The rapid escalation of AI workloads has turned once-routine cooling and power delivery into the primary bottlenecks for data center scalability. Addressing these constraints, Arthur Wang, Director of Product Management at xFusion, introduced a shift in operational metrics: WATT → FLOPS → TOKEN → AGENT → VALUE. This approach moves beyond hardware volume, evaluating how effectively electricity is converted into autonomous agent performance.

To support this shift, xFusion is showcasing three core engineering advancements at the San Jose McEnery Convention Center. The company’s direct liquid cooling solutions, featuring diamond-copper composites, now manage heat loads exceeding 4,000 watts per chip. Their FusionPod for AI V8 system provides a modular 155 kW rack design that integrates into existing data center footprints without requiring secondary piping modifications.

Power and connectivity are also undergoing radical architecture changes. xFusion is moving away from traditional 54V distribution in favor of an 800V high-voltage DC system, designed to minimize line losses across megawatt-class configurations. Complementing this, the DaYu 2.0 Super Node supports up to 256 XPUs across dual cabinets, utilizing optical interconnects to bypass the physical signaling limits of traditional copper. By integrating these systems, xFusion aims to achieve computing-power synergy, optimizing the energy path from the power grid directly to the processor.