AI servers are moving from 12V to 48V power distribution because higher voltage delivers the same power with less current, reducing resistive losses and the amount of copper required inside a rack. As accelerator systems approach much higher power densities, the same engineering principle is driving a further transition toward 800V DC distribution, while lower voltages remain necessary closer to the processors.
This is changing more than the power supply. The shift affects how racks are constructed, where electricity is converted, how faults are isolated, which components become more valuable, and how much computing equipment can fit within a given physical and electrical footprint.
By October 2026, the transition had become a coordinated industry effort. NVIDIA, Google and Microsoft were working through the Open Compute Project (OCP) to standardize 800V DC interfaces, and NVIDIA had contributed additional rack-level designs at the October 2026 OCP Global Summit. The technology that made 48V attractive is now exposing its own limits.
The arithmetic: four times the voltage, one-quarter the current
The physical argument begins with two relationships: electrical power equals voltage multiplied by current (P = VI), while resistive power loss equals current squared multiplied by resistance (Ploss = I2R).
For a fixed amount of delivered power, increasing distribution voltage from 12V to 48V reduces the required current to one-quarter. If conductor resistance remains unchanged, resistive losses fall to one-sixteenth of their previous level.
Consider a hypothetical 100kW electrical load distributed at different voltages:
| Distribution voltage | Current at 100kW | Relative conductor loss* | Architectural context |
|---|---|---|---|
| 12V DC | 8,333A | 16.0× | Legacy low-voltage distribution |
| 48V DC | 2,083A | 1.0× | High-density rack distribution |
| 800V DC | 125A | Not directly comparable | Emerging facility-to-rack distribution |
*Illustrative calculations at the same delivered power. The 12V and 48V loss comparison assumes identical conductor resistance; 48V is the 1.0× reference. The 800V architecture operates over a different distribution boundary and should not be compared as though it used an identical physical circuit. The 12V case is a mathematical illustration, not a recommended 100kW rack design.
The improvement is substantial, but its boundary matters. A sixteenfold reduction in conductor loss does not mean sixteen times greater power-supply efficiency or sixteen times lower data-center electricity consumption. Conversion losses, cooling systems, connector resistance, redundancy and other loads remain.
Conductor design also changes when voltage rises. Engineers may use less copper, reduce conductor resistance, free space for other equipment, or combine these benefits. They do not necessarily retain identical busbars.
The practical economic gain comes from making high-power distribution manageable within the rack’s space, thermal and materials constraints.
A server rack contains several voltage domains
The term “48V architecture” describes a distribution level within a larger power-delivery network. It does not mean a GPU operates at 48V.
Electricity typically reaches a data center as medium-voltage alternating current. Transformers and power-conversion equipment condition that electricity before it reaches server hardware. Inside the system, progressively lower voltages supply processors, memory, networking components and auxiliary electronics.
Three simplified architectures illustrate the evolution:
| Architecture | Simplified power path | Principal change |
|---|---|---|
| Traditional 12V | AC input → PSU → 12V bus → VRM → processor | Low-voltage distribution across the server |
| 48V rack | AC input → power shelf → 48V bus → IBC → VRM → processor | Higher-voltage rack distribution and local conversion |
| Emerging 800V DC | AC/grid conversion → 800V DC → rack converter → intermediate bus → VRM → processor | Higher-voltage distribution closer to or through the compute rack |
Conceptual architectures. Exact intermediate voltages, conversion stages, isolation arrangements and backup systems vary by implementation.
The 48V label itself is nominal. Some equipment operates around 50V or 54V, while individual power modules support a range of input voltages. Monolithic Power Systems, for example, offers intermediate bus converters that accept 40–60V and generate lower-voltage outputs.
Consequently, specifications describing 48V, 50V and 54V systems may refer to closely related power architectures, although their operating limits and component requirements are not necessarily identical.
Inside a 48V AI server: the conversion chain
Moving to 48V relocates part of the engineering challenge from the distribution bus to the converters near the load.
Power shelves and busbars
Instead of placing independent AC-to-DC supplies inside every server tray, a rack can consolidate power conversion into a shared power shelf. Multiple power-supply units feed a common DC bus, while busbars distribute electricity to accelerator and server trays.
This arrangement supports modular power capacity and redundancy. It also reduces the need to route very large currents through multiple individual cable assemblies.
The busbar, however, still carries substantial current. At 100kW, a nominal 48V supply corresponds to approximately 2,083A before accounting for conversion and distribution losses. Connections, contact resistance and thermal performance remain critical.
Intermediate bus converters
An intermediate bus converter, or IBC, reduces the rack voltage to a level suitable for downstream regulators. A common arrangement converts a 48V-class input toward 12V, although alternative architectures use approximately 6V or other intermediate voltages.
Monolithic Power Systems’ 2024 technical portfolio illustrates this design choice. It includes 4:1 converters for approximately 54V-to-13.5V conversion, as well as higher-ratio modules that produce lower intermediate voltages.
Converter topology becomes important. High-frequency switching can reduce magnetic-component size, but semiconductor switching losses, inductance, thermal resistance, electromagnetic interference and output regulation introduce competing constraints.
The appropriate converter is therefore determined by system-level efficiency and power density, including its behavior under partial load and rapid load changes.
Voltage regulators at the processor
The final stage supplies very low voltage at high current to processors and accelerators. Digital multiphase voltage regulators distribute this current across multiple power stages and respond to fast changes in computational workload.
MPS documents power modules with output-current ratings reaching 170A and emphasizes their use in compact, thermally constrained processor power systems.
Multiple stages can be operated in parallel because the total current demanded by a high-performance accelerator can greatly exceed what one module can safely supply.
At this point, the physical distance between the regulator and the processor becomes as important as the voltage-conversion efficiency.
The next constraint is the distance to the processor
Higher rack voltage reduces upstream current. Near the GPU, very low operating voltage still requires large current.
That creates losses across circuit-board traces, package connections, connectors and other elements in the power-delivery network. The faster current changes, the more difficult it becomes to maintain a stable processor voltage.
Traditional lateral power delivery places voltage regulators around the processor on the circuit board. As current rises, the length and electrical characteristics of the path between regulator and load become more demanding.
One response is vertical, or Z-axis, power delivery. Regulators can be positioned on the opposite side of the board, directly beneath the processor, shortening the electrical path. MPS describes this approach as capable of materially reducing power-delivery-network losses, although the result depends on board construction and system design.
Shorter delivery paths can reduce the need for compensating capacitance in some locations. They also create demanding packaging and cooling requirements because high-power conversion equipment is placed near an already hot processor.
Capacitors remain important across these voltage domains. Different technologies provide bulk energy storage, high-frequency filtering and transient response. The commercial implications for ceramic capacitors are examined in Sector Foundry’s analysis of MLCC demand in AI servers.
The power architecture is therefore increasingly co-designed with the accelerator board, package and cooling system.
Why are AI racks moving beyond 48V?
The advantage of 48V eventually encounters the same constraint that made 12V unattractive: current increases with power.
Texas Instruments’ November 2025 architecture review uses a hypothetical 1MW rack to illustrate the limit. Distributing 1MW through a 50V bus requires approximately 20,000A. At 800V, the same power corresponds to 1,250A.
Moving such currents through a rack changes the amount of copper required, the size and number of connectors, the protection system, the available compute space and the thermal-management burden.
Higher-voltage DC distribution can address those problems, but implementation is not simply a matter of increasing the output of the existing power shelf.
A new facility-to-rack architecture can move some AC-to-DC conversion outside the computing rack. A dedicated side power rack may convert existing facility AC distribution to an 800V DC bus, allowing adjacent racks to receive higher-voltage power while preserving much of the existing building electrical infrastructure.
A longer-term design can convert medium-voltage AC to DC more directly through centralized transformer-rectifier equipment or solid-state transformers. This could simplify the power path and reduce the amount of power-conversion equipment occupying computing space.
These alternatives have different capital requirements and implementation timelines.
In August 2026, NVIDIA described three deployment stages: an 800V-compatible power rack for existing facilities, row-level power centers designed to distribute power across larger groups of racks, and facility-scale power blocks for new developments.
At the October 2026 OCP Global Summit, NVIDIA also contributed Vera Rubin rack and compute-tray designs to the open ecosystem. The direction increasingly concerns repeatable electrical interfaces and integration across multiple suppliers.
These developments should be interpreted as architecture roadmaps and ecosystem commitments. They do not establish that the installed data-center base has already converted to 800V.
Higher voltage creates a different protection problem
Moving to 800V brings new engineering costs that are less visible than reduced copper consumption.
Direct-current fault interruption is particularly demanding. Unlike AC systems, DC current does not naturally pass through zero every cycle. Breaking a high-energy DC fault therefore requires appropriate protection devices, insulation, sensing and coordinated control.
A June 2026 technical review led by researchers affiliated with Oak Ridge National Laboratory identifies fault interruption, grounding arrangements, rack hot-swapping and interactions between facility and processor power systems as unresolved design challenges.
The protection strategy must ensure that a fault in one server tray or rack can be isolated without unnecessarily interrupting neighboring racks. Battery backup and capacitor-based energy buffers create additional considerations because stored energy must be managed during faults and switching events.
Grounding also affects personal safety, electromagnetic compatibility and fault tolerance. Different DC grounding configurations provide different trade-offs, and no single arrangement can be assumed appropriate for every facility.
These challenges create demand for DC-rated breakers, solid-state protection, sensing systems, high-voltage semiconductors and system-level controls. They also mean that the electrical infrastructure requires extensive qualification before it can be deployed broadly.
For an operator, the relevant comparison is the installed and operating economics of the complete electrical system, including protection and maintenance, rather than conductor losses alone.
Where the industry economics change
The voltage transition affects several layers of the AI infrastructure supply chain.
Power semiconductors and controllers
Different voltage domains require different devices. Silicon-based power stages, advanced MOSFET processes, gallium-nitride devices and silicon-carbide technologies can compete or complement one another depending on voltage, current, frequency, conversion topology and thermal conditions.
Texas Instruments, Infineon, Monolithic Power Systems, Analog Devices and other suppliers participate in different parts of this system. Higher voltage creates opportunities in power switching, control, sensing and protection; higher processor current simultaneously raises demand for efficient multiphase regulation.
Market growth does not translate uniformly into additional revenue for every device category. More integrated modules may reduce the number of discrete components, even as the value of the complete power-conversion function increases.
Magnetics, connectors and passive components
Converters require inductors, transformers or alternative energy-transfer networks, depending on their topology. Compact planar magnetics can be attractive in power-dense applications because their geometry can support low-profile construction and thermal integration.
At the same time, 48V and 800V architectures alter requirements for capacitors, connectors, busbars and protection hardware. Some specifications become more demanding while other components are consolidated or removed.
Capturing value consequently depends on where a supplier’s product sits in the electrical architecture and whether it satisfies the next generation’s qualification requirements.
Data-center infrastructure and operating economics
For facility operators, the main potential benefits are lower distribution losses, reduced conductor volume and more space available for computing equipment. Centralizing conversion can also change how redundancy, backup power and maintenance are organized.
Those benefits have to be evaluated against conversion efficiency at different loads, equipment acquisition costs, installation complexity, electrical safety, cooling requirements and the cost of replacing existing infrastructure.
The economics are also linked to accelerator deployment. If power infrastructure takes less physical space, an operator may accommodate more computing capacity within an otherwise constrained facility. The value of that additional capacity depends on utilization, networking, cooling and the economics of the workloads running on it.
For component manufacturers, customer architecture selection can influence the timing and size of future capacity investments. Sector Foundry’s analysis of Samsung Electro-Mechanics’ orders and investment provides an example of how AI-related component commitments can interact with capital allocation, although those orders do not establish direct exposure to every new power-conversion technology.
The threshold for the next architecture
A 48V rack remains a practical design when its busbar current, thermal requirements and conversion costs are manageable. At much higher densities, distributing electricity at 800V can become more compelling because it reduces current and allows conversion equipment to be repositioned.
The crossover point is not a universal rack-power number. It depends on conductor length, rack configuration, conversion topology, redundancy, cooling, protection requirements and the cost of electrical infrastructure.
The strongest evidence for the next stage will be commercial deployment of interoperable 800V equipment in operating facilities, with measured end-to-end efficiency, acceptable protection performance and lower total infrastructure cost per unit of useful compute.
Until those results are available at scale, 48V and 800V should be evaluated as complementary layers in an evolving electrical architecture, with each deployment determined by the power density and economics of the system it serves.
Sources & Methodology
Data as of October 26, 2026.
This analysis compares rack-level power distribution, intermediate bus conversion and emerging facility-to-rack DC architectures. The 100kW current examples use I = P/V; the conductor-loss comparison uses P = I²R at equal conductor resistance. These are illustrative engineering calculations, not measured end-to-end data-center efficiency. Equipment specifications and deployment roadmaps are attributed to their publishers, while the 2026 research preprint describes proposed architectures and unresolved technical issues.
Primary sources
- Monolithic Power Systems, “48V Datacenter Solutions,” Q4 2024
- Texas Instruments, “Data Centers Evolve to Meet AI’s Massive Power Needs,” November 2025
- Open Compute Project, “Powering the Next Era of AI,” 2026
- NVIDIA, “Why Scaling AI Compute Performance Requires a New Power Architecture,” August 11, 2026
- NVIDIA, “Amid Ecosystem Adoption, NVIDIA Opens Vera Rubin Architecture,” October 12, 2026
- Lee et al., “Toward Next-Generation AI Data Centers: Power Delivery Architecture Shifts, Emerging Technologies, and Challenges,” June 23, 2026, research preprint
Other sources
- Mouser Electronics, “Why Data Centers Are Moving to 48V Power,” September 4, 2025
- Payton Group, “The Rise of 48V Architecture in Data Centers,” October 11, 2026
- GlobalFoundries, “Powering the Next Architecture Shift: Inside GF’s Approach to 48V,” August 20, 2026
- Infineon Technologies engineers, “Optimizing the 800V to 48V/50V Power Path for AI Data Center Servers,” May 4, 2026