800VDC: Power Architecture for the AI Rack Era

Abstract

Design power per rack in hyperscale AI data centers will move from around 22 kW in 2018 to 600 kW by 2027. That is not an incremental change to the existing power architecture. It is a break. This paper walks the energy budget from the high-voltage utility feed to the GPU package, accounts for every conversion loss along the path, and compares the conventional six-stage 415 VAC architecture with the four-stage 800 VDC architecture HyperNext is building. The end-to-end efficiency gap is roughly twenty percentage points. Conventional delivers 75 percent of grid energy to the chip. 800 VDC delivers 95. When inference is billed by the token, twenty points compounds into real money. We cover where the savings come from, what the architecture costs in capital and design effort, and why the rack-scale GPU systems shipping in 2027 and 2028 were architected around DC delivery from the start.

Contents

1. 011. The rack power budget is not the same problem anymore
2. 022. The conversion arithmetic
3. 033. What this costs
4. 044. Why the rack-scale systems require it
5. 055. What we recommend
6. 066. DC fault behaviour and protection
7. 077. Cable sizing and thermal analysis
8. 088. Worked example: 100 MW DC distribution sizing
9. 099. Standards landscape and references

Key findings

• Conventional 415 VAC delivers 75 paise of every grid rupee to the GPU.
• HyperNext 800 VDC delivers 94 paise of every grid rupee to the GPU.
• The difference is structural. It comes from removing conversion stages, not from squeezing efficiency out of stages that remain.
• For inference workloads billed at INR 0.6 per million output tokens, the recovered nineteen percent generates approximately INR 95 crore per year of additional revenue per gigawatt of facility capacity.