Chinese chipmaker Iluvatar CoreX has unveiled a supernode designed for large-scale AI training and inference, packing 144 GPUs into one cabinet. The company says its Tiangai 300 IXL144 can reduce the cost of generating tokens by 40 per cent compared with mainstream international alternatives, a claim that could matter as AI computing demand grows and access to advanced chips remains constrained.
Watch Desk analysis
What happened
The system was unveiled at the 2026 China International Industry Fair in Shanghai on Monday, according to the South China Morning Post. The paper reports that the cabinet also contains 18 CPUs and around 10.1 terabytes of memory. Iluvatar describes the system as integrating computing, storage, networking, management and software frameworks.
Read the South China Morning Post report.
Why it matters
A dense, integrated system is one way to scale AI computing without relying on a single enormous chip. The reported 40 per cent cost reduction, if borne out in comparable workloads, would be a substantial selling point for organisations training or running large models.
Our read
The hardware figures make this a notable AI-computing launch; the cost claim is the part buyers will want to interrogate. “Valid tokens” and “mainstream international alternatives” need a clear comparison before that percentage can guide a purchasing decision. A big cabinet is easy to count. A fair benchmark is the harder trick.
What to watch
- Which GPUs and CPUs the system uses, and how its components are configured.
- What workloads and comparison systems underpin the claimed 40 per cent cost reduction.
- Whether customers or independent evaluators report results from deployed systems.
Discussion spark: What evidence should buyers require before treating a vendor’s claimed AI-computing cost saving as a real advantage: published benchmark methods, independent tests, or results from their own workloads?
Sources and evidence
- Iluvatar CoreX’s first AI computing supernode squeezes most GPUs in 1 system – South China Morning Post (13 October 2026, 09:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.