Iluvatar CoreX Unveils Tiangai 300 Supernode With 144 GPUs in a Single Cabinet

Chinese AI accelerator developer Iluvatar CoreX has unveiled its first supernode, the Tiangai 300 IXL144. A single cabinet holds 144 GPUs and 18 CPUs. The company calls this an industry record for GPU density in one system. The announcement came on Monday, October 12, 2026, at the China International Industry Fair (CIIF) in Shanghai. According to SCMP, the system is designed for training and inference of large AI models.

What is inside the cabinet

A supernode is generally treated as a single compute unit: dozens of accelerators are linked so fast that software sees them as one large GPU. This approach is needed when a model does not fit into the memory of a single server and data constantly moves between cards. The more accelerators sit in one domain, the less traffic spills onto the slower network between racks.

Key stated specifications:

• Name: Tiangai 300 IXL144
• Accelerators: 144 GPUs in one cabinet
• Processors: 18 CPUs
• Purpose: training and inference of large AI models
• Performance: about 10.1 claimed (the units are cut off in the available excerpt of the source material)
• Debut: October 12, 2026, CIIF, Shanghai

The ratio of 8 GPUs per CPU matches the classic eight-accelerator server layout. The cabinet may contain 18 such nodes, but that is our assumption, not data from the manufacturer.

Why density matters

For comparison, Nvidia's GB200 NVL72 rack combines 72 GPUs and 36 Grace CPUs. Among Chinese systems, Huawei's CloudMatrix 384 is the best known: it pools 384 Ascend accelerators but spreads them across several cabinets. Iluvatar CoreX is betting on a different approach: fitting as much as possible into a single enclosure. That has practical advantages. It means fewer cables and points of failure, simpler installation, and shorter data paths.

There is also a cost. Dense packing means high power consumption and strict cooling requirements at the level of a single rack. Not every data center in China or abroad is ready to host such a cabinet without reworking its engineering infrastructure.

What is left out

The record claim needs caveats. Comparing systems by GPU count is only valid if accelerators are counted the same way. A dual-die module can be counted as one GPU or as two. Comparability also depends on other parameters: memory capacity and bandwidth, interconnect speed between cards, and supported compute formats (FP8, FP16 and others).

The source material also gives no data on price, delivery dates, first customers, or the process node and manufacturer of the chips themselves. For the Chinese market, these are not minor details: access to advanced manufacturing remains limited, while domestic demand for Nvidia alternatives is high.

In practice, there are five things to look at when assessing the Tiangai 300 IXL144:

1. The units and measurement conditions behind the claimed "about 10.1". For comparability, numerical precision and mode (training or inference) matter. 2. The topology and speed of the interconnect inside the cabinet. This determines whether a 144-GPU system behaves as a single whole. 3. Software maturity: support for popular frameworks and model-serving tools. For a new accelerator, this often matters more than the hardware. 4. Power consumption and cooling type, liquid or air. 5. Real pilot deployments at data center and cloud operators. So far this is a trade-show debut, not confirmed production use.

If Iluvatar CoreX discloses these figures and shows working clusters, the company will establish itself among Chinese suppliers able to sell not just individual cards but complete rack-scale systems. For now, the main fact remains the same: 144 GPUs in one cabinet and a record claim.