructure and product R&D over the past four quarters, and in February announced an additional 380 billion CNY over the next three years for cloud and AI hardware, with chip development as a primary focus.

Baidu’s chip strategy is similarly long‑term; it formed a chip R&D team in 2011, released its first Kunlun chip in 2018 for autonomous driving and cloud inference, achieved several‑fold performance gains with the second‑generation Kunlun, and now employs the Kunlun P800 chip directly for training its Wenxin large model, creating a “dual‑wheel” model of algorithms and compute.

Both firms have not completely abandoned Nvidia. In cutting‑edge large‑model research, Nvidia remains indispensable, yet employees who have used Alibaba’s home‑grown chips report performance comparable to Nvidia’s top‑of‑the‑line H20 chip, suggesting the “critical point” for domestic substitution is approaching.

Nvidia, meanwhile, is seeking a way out of the export curbs; its CEO Jensen Huang said negotiations with the White House on a “downgraded” next‑generation chip for China are ongoing, and reports claim Nvidia may have agreed to remit 15 % of its H20 sales in China to the U.S. government.

In the short term, this rivalry will intensify the geopolitical dimension of the global AI industry; in the long term, as Alibaba, Baidu and others mature their chip capabilities, China’s AI sector could see a reshaped compute landscape, with the ability to build a complete domestic AI‑compute chain becoming a decisive factor for its future progress.

Source: https://www.reuters.com/world/china/alibaba-baidu-begin-using-own-chips-train-ai-models-information-reports-2025-09-11/