Power Constraints Have Redefined Data Center Scaling
The transition to rack-scale AI systems has made power delivery and thermal management the primary bottlenecks for growth.
Updated on Sept. 29, 2026 in Data Centers
Live Poll
Do you worry that large-scale AI data centers will negatively affect energy availability in your area?
Large-scale AI data centers now routinely consume more than 100 MW of power, forcing an industry-wide shift in site selection and infrastructure design. This transition emphasizes power access and thermal efficiency over traditional land availability as global investment in data centers accelerates.
Why it matters
The massive energy demand of continuous inference and sovereign AI models has moved the bottleneck from pure processor performance to physical infrastructure. As a result, the industry is prioritizing integrated rack-scale power delivery to manage the heat generated by these high-density computing loads.
Large AI data centers currently exceed 100 MW of power consumption, necessitating complex multi-stage conversion systems from utility grids down to rack-level power shelves. These systems now utilize a mix of silicon, silicon carbide, and gallium nitride power electronics to manage thermal loads.
The players
Ang Wee Seng
An industry analyst focused on the intersection of power electronics and the semiconductor supply chain.
The details
Modern AI infrastructure organizes accelerators, networking, and power delivery into integrated rack-scale systems where the rack serves as the primary unit of design. Because every watt consumed generates heat, cooling systems are now inextricably linked to power delivery, creating a hard physical limit on scaling. To navigate these constraints, power electronics are evolving to include materials like silicon carbide and gallium nitride to improve conversion efficiency.
Timeline
July 2026: Ang Wee Seng discussed the infrastructure shift in an EE Times ASEAN e-guide.
September 29, 2026: The industry trends were presented at EE Power Asia 2026.
The Tech Race
The global surge in AI infrastructure investment is tied directly to the transition toward sovereign AI and continuous inference requirements. This shift follows a trend where site selection is dictated by regional grid capacity rather than conventional data center location factors.
Companies and developers must now account for power scarcity as a primary business constraint when deploying high-density AI models. This will likely lead to a concentration of infrastructure in regions with high-capacity power access, affecting the geographic rollout of future AI services.
The takeaway
The move toward rack-scale integration shows that the physical limits of thermodynamics are now the most significant challenge for AI scaling. Watch for future performance benchmarks regarding silicon carbide adoption in rack power shelves to see if these components can keep pace with increasing watt requirements.
Further reading
For more on the hardware architecture driving current growth, see Data Centers.
Source note: This article includes information reported by EE Times.
Live Poll
Do you worry that large-scale AI data centers will negatively affect energy availability in your area?
