Moore Threads packs 256 GPUs into its MTT C256 system

Moore Threads unveiled the MTT C256 at WAIC 2026, a data-center system that links 256 GPUs across two standard racks.

The setup uses a single-layer Scale-up network and is said to deliver sub-microsecond all-to-all latency.

The company also demonstrated training a 236-billion-parameter mixture-of-experts model on more than 25 trillion tokens.

The system points to advances in dense GPU interconnects and scalable infrastructure for large AI training workloads.