ET Datacenters argues that AI workloads are reshaping power efficiency in data centres.

Artificial intelligence (AI) workloads are forcing data centres to reconsider long-standing assumptions about power delivery and efficiency. The deployment of large GPU clusters, such as NVIDIA's 72-GPU Vera Rubin model, is increasing rack power densities from tens to hundreds of kilowatts, with some designs projected to exceed one megawatt per rack.

These AI workloads behave differently from conventional applications, causing rapid utilization fluctuations from approximately 30 per cent to full in milliseconds during training cycles. This dynamic behaviour challenges traditional power architectures optimised for stable workloads.

**Rethinking Power Efficiency**

Historically, power conversion systems focused on peak efficiency, or the maximum efficiency achieved at a specific operating point. However, AI workloads rarely operate in a steady state, spending more time at partial loads where efficiency often declines.

Even a one- or two-percentage-point drop in efficiency can lead to substantial energy losses across hundreds or thousands of GPU racks. This inefficiency increases heat generation, cooling requirements, and operational costs.

**High-Voltage Distribution and Flat-Load Efficiency**

Traditional 48-54V DC busbar architectures struggle to scale with increasing rack power requirements. Lower voltages demand higher currents for the same power, leading to resistive losses, larger cabling needs, and increased material costs, such as 200 kilograms of copper for a 1MW rack.

Data centre operators are now considering two strategies: adopting higher-voltage power distribution architectures, such as ±400V or 800V DC models, and maintaining flatter efficiency profiles across the full operating range. Higher voltages reduce current, copper requirements, and resistive losses, supporting greater rack densities.

Flat-load efficiency ensures consistently high performance across the fluctuating load ranges typical of AI workloads. This approach helps reduce energy losses, lower heat generation, and improve the predictability and scalability of power systems. Philip Zuk, senior vice president and general manager, AI and Datacenter, with OmniOn Power, states that these combined approaches help design power systems for how AI actually behaves.