Baidu has successfully trained a key version of its flagship ERNIE 5.1 model on a fully domestic hardware and software stack, a major milestone in its bid to create a self-reliant AI ecosystem.
Back
Baidu trains ERNIE 5.1 model on its domestic Kunlunxin chips
Shen Dou, president of Baidu's AI Cloud Group, announced at the Create 2026 conference that its domestic Kunlunxin AI accelerators have achieved large-scale deployment, enabling the full training of its most advanced language model and reducing reliance on foreign hardware.
"On the fully domestic Kunlunxin cluster, the company has successfully completed training of a key version of ERNIE 5.1," Shen said, highlighting the performance of the company's vertically integrated AI infrastructure.
The training was achieved on clusters of Baidu's Kunlunxin P800 chips, which have completed large-scale validation with "multiple ten-thousand-card clusters" delivered since last year. The system achieved an overall effective training rate of 97% and linear scalability on a ten-thousand-card cluster that surpassed 85 percent.
This achievement is a direct challenge to the dominance of foreign chipmakers like Nvidia in China's AI development. By proving out a viable, domestically produced AI stack from hardware to model, Baidu strengthens its position in the enterprise cloud market and aligns with Beijing's strategic goal of technological self-sufficiency.
Tianchi Supernode to Launch in June
Further expanding its hardware capabilities, Baidu will launch its Tianchi 256-card supernode in June. This new system, based on the Kunlunxin architecture, boasts a 25% increase in throughput performance compared to the previous generation. The supernode has already been adapted for mainstream Chinese large language models, including DeepSeek, GLM, and MiniMax, in addition to Baidu's own ERNIE, positioning it as a versatile platform for the broader Chinese AI industry.
The Self-Reliance Strategy
Baidu's progress with Kunlunxin is critical amid ongoing US-China tech tensions and restrictions on the export of advanced semiconductors to China. While companies like Nvidia and AMD face hurdles in supplying the Chinese market, Baidu is building a competitive alternative. This vertical integration—controlling the chips, the cloud platform, and the foundational models—could offer Chinese enterprise clients a more stable and secure supply chain for their AI initiatives. The successful training of ERNIE 5.1, a direct competitor to models from Alibaba and other domestic rivals, on its own silicon serves as a powerful proof-of-concept for this strategy. The high training efficiency and scalability metrics suggest that Baidu's domestic hardware is reaching a level of maturity capable of handling cutting-edge AI workloads.
This article is for informational purposes only and does not constitute investment advice.
