Data-center power is moving from "in-rack 48V" to "800V DC"
As power consumption per AI server rack exceeds 100kW and heads toward 1MW, the in-rack 48/54V DC power architecture that has been the standard is nearing its limit. Current equals power divided by voltage, so if power rises tenfold without raising voltage, the current through cables and busbars also rises tenfold, pushing conductor cross-section, copper loss, and heat generation to impractical levels.
The industry's answer to this wall is a sharp increase in the power voltage before it enters the rack: 800V DC (HVDC). NVIDIA describes 800V DC as "the optimal architecture for next-generation power distribution" and says it will lead the transition toward the 1MW rack generation after 2027 (NVIDIA "800 VDC Architecture for AI Data Centers"). This is not simply a story of improving efficiency by a few percent. It changes the backbone of data-center power delivery.
Why 800V: lower current and fewer conversion stages
In a conventional configuration, AC delivered to the facility passes through UPS and PFC stages to become a DC bus, is stepped down to around 54V inside the rack, and is then stepped down again on the server board to the 1V class. The problem is where each voltage is created and how many times power is converted. Every conversion adds a few percent of loss, and conductor loss becomes dominant in low-voltage, high-current sections.
The goal of 800V DC is twofold. First, raise the voltage entering the rack and substantially reduce current at the same power level. Second, reduce the number of conversion stages from AC onward. NVIDIA's technical blog says an 800V DC architecture can improve end-to-end efficiency by up to 5%, reduce maintenance cost by up to 70%, and lower cooling load (NVIDIA Technical Blog). A few percent across the full power chain translates into megawatt-scale facility differences at a 10,000-rack scale.
The broader industry is moving in the same direction. The OCP (Open Compute Project) concept for a separate sidecar power rack, led by hyperscalers, raises distribution from in-rack 48V DC to ±400V or 800V DC, enabling 100kW to 1MW IT racks. NVIDIA is also presenting an 800V ecosystem in line with this direction (NVIDIA Technical Blog), making this less a single-vendor proprietary design than an industry-wide transition axis.
High-current limits
As rack power moves from 100kW toward 1MW, staying at 48/54V makes current enormous and pushes conductor size, copper loss, and heat to impractical levels. Higher voltage lowers current.
Conversion stages and losses
Losses accumulate across AC-to-DC-to-rack-to-board conversion. 800V and DC distribution reduce stages and can improve end-to-end efficiency by up to 5% (NVIDIA).
Standardization and interoperability
OCP's separate power-rack concept points from 48V toward ±400/800V. Hyperscaler-led work is forming a transition axis that is not tied to one vendor.
Products are moving on a 2026 sample and 2027 production schedule
800V DC is no longer just a concept. Compatible products and reference designs have been announced with concrete dates.
NVIDIA says it is developing a 660kW monopolar 800V reference design, with air-cooled samples and production planned for mid-2026 and a liquid-cooled "VR Ultra" derivative planned for sampling in the second half of 2026 (NVIDIA Technical Blog).
Power-semiconductor and power-system suppliers are moving in step. Texas Instruments announced in March 2026 a "complete 800V DC power architecture" developed with NVIDIA (TI Newsroom), while STMicroelectronics has presented 6kW, 12kW, and 20kW power-supply boards (ST Blog).
On the facility side, Schneider Electric is presenting 800V DC optimization beyond 1MW racks and sidecars at GTC 2026, showing that semiconductor, power-supply, and facility layers are all updating around the same 800V direction (NVIDIA On-Demand GTC26). In other words, this is advancing as a supply-chain-wide ecosystem transition rather than one company's roadmap.
The scale of change is moving by orders of magnitude. Voltage moves from 54V to 800V, and target power moves from kW to MW. Efficiency and maintenance-cost gains, which NVIDIA states as up to 5% efficiency improvement and up to 70% maintenance-cost reduction, become material to total site cost.
New design problems from the transition: insulation, protection, and power devices
800V is not an all-upside move. High-voltage DC brings harder safety and protection design in exchange for efficiency.
First is insulation and creepage distance. At the 800V class, insulation design for boards, connectors, and cables, arc countermeasures, and whether hot-plugging is allowed become new constraints. Second is protection. DC has no current zero crossing like AC, making interruption difficult and putting DC breakers and fast protection at the center of design. Third is power-device selection. SiC is advancing in high-voltage conversion stages, while GaN is advancing in high-frequency step-down stages inside the rack, each in the voltage and frequency range where it is strongest. A change in power architecture connects directly to wide-bandgap semiconductor adoption decisions.
Insulation and creepage design
800V-class insulation distance, arc countermeasures, and hot-plug limits constrain connector and board design. Low-voltage assumptions cannot simply be reused.
DC protection and interruption
DC is hard to interrupt because it has no current zero crossing. DC breakers, fast protection, and ground-fault detection become central to safety.
SiC/GaN partitioning
High-voltage conversion stages favor SiC, while high-frequency in-rack step-down favors GaN. 800V directly affects WBG device selection.
Procurement lead times
800V-compatible power supplies and interruption equipment are part of a new supply network. Lead time from design freeze to production and multi-sourcing affect business plans.
What each role should check next
800V DC has lifted AI data-center power from an in-rack topic to a design theme spanning facility and grid. The next question differs by role.
- Power and circuit design: How much of a 48/54V-based design must be rebuilt for 800V? SiC/GaN partitioning by voltage range and DC protection implementation become selection points.
- Facility and infrastructure: Should the design assume sidecar-style separate power racks? How should heat removal, distribution, and protection for 1MW racks be aligned with facility design?
- Procurement and technology planning: Which design axis, NVIDIA or OCP, should the program align with? In line with the 2026 sample to 2027 production schedule, when should lead times and multiple sources for 800V components be secured?
The common question is whether the current 48V design can be carried straight into 1MW scale. In many cases the answer is that some layers cannot. Which layer shifts to 800V first will be a central battleground in data-center power design over the next two to three years.