CCTV News reported on August 31 that throughout 2026, multiple major AI models have been quickly adapted by domestic computing platforms following their release, revealing both challenges and opportunities in the "two-way pursuit" between large models and computing power platforms.
Recently, an anonymous model tested on OpenRouter became the most popular model of the week, later revealed to be the latest model from Zhipu AI. Notably, all test-period request traffic was supported by domestic chips. Earlier, Moonshot AI released the Kimi K3 model with 2.8 trillion total parameters, and within less than a day, multiple domestic computing platforms announced completed adaptation.
A domestic computing enterprise representative explained that the core challenge comes from the massive time gap between years and months. Chip industry cycles are calculated in years, while large models release new versions almost monthly, with new model structures, operators, and precision requirements constantly testing computing chips.
The "15th Five-Year Plan" explicitly calls for strengthening computing infrastructure support and accelerating the cultivation of independently controllable, collaboratively operating software and hardware ecosystems. The National Digital Research Institute's deputy director Chen Dong emphasized that the domestic computing chain's value is not just "cheaper" but "more controllable" - with controllability across cost, data, and supply chains forming the true competitive moat.
On the application side, domestic large model and computing power vendors are actively collaborating, with demand continuously increasing. One model company architect noted they are working with nine domestic computing vendors on adaptation and building their own domestic computing cluster, expected to be completed this quarter.
Tsinghua University Professor Chen Wenguang noted that the relationship between models and computing power is a complex interaction where model designs can influence chip designs and vice versa, making the combination more efficient and capable of triggering innovation across industries.
Source: CCTV News
