Moore Threads, a Chinese GPU developer, demonstrated its MTT C256 system at WAIC 2026, linking 256 GPUs into a single data-center-scale computing unit.

The system uses a one-layer Scale-up network for all-to-all communication across the 256 cards and is housed in two standard racks. Moore Threads says the network achieves sub-microsecond latency.

The company also demonstrated training for a 236-billion-parameter mixture-of-experts model using more than 25 trillion tokens. [IT Home, in Chinese]