Power constraints now dictate AI infrastructure economics, with energy costs reaching forty percent of operating expenses and grid limits capping facility expansion. At the gigawatt scale, performance per watt determines token production and profitability. NVIDIA addresses this with a full-stack optimization strategy leveraging hardware co-design, software frameworks, and real-time management to maximize efficiency within fixed power envelopes. Inference drives revenue, making throughput per watt the primary target. NVIDIA architectures improved inference efficiency by a factor of one million across six generations. The GB200 NVL72 platform uses direct-to-chip liquid cooling and power smoothing to stabilize current spikes, enabling additional GPU deployment within budget limits. Narrow-precision formats like NVFP4 further cut energy use while maintaining accuracy. Mixture-of-experts architectures improve efficiency by activating only a fraction of parameters per token. NVIDIA Dynamo and TensorRT-LLM orchestrate these gains, scaling reasoning models across clusters while minimizing latency and infrastructure overhead. Training optimization tackles a separate bottleneck. Distributed training wastes energy through unbalanced utilization, as faster nodes idle waiting for slower critical-path tasks. Research with the University of Michigan’s ML.ENERGY Initiative introduced coordinated speed tuning for Megatron-LM, NVIDIA’s open-source training framework. Throttling faster GPUs to prioritize the critical path reduces training energy by twenty-five percent without extending iteration times. This shifts training onto a better energy-performance frontier, freeing power for additional model runs or shifting capacity toward inference workloads. NVIDIA DSX provides the operational backbone for energy-aware factories. The platform integrates telemetry, dynamic power allocation, and cooling management to eliminate stranded capacity. DSX MaxLPS raises liquid cooling inlet temperatures to improve power usage effectiveness, while reallocating power across racks to match workload demands. DSX Flex extends optimization to the electrical grid, synchronizing factory operations with external energy signals and carbon intensity metrics. Aligning scheduling with available cooling and power zones allows operators to prioritize high-revenue token generation under strict caps. These optimizations transform constrained power into a competitive advantage. Adopting full-stack management yields up to 2.6 times more tokens per second per megawatt than conventional deployments. As power limits restrict expansion, treating token economics and performance per watt as foundational design parameters will dictate infrastructure viability. NVIDIA’s integrated approach establishes a scalable pathway for sustaining AI growth within finite energy resources.