At the Yunqi Conference, Alibaba unveiled its plan to train models with 5-10 trillion parameters, released the Zhenwu V900 chip, and set a target of over 20GW of data center capacity by 2032. Grok 4.7 was released, with standard API pricing remaining

2026-09-22

At the Yunqi Conference, Alibaba unveiled its plan to train models with 5-10 trillion parameters, released the Zhenwu V900 chip, and set a target of over 20GW of data center capacity by 2032. Grok 4.7 was released, with standard API pricing remaining the same as the previous generation. Meanwhile, OpenAI partnered with an independent mathematical advisory group, NVIDIA launched a certification program for power supply and cooling equipment, and Meta announced a new submarine cable project. Model Pulse focuses on the surge in GLM usage and the latest independent evaluations from Grok and MiMo.