AI racks to hit 1 megawatt per unit by 2028–2030, reshaping power infrastructure

Key takeaway

• AI data center racks are consuming power at unprecedented rates, with individual units projected to exceed 1 Megawatt per rack by 2028–2030—far beyond the 140 kW of earlier NVIDIA generations.

• This shift is forcing a fundamental redesign of data center electrical architecture, with a move toward 800V HVDC systems and advanced wide-bandgap semiconductors to manage thermal and energy-delivery challenges in increasingly dense AI compute footprints.

3 Key Points

FAQ

What power levels are AI racks currently reaching?

NVIDIA Hopper and Blackwell server frames like the GB300 NVL72 reached around 140 kW. The newer Vera Rubin and Feynman generations are pushing to 600 kW+ for Rubin Ultra, with projections to exceed 1 Megawatt (1,000 kW) per rack by 2028–2030.

What is driving this increase in power consumption?

The rapid growth of generative AI and large language models (LLMs) has fundamentally changed assumptions. Modern AI clusters integrate more AI GPU/ASIC dies and multi-terabit SerDes switching fabrics into dense footprints, forcing architects to simultaneously optimize computing density, memory bandwidth, interconnect capacity, thermal management, and energy delivery within compact rack-scale systems.

What technology is being adopted to handle higher power demands?

The industry is transitioning to 800V HVDC power systems. Wide-bandgap semiconductors, specifically silicon carbide (SiC) and gallium nitride (GaN), are joining the supply chain because they offer competitive advantages over traditional silicon devices in managing the higher power and thermal requirements of AI data center infrastructure.