±400 VDC vs 800 VDC
Rack current drives copper mass, breaker size and I²R loss, which is why DC distribution is on every high-density roadmap. ±400 VDC (two rails around a grounded midpoint) and 800 VDC (two wires) put the same full-span voltage across the load, so they draw the same current for the same power. What separates them is grounding, how a ground fault behaves and is cleared, how the bus is isolated for service, and which racks ship for each. Against AC there is no universal halving of current: compare the whole route, conductor counts and conversion losses.
| Axis | ±400 VDC | 800 VDC |
|---|---|---|
| Topology | two rails around a midpoint grounded solidly or through high resistance; a ground-fault detector is required either way | two wires with one end bonded, high-resistance grounded or isolated from earth; each scheme needs its own fault study |
| Reference design | OCP Mt. Diablo 1.0.0 (Microsoft, Meta, Google) specifies the midpoint bus and also an 800 VDC two-wire PSU option (isolated from PE), so check which grounding option the vendor offers | NVIDIA's Kyber-class roadmap: ~600 kW per rack on 800 VDC, a roadmap figure rather than an installed rating |
| Copper and current | the same ideal current as the two-wire option at equal DC power; count any midpoint load and return conductor separately | NVIDIA quotes +157% power through the same wire gauge for three-wire DC against its four-conductor AC reference, which is not a like-for-like comparison with the midpoint bus |
| Voltage to ground | a solidly grounded midpoint holds each rail at ±400 VDC to earth in normal operation; high-resistance grounding changes what the first fault does, so study it separately | bonding one end puts the other pole at the full bus voltage to earth; isolated and high-resistance schemes need their own first-fault study before you count on riding through a fault |
| Conversion chain | an AC-fed sidecar keeps the upstream transformer and any UPS, then converts AC to DC and DC to DC; compare every package at the same output point | a medium-voltage solid-state transformer can replace the transformer and rectifier, but the rack's DC/DC stage remains. A prototype reached ~98%; SemiAnalysis models ~87% utility-to-VRM for the full chain, a model rather than a test result |
| Parts maturity | EV and industrial breakers, connectors and insulation are candidate supply; test DC interruption, isolation, thermal duty and the service procedure at your voltage before counting on them | the same tests at full pole-to-earth voltage; listings and delivery dates decide whether parts arrive on schedule |
| When AC stays | AC stays when the rack roadmap fits the in-rack low-voltage busbar and passes its current, heat, fault and service checks at lower cost | the same test applies; AC also stays for other loads whose interface and lifecycle cost favor it |
Quantitative cells are the guide's canonical figures — each is date-stamped and sourced in the numbers register and derived in the chapters below.
How the decision falls
Stay on AC until your rack roadmap outgrows the in-rack low-voltage busbar: below that point, AC busway is cheaper, mature and fully supplied (Chapter 4.7). Once racks cross it, pick the DC interface your racks ship with: OCP Mt. Diablo (the midpoint bus or its two-wire option) for OCP-aligned racks, NVIDIA's 800 VDC for Kyber-class racks. Both draw the same current for the same power, so the copper does not decide; grounding, fault clearing, service isolation and parts delivery do.
What would flip it: Switch interfaces if the chosen one fails its fault or service-isolation test or misses its delivery date, or if the other delivers the same usable capacity at lower lifetime cost. Credit a solid-state transformer's efficiency or footprint gain only for the whole route: it does not remove the rack's own DC/DC stage.
Model this fork with your own numbers: Build-cost calculator →
Full derivations, worked examples, and the numbers behind this matrix: The DC power revolution (Ch 4.7) · The I²R wall: busway to rack (Ch 4.6) · Transformers and the SST (Ch 4.4)