Maintain "Buy" rating on ILUVATAR COREX (09903); releases first supernode solution Tiangai 300 IXL144

Shenwan Hongyuan Group released a research report maintaining its earnings forecasts for ILUVATAR COREX (09903), expecting net profit attributable to parent of 1.03, 16.96, and 32.64 billion yuan for 2026-2028, corresponding to 629x, 38x, and 20x PE, and maintaining a "Buy" rating; the bank believes that the company's GPGPU products focus on TCO and computational efficiency adaptation for business scenarios, possess competitive advantages, and are expected to fully benefit from the high prosperity of domestic computing power. Shenwan Hongyuan Group's main points are as follows: ILUVATAR COREX releases its first 144-card supernode solution On October 12, 2026, ILUVATAR COREX showcased its first systematic supernode solution, Tiangai 300 IXL144, at the 2026 China International Industry Fair. The solution can integrate 144 GPUs and 18 CPUs in a single cabinet, equipped with approximately 10.1TB of memory. The single-cabinet supernode solution is one of the mainstream supernode solutions and is recognized by major customers Currently, supernode design solutions include multi-cabinet solutions represented by Huawei CloudMatrix384 and single-cabinet solutions represented by Alibaba Panjiu 128. Compared with multi-cabinet solutions: single-cabinet solutions increase GPU density, improving data center space resource utilization, but at the same time place higher demands on liquid cooling, power supply, load-bearing and other infrastructure, potentially requiring significant infrastructure renovation; in single-cabinet solutions, GPUs and switches are generally connected via orthogonal plugging, reducing copper cable usage, with Scale-up not using optical interconnects, and the per-chip interconnect network cost may be relatively lower. Based on the supernode solutions currently launched by internet vendors (Alibaba Panjiu, ByteDance Dayu, etc.), first-generation supernode products tend to favor single-cabinet designs. Supernodes aggregate more AI chips through Scale-up networking, improving the system performance of domestic AI chips in inference scenarios Under the trend of model parameter counts and the increasing number of experts in MoE models, inference demands for computation, memory, and interconnect bandwidth will continue to rise, requiring computation and memory to be co-located within a tightly integrated supernode to reduce communication latency and maintain high throughput. Therefore, the capacity of supernodes is important for maintaining the fine-grained locality and performance characteristics required by MoE models. A larger resource pool can provide greater deployment flexibility for non-uniformly sized jobs, and under actual workload distributions, scaling up supernode size can significantly improve system throughput and efficiency. The domestic supernode ecosystem is gradually maturing, with computing power infrastructure accelerating deployment in the second half of 2026. Currently, the domestic supernode ecosystem is gradually moving toward open source, with internet giants, AI chip companies, and server companies jointly developing supernode products, formally entering the supernode era from the demand side. Domestic computing power prosperity remains high, and under supply shortage the company's performance may exceed expectations Domestic foundation models are ushering in a new round of intensive iteration. DeepSeek released V4.1-Flash and open-sourced the full set of infrastructure components for the Ascend platform, and Moonshot previewed Kimi K3.1, among others. Domestic models continue to narrow the gap with overseas leading models in Agent Engineering, Coding, and complex reasoning capabilities. Model iteration drives continued rapid growth in inference demand, while domestic computing power supply remains persistently tight. The bank believes that against the backdrop of supply shortage in computing power, Tiangai 300 IXL144, as the company's first single-cabinet 144-card supernode solution, is positioned to address the core inference computing power needs of internet giants, and the company's order visibility is expected to improve significantly, with future performance potentially exceeding expectations. Risk warnings Risks of uncertainty in technology; technology iteration progress falling short of expectations; risks of supply chain instability.