Key Takeaways:
KNOWLEDGE ATLAS (02513.HK) boosted its AI computing power, increasing average GPU inference throughput by 15 percent after deploying the new ZCube network architecture in its live production environment, a move that enhances its competitive standing as a key infrastructure partner for hyperscalers like Alibaba Cloud.
"In response to increasingly severe structural network congestion challenges... ZCube achieved breakthroughs purely through architectural optimization," the company said in a statement. The research was conducted jointly with Yuxun Network and Tsinghua University. In benchmark tests within the GLM-5.1 coding production environment, the company reduced capital expenditure on switches and optical modules by 33 percent.
Beyond the cost savings and throughput gains, the architecture also improved latency, with Time to First Token (TTFT) P99 dropping by 40.6 percent. The improvements were achieved with no changes to the existing GPUs, software stack, or applications, highlighting the efficiency of the network-level optimization. The company did not disclose the total MW capacity of the upgraded environment.
This architectural upgrade strengthens KNOWLEDGE ATLAS's position in the high-stakes AI infrastructure market. As a partner to major players like Alibaba Cloud on its Bailian platform, per a recent CMSI report, demonstrating superior performance-per-dollar is critical. The 33 percent capex reduction on networking hardware directly translates to higher margins and a more competitive pricing structure, potentially allowing KNOWLEDGE ATLAS to win larger deals from enterprises and cloud giants grappling with the soaring cost of building out AI capabilities.
This article is for informational purposes only and does not constitute investment advice.
