d-Matrix Debuts Megawatt-Scale AI Inference Clusters

The Squadron clusters, developed alongside Arista, Broadcom, and Supermicro, utilize a memory-centric architecture designed to integrate into existing air-cooled data centers without requiring expensive retrofits. By offering a modular design—ranging from individual Corsair servers to 16-rack clusters—the company provides a pathway for enterprises to scale inference compute while bypassing the liquid cooling requirements often associated with high-density AI clusters.

Beyond hardware, the company is proving the practical application of Attention-FFN Disaggregation (AFD). During a demonstration at the OCP Global Summit, d-Matrix paired its Squadron systems with traditional GPU servers. This heterogeneous setup achieved a throughput of 20,415 tokens per second across 64 concurrent users, effectively offloading intensive workloads to the more efficient Corsair racks. This marks the first time AFD has been validated on commercial silicon, signaling a shift in how AI labs may balance their compute environments to optimize for both speed and energy consumption.