Measure useful work per joule
Peak compute throughput is not the same as useful application work. A deployment model should include memory, networking, cooling and idle behavior, with task quality held constant. At the edge, battery life adds another constraint to model and radio scheduling.
Place work where the system benefits
Local processing can reduce transmitted data or preserve an offline capability, while cloud services can host larger models. The best split depends on model accuracy, update frequency, network conditions and device resources. Edge processing does not automatically reduce total energy.
Keep infrastructure data separate from product claims
These figures describe global data centres, not FuuVeXia infrastructure capacity or product savings. Demonstrating a reduction requires an application-specific baseline, workload, measurement boundary and comparable quality of service.