Busbars, 54 Volts and the Road to 800 V DC: How AI Racks Are Rewriting Data Centre Power
For thirty years, powering a server rack meant two power strips and a row of plugs. An AI rack does not work that way. It has shelves of power supplies feeding a thick copper bar that runs the height of the rack, and the servers clip onto it.
That change has already happened. A bigger one is coming. NVIDIA and the power industry are preparing to move data centres to 800 volts DC, and a power design approved today may be a generation behind before the building's lease runs out.
Three eras of rack power
| What | π Traditional rack | β‘ Today's AI rack | π Next: 800 V DC |
|---|---|---|---|
| Power per rack | Around 10 kW |
132 to 140 kW |
Up to 1 MW |
| How power reaches servers | AC power strips and cables | Power shelves and a DC busbar at about 50 to 54 V | 800 V DC delivered to the rack, stepped down near the chips |
| Where AC becomes DC | Inside each server | In the rack's power shelves | Once, at the edge of the data centre |
| Status | π’ Everywhere | π’ Shipping | π‘ Planned for volume from 2027 |
Sources: HPE NVL72 QuickSpecs, NVIDIA on its 800 V DC architecture. The bar is scaled to 1 MW.
How today's AI rack is powered
Take HPE's NVL72 as the example. Its QuickSpecs list eight power shelves of 33 kW each, feeding a 50 V DC busbar, with the rack supplied by 400 V, 60 A power whips. The shelves convert the building's AC into low-voltage DC. The busbar carries it to every compute tray.
There is a detail here that buyers miss. Some configurations ship with six shelves and some with eight. Fewer shelves can still power the rack, but with less redundancy. It is the first thing to check when comparing quotes.
Why 54 volts is running out of road
Power is voltage multiplied by current. Keep the voltage low and every extra kilowatt arrives as more current. Current is what needs copper.
| Problem at megawatt scale | NVIDIA's figure |
|---|---|
| π₯ Copper | A single 1 MW rack at 54 V would need up to 200 kg of copper busbar |
| π¦ Space | Power shelves would take up to 64U, more than the whole rack, leaving no room for compute |
That second row is the decisive one. At megawatt scale a 54-volt rack would be entirely power supplies. The approach does not get harder. It stops working.
What 800 V DC changes
Raise the voltage roughly fifteen-fold and the current drops by the same factor. Thinner conductors carry the same power. NVIDIA's design converts grid power to 800 V DC once, at the edge of the facility, and distributes that to the racks.
| NVIDIA's claimed benefit | Figure |
|---|---|
| π Rack power range on one infrastructure | 100 kW to over 1 MW |
| π₯ Copper, versus 415 V AC distribution | About 45% less |
| β‘ End-to-end power efficiency | Up to 5% better |
| π§° Maintenance cost | Up to 70% lower |
These are NVIDIA's projections for an architecture that is not yet in volume production.
The timeline is public. NVIDIA says full-scale production of 800 V DC data centres will coincide with its Kyber rack-scale systems in 2027. Power vendors are lining up behind it; Vertiv, for one, has announced an 800 V DC portfolio. At least one facility, a Foxconn data centre in Taiwan, is reported to be operating on 800 V DC already.
Sources: Data Center Dynamics, NAND Research on the 800 V DC transition
What this means for a buyer today
The good news: nothing you buy this year needs it
GB300 and GB200 NVL72 racks run on the power-shelf-and-busbar design. Third-party analysis indicates that the Vera Rubin generation raises density but does not itself require 800 V DC. So the racks on sale now, and the next ones, fit facilities built for today's approach.
The awkward news: the generation after may not
If megawatt-class racks arrive on the published timeline, a hall designed around 140 kW AC-fed positions will not host them without major electrical work. That is not a reason to delay a purchase with a two-to-three-year payback. It is a reason to be careful about ten-year assumptions.
The honest uncertainty
Megawatt racks are a hyperscale concern first. Whether and when they matter to an enterprise in Singapore is an open question. High-voltage DC also brings its own safety requirements, equipment and skills, and local practice will take time to catch up. Treat 2027 as a vendor roadmap date, not a deadline.
Questions to ask before signing off a power design
| Ask | Why |
|---|---|
| β‘ What is the maximum power per rack position, including peaks? | Nominal figures understate what the feed must carry |
| π How many power shelves, and is that redundant? | Six versus eight changes resilience and feeds required |
| π How much headroom is there above today's rack? | The next generation is expected to draw more |
| π What is the operator's view on high-voltage DC? | Tells you whether they are planning or waiting |
| π What is the contract term versus the hardware's useful life? | Avoid committing a building for longer than the racks it suits |
The broader lesson is one this series keeps returning to. In rack-scale AI, the chip roadmap now dictates the building roadmap. The people who need to read NVIDIA's announcements are no longer only in IT.
Related on SourceIT
SourceIT Approved
Researched and verified by the SourceIT team for Singapore buyers.




