Moore Threads packs 256 GPUs into its MTT C256 system

Moore Threads unveiled the MTT C256 at WAIC 2026, a data-center system that connects 256 GPUs in a single unit.

The company says the one-layer Scale-up network enables sub-microsecond communication latency across the cluster.

Moore Threads also said it used the system to train a 236-billion-parameter mixture-of-experts model on more than 25 trillion tokens.

The demo highlights how domestic GPU makers are pushing toward larger, tightly coupled AI training infrastructure.