What it is

AFCOM’s 10th anniversary State of the Data Center report documents average rack density reaching 27 kW per rack (69% YoY increase from 16 kW), driven by AI workloads. The article traces density evolution from 1 kW (1988) to today’s AI-capable systems, with NVIDIA roadmaps projecting 600 kW per rack for upcoming platforms like Rubin Ultra NVL576.

Why it matters

Facilities managers face a power distribution constraint shift from incremental scaling to backbone redesign: a single NVIDIA DGX H100 draws 10+ kW, and upcoming rack-scale systems approach 600 kW. This affects power delivery design, cooling capacity, and whether existing infrastructure can support AI deployment without campus-level power producer roles. The 74% planning AI-capable infrastructure must decide now if their distribution gear, thermal systems, and service capacity can handle 10x–20x density jumps.

Evidence from source:

  • Average rack density reached 27 kW per rack, up 69% YoY from 16 kW and up from 6.1 kW in earliest study edition.
  • Single NVIDIA DGX H100 system can draw 10 kW or more at node level; NVIDIA Rubin Ultra NVL576 rack expected to approach 600 kW threshold.
  • 74% of respondents plan to deploy AI-capable infrastructure; 72% expect AI workloads to significantly increase capacity requirements.

Open questions

  • What specific power distribution architecture changes (busway, direct feed, modular UPS) are operators selecting for 100+ kW and 600 kW rack targets?
  • How are commissioning and monitoring strategies adapting to detect thermal or power faults at these densities before catastrophic failure?