At the 2026 China International Industry Fair in Shanghai on October 12, domestic GPU company Iluvatar CoreX (09903.HK) unveiled the first solution in its systematic super-node portfolio — the TianGai 300 IXL144 — described as the highest-density single-cabinet GPU super-node in the industry. The system integrates 144 GPUs and 18 CPUs in one cabinet with roughly 10.1TB of memory.
As large models and agents move into production, the super-node has become the main unit of AI infrastructure. IDC data show China's intelligent computing scale is expected to reach 2,576.5 EFLOPS in 2026, up 87.9% year on year. Moving from technical validation to large-scale procurement, customers need to settle three accounts: whether existing models and code must be rewritten, whether the machine room needs dedicated modifications to power, load-bearing and cooling, and whether cost per million effective tokens can fall while capacity scales on demand without stranding.
The TianGai 300 IXL144 is designed against those three questions. Unlike mainstream whole-cabinet super-nodes that require purpose-built facilities with dedicated liquid cooling and power — building the room first, then the cabinet — the IXL144 adopts a single-cabinet form close to standard rack specifications, so it can enter existing facilities where power distribution, load capacity and cooling conditions match. Flexible networking topologies allow phased planning and stepwise expansion, avoiding over-investment upfront.
High-bandwidth full-cabinet peer-to-peer interconnect at roughly 100-nanosecond single-hop latency, native load/store and cross-node direct memory access, and the 10.1TB memory pool together let all 144 GPUs operate as a unified high-bandwidth domain with consistent memory semantics, leaving room for model weights, MoE experts and long-context KV cache. Its general-purpose GPU architecture is compatible with the mainstream CUDA ecosystem, allowing reuse of existing models and frameworks. According to company estimates, cost per million effective tokens falls about 40% compared with mainstream international solutions. In a financial product valuation scenario, curve interpolation over one million rows runs about four times faster than CPU-based approaches.
Iluvatar CoreX vice president Zou Xuan said a super-node is not a single product but a systematic engineering effort, and the real contest comes down to system economic efficiency. The TianGai series has already achieved deployment at 10,000-card scale.
Source: Xinhua Finance / 21st Century Business Herald / Southern Plus
