KNOWLEDGE ATLAS (02513) Next-generation layout is becoming clearer: 1 GW computing power and top-notch infrastructure are successively implemented.

According to Bloomberg, KNOWLEDGE ATLASAI (02513, Z.ai) has successfully established a 1GW domestic AI computing data center using entirely domestically-produced AI chips. At the same time, KNOWLEDGE ATLAS has officially completed the acquisition of domestic AI heterogeneous computing software company XCore Sigma. The latter originated from the Institute of Computing Technology of the Chinese Academy of Sciences, and has long been involved in the research and development of heterogeneous computing software stack and compilation optimization, being recognized as one of the top AI Infra teams in China. Analysts believe that these two actions respectively complement the two key capabilities of computing supply and computing release. The former provides the computing resources needed for large-scale model training, while the latter improves the utilization rate of heterogeneous chips, reduces inference costs and improves model deployment efficiency through basic software capabilities such as compilers, runtime, and inference engines. Recently, discussions in the market about KNOWLEDGE ATLAS's next-generation basic models have been heating up. Some believe that with the support of large-scale computing power, a mature Infra system, and long-term accumulation of post-training capabilities, KNOWLEDGE ATLAS's next-generation basic models will continue to evolve towards larger parameter scales and higher levels of intelligence, while also considering inference efficiency and engineering deployability. Industry insiders further point out that the competition among global Frontier AI companies is gradually evolving from single-model capabilities towards system competition in models, computing power, Infra, and ecological capabilities. KNOWLEDGE ATLAS has recently been continuously enhancing its underlying capabilities and advancing its model, Agent, MaaS, and industrial ecological layout, indicating that it is building a complete competitive system of a basic model company, rather than just focusing on competition around individual model releases.