Zhipu AI’s Infra Agent Speeds Up Sovereign Hardware Clusters Using Recursive Optimization

Zhipu AI founder Tang Jie announced that the company's GLM-5.3-driven 'Infra Agent' delivered a 3.2x end-to-end throughput improvement on a production-grade inference system running on a cluster of more than 100,000 domestic Chinese AI chips in under two weeks. The agent autonomously identified and resolved bottlenecks spanning operator precision issues, Python/C++ concurrency, and kernel optimization, which Zhipu AI said demonstrates progress toward 'recursive self-improvement' in AI on large-scale domestic hardware.

Zhipu AI’s Infra Agent Speeds Up Sovereign Hardware Clusters Using Recursive Optimization Professor Tang Jie, founder of Zhipu AI, announced that their GLM-5.3 driven 'Infra Agent' achieved a 3.2x end-to-end throughput improvement for a production-grade inference system built on a cluster of over 100,000 domestic Chinese AI chips in under two weeks. AsiaAI Publisher · September 17, 2026 · 2 min read · Source: 量子位 QbitAI · Issue 99 East Asian Technology Intelligence Japan & China tech news — translated, contextualized, and delivered for Western readers. Free. Unsubscribe anytime. This story ran in Issue 99, alongside three other stories. AI & Machine Learning Professor Tang Jie, founder of Zhipu AI, announced that their GLM-5.3 driven ‘Infra Agent‘ achieved a 3.2x end-to-end throughput improvement for a production-grade inference system built on a cluster of over 100,000 domestic Chinese AI chips in under two weeks. This agent autonomously identified and resolved performance bottlenecks from scratch, spanning operator precision issues, Python/C++ concurrency, and Kernel optimization. This development from Zhipu AI, a leading Chinese LLM developer, demonstrates progress in AI’s ability to optimize its own underlying infrastructure, even on complex, large-scale domestic hardware. It highlights China’s ambition to achieve ‘recursive self-improvement’ RSI in AI, aiming for systems that can autonomously design and train their successors.