The power challenge in AI data centers is no longer limited to GPU power consumption or accelerator efficiency; it has extended to the entire path that carries electricity from the grid to the gates inside chips. As server-rack power rises from approximately 125 kilowatts today to 500 kilowatts and then one megawatt, infrastructure designers are turning to 800VDC engineering to reduce the number of power-conversion stages, lower losses and heat, and save copper and space inside data centers.
Peter Wawer, head of the Green Industrial Power division at Infineon Technologies, said that server-rack power is currently approximately 125 kilowatts before rising to 500 kilowatts and then to an 800-volt DC-DC conversion architecture when reaching the megawatt level. The chain typically begins with a medium-voltage grid of up to 35 kilovolts, where high-voltage AC must be converted to suitable levels, then to direct current, and ultimately to less-than-one-volt levels required by GPUs and other accelerators.
From 48 Volts to 800VDC
Traditional data centers have generally relied on 48-volt or 12-volt architectures. Electricity passed through multiple stages, including an uninterruptible power supply (UPS), a power-conditioning and distribution unit (PDU), and then conversion to 48 volts before reaching levels such as 1.2 or 0.7 volts inside the graphics processor. Steven Lee, director of power-electronics design software products at Keysight Technologies, says this approach was suitable when data-center requirements were far below the levels imposed by AI workloads.
In the proposed 800VDC architecture, some AC stages can be integrated into a single AC-DC interface using silicon carbide. Instead of moving sequentially through numerous stages, the path becomes shorter, reducing switching losses and the heat associated with each stage. Pavani Gottipati, director of applications engineering at Synopsys, believes that the transition by component suppliers and data-center operators from the conventional 415VAC architecture to 800VDC requires rethinking component design, including the introduction of solid-state transformers.
Increasing the voltage does not itself increase the required power, but it lowers the current needed to transmit the same amount of power. For this reason, Pradeep Shenoy, a computing power technology specialist at Texas Instruments, believes that remaining at around 50 volts will not be practical at the new power levels, even with liquid-cooled busbars. He also notes that moving through intermediate levels such as 400 volts may not be attractive if rack power continues to rise rapidly, because facilities may have to redesign again to reach 800 volts.
Higher Efficiency, Less Space and Copper
The 800VDC architecture enables higher power density, lighter transformers, and less space for power-distribution components. At the same power level, the higher voltage produces lower current, allowing thinner wires to be used instead of the thick conductors required at 48 volts. This is especially important in AI data centers, where transmitting power at high currents may consume large quantities of copper and impose thermal and mechanical constraints on racks.
Even modest improvements in the efficiency of each stage accumulate across the conversion chain. Shenoy said that moving from a 12-volt input to 6 volts at one power-conversion stage could provide approximately a 2% efficiency gain and enable approximately 30% more power within the same area. He added that a one-percentage-point increase in overall efficiency could have a significant operational and financial impact for cloud-service providers and data-center operators.
The benefits are not limited to data centers. Puneet Sinha, senior director and global head of battery technology at Siemens EDA, pointed to a trend among several companies toward 800-volt battery architectures, with potential charging benefits, although this also imposes new requirements on inverters and the electronics associated with the system.
The Role of SiC, GaN, and Solid-State Transformers
Solid-state transformers are gradually replacing traditional iron-core transformers in some 800VDC designs. A solid-state transformer can be connected directly to the high-voltage AC grid, then output low-voltage AC or convert directly to 800VDC, depending on the selected architecture.
Wide-bandgap materials such as silicon carbide (SiC) and gallium nitride (GaN) provide higher switching speeds and lower thermal losses, while also enabling smaller and lighter magnetic components. Compared with traditional wire-wound transformers, planar transformers help reduce height and use space more densely inside racks.
However, transitioning from silicon to SiC or GaN is not a direct replacement. Higher switching speeds increase electromagnetic-interference and noise challenges and require stable control loops, different gate-drive signals, and suitable magnetic design. Material selection remains linked to the voltage class and application; IGBT transistors may remain suitable for some slower-switching converters, while Wawer believes that high-voltage, fast-switching solid-state transformers are particularly well suited to SiC. GaN is also used because of its efficiency, power density, and ability to operate across a wide voltage range.
What Changes in Practice for the Architecture and Operation?
Power requirements are changing even closer to the loads themselves. Power-management integrated circuits (PMICs) have become central components in system performance rather than background elements, because they generate power rails closer to the processor and memory, provide more precise voltage regulation, real-time measurement, and integrated sequencing, while reducing conversion losses and supporting rapid load changes.
The memory subsystem is also under additional pressure as DRAM speeds increase. Any voltage drop may cause errors in memory timing and consequently data corruption or other server failures. For this reason, memory-power solutions have moved to PMICs integrated into DDR5 modules, while challenges related to load current, voltage-regulation accuracy, and response to sudden changes remain.
At the chip level, redesigning the power network alone is not sufficient. Shenoy emphasizes that increasing the processor's computational efficiency remains a fundamental path, while companies such as Imagination Technologies are seeing demand to expand efficient GPU technologies from edge and automotive applications to larger scales targeting data centers. John Weil of Synaptics notes that the evolution of AI models, including vision and language models, is increasing demand for greater capacity over short periods.
By contrast, the transition to 800VDC presents a major safety challenge, particularly when upgrading existing data centers. Hoa Tram, a senior product engineer at Cadence, says that disconnect equipment, breakers, fuses, and isolation switches designed to interrupt AC faults must be replaced with equipment rated for direct current. New measurement and alarm systems must also be integrated, and personnel must be trained to handle the different equipment.
Design variations further increase complexity; some facilities may use an 800-volt unipolar bus, while others may prefer a ±400-volt bipolar architecture, affecting interoperability, protection systems, and the training and certifications required for personnel. At the same time, some data centers may remain on a 48-volt architecture while benefiting from more efficient converters. For example, a laboratory prototype developed by engineers at Binghamton University demonstrated a single-stage point-of-load converter with 10% to 12% higher efficiency and a slew rate approximately twice as fast as the previous rate.
These developments show that 800VDC is not a standalone component, but rather a redesign of the power chain from the grid to the gate inside the chip. This vision requires more efficient converters, materials, semiconductors, and power-management circuits, along with updates to protection, cooling, and operation. However, the slow construction of infrastructure and power grids may limit the speed of deployment, even as demand rises from tens or hundreds of kilowatts at the rack level to megawatts, and then to tens or hundreds of megawatts at the data-center level.