Startup Boosts Scale-Up to 1000+ GPUs in a Single Domain

Startup Delos Data is advancing GPU scale-up capabilities beyond 1000 GPUs in a single domain through its cluster management software and a new server architecture, aiming to enhance GPU utilization a... Read original →

Delos Data’s breakthrough in scaling beyond 1,000 GPUs within a single domain signals a pivotal shift from brute-force compute stacking to intelligent resource orchestration. Its Mosaic software and modular hardware directly challenge NVIDIA’s closed NVLink ecosystem—if NVIDIA fails to offer topology flexibility, its inference-market premium will erode. Technically reliant on TSMC (Taiwan, China) 3nm and EUV nodes, the architecture faces supply chain risk: U.S. export controls on OSFP optical interconnects could inflate deployment costs by over 15%. Competitors like Supermicro and Inspur will likely accelerate disaggregated server rollouts, while hyperscalers such as AWS may fast-track in-house alternatives to A100 clusters. Within 18 months, the market will bifurcate—training systems retain tight coupling for throughput, while inference infrastructure fully embraces heterogeneous, elastic designs, making GPU utilization—not just card count—the new KPI.