800 V DC Server Racks vs 48 V and +/-400 V: Current, Losses, Density

Data-center power voltage: from 48V to 400/800V

As rack power rises, data-center power architectures are being discussed in terms of higher voltage. Because current equals power divided by voltage, delivering the same power at a lower voltage increases current and expands conductor cross-section, copper loss, and heat. As AI racks exceed 100kW and move toward 1MW, the question is how to choose among 48V, +/-400V, and 800V.

48V DC: the conventional in-rack standard

In-rack 48V DC has been the standard power architecture. Power supply units(PSUs) create a 48V bus, and DC-DC converters on the server board supply around 1V to each device. Its strengths are field experience and mature components. At higher rack power, however, staying at 48V makes current enormous, exposing limits in conductors, copper loss, and space. That is the starting point for moving to higher voltage.

+/-400V / 800V DC: the shift to HVDC

Hyperscalers are driving the move to raise the voltage entering the rack to high-voltage DC(HVDC). In the disaggregated(sidecar) power-rack concept described by OCP(Open Compute Project), power delivery shifts from in-rack 48V DC to +/-400V or 800V DC, enabling 100kW-to-1MW IT racks. +/-400V has relatively lower insulation and protection hurdles than 800V, while 800V lowers current further and is easier to scale toward the 1MW class.

NVIDIA's technical blog says an 800V DC architecture can improve end-to-end efficiency by up to 5% and reduce maintenance cost by up to 70% (NVIDIA Technical Blog). NVIDIA also says it will lead the transition to 800V for the 1MW rack generation from 2027 onward.

The current-reduction effect is simple but large. Delivering 100kW at 48V implies roughly 2,083A, while 400V lowers that to 250A and 800V to 125A. Lower current makes it easier to manage conductor size, busbars, connector heating, and voltage drop. The tradeoff is that higher voltage makes insulation distance, arc control, DC breakers, maintenance procedures, and qualified component sourcing harder. In other words, 800V is not a quick replacement just because it is efficient; it is an architecture change that must be paired with rack-density and facility-level design updates.

Note: theoretical current = power ÷ voltage between conductors (conversion losses and power factor ignored). A +/-400V system is 800V pole to pole, but holds the voltage to ground at 400V to ease insulation and protection requirements.

NVIDIA says that once rack power exceeds 200kW, pushing more current through 54V DC distribution hits a physical limit in copper busbar weight, volume, and heat. The 200kW row in the table is that threshold; at 1MW, a 48V-class system would need more than 20,000A, which is no longer realistic for in-rack distribution.

800 V DC rack adoption (as of October 2026)

For 1MW-class racks, the question is shifting from whether 800V is the right architecture to whether the equipment it needs is available. As of October 2026, products and deployment sites have appeared at the facility, rack, and board levels.

Compared with 48V and +/-400V, 800V's main weakness, the lack of compatible equipment, is being closed between late 2026 and 2027. What remains is how facilities design DC interruption, insulation, and maintenance procedures, and which combination of vendors they buy from. Vendor rollout plans are covered in 800VDC Rollout: Vertiv, Foxconn Lead the Way, and the in-row power system in Delta: 660 kW 800VDC In-Row Power.

How to choose: where the transition stands

The voltage choice depends on the rack power range, design tolerance for insulation and DC protection, and procurement lead times for 800V-compatible components. The current state is that 48V remains the proven standard, while the transition toward +/-400/800V is beginning. Facilities, power supplies, and devices are updating their designs in the same direction. As of October 2026, a practical line is this: racks planned above 200kW cannot be served by stretching 48V, and new 1MW-class builds for 2027 onward should assume 800V and work backward from equipment lead times.

In practice, it is better to treat 48V, +/-400V, and 800V as architectures for different facility refresh points rather than as a single winner-take-all choice. Existing halls can retain 48V, new high-density zones can adopt +/-400V or 800V, and the distribution voltage can rise with server generations. For procurement and architecture teams, the practical work is to connect each voltage choice to supplier availability, protection design, and the facility refresh calendar. Related articles cover the overall architecture and transition schedule in more detail.

For the background, product timeline, and new design problems of the 800V DC shift itself, see The 800V DC power shift for AI data centers.

Reference FactCards