Moore Threads shows MTT C256 super node, connecting 256 GPUs in one network layer

At the 2026 World AI Conference (WAIC), Moore Threads showed its MTT C256 super node for the first time. It uses a new one-layer network to connect 256 GPUs into one supercomputer, beating the usual 64-card limit and cutting communication delays to under a microsecond.

With a compact design, the MTT C256 needs only two standard cabinets to connect all 256 cards, saving space and boosting GPU density. It also works with existing AI hardware and software, making it easy to scale from 10,000 to 100,000 cards.

The company also showed results from training large models on its own chips, including a 236-billion-parameter MoE model with over 25TB of training data, a 5D world model for Peking University, and a robot brain model for the Beijing Institute for Artificial Intelligence.