Zhipu GLM Boosts Inference Throughput by 3x in Two Weeks Using 100K Chinese AI Chips

Zhipu GLM announced on its official blog that it built an inference system using a cluster of more than 100,000 Chinese AI chips and increased end-to-end processing speed to three times the initial baseline within two weeks. The company explained that the GLM-5.3-based infrastructure agent handled design, adjustment, and optimization of the inference infrastructure. During this process, hardware utilization efficiency and single token cost reached levels comparable to mainstream NVIDIA GPUs.

Zhipu GLM introduced this case as an engineering example of recursive self-improvement (RSI), where the model improves the system it runs on.