Source

Tags: ai-power, hvdc, power-semiconductors, sic, gan, data-centers, voltage-stack, war-of-currents, transformers, skin-effect, nuclear-grid-deliverability

• Source: The AI Power Crisis — Part 1
• Author: Nutty (@NuttyCLD)
• Date published: 2026-03-30
• Date captured: 2026-05-01

TL;DR

The frontline of the GPU war has shifted from compute speed to power delivery. NVIDIA's own analysis says ~30% of all power is lost to conversion, distribution, and cooling before it ever reaches the GPU — at a 1 GW data center, that's 300 MW of "parasitic energy" turning into heat. The 140-year-old War of Currents is reigniting because power semiconductors (SiC/GaN) finally make HVDC competitive with AC. Generation isn't the bottleneck — delivery is. Three Mile Island can't connect until 2031 even if it restarts in 2027.

Key takeaways

• Power, not compute, is the new bottleneck. H100 = 700W, B200 = 1,000W+, Vera Rubin = 2,300W per GPU. Racks went from 5-15 kW (traditional) → 120 kW (Blackwell GB200/GB300) → 600 kW (Kyber 2027) → 1 MW design target.
• 30% parasitic energy. Power loss between grid and GPU breeds more loss (heat → cooling → more heat).
• The grid is "a civilization built on transformers." Transformers are passive (no moving parts, near-zero loss) but only work for AC. That's why AC won the 1890s War of Currents — DC voltage conversion required active power electronics that didn'...