AI Power Supply in the Compute Race: How Low-Voltage GaN Supports 800V DC Architecture

Computing Power Race, Power Supply Looks to GaN Source: Semiconductor Horizons Published: 2026-10-30 12:07 AI GaN-generated poster GaN is an important support for AI power supply infrastructure, and is especially suited to low-voltage, high-frequency power conversion scenarios.

Artificial intelligence is reshaping computing systems, but behind every new generation of GPUs and accelerators lies a less conspicuous challenge: how to deliver more power in a smaller space while achieving higher efficiency and faster dynamic response. As AI processor power consumption moves from hundreds of watts to the multi-kilowatt level, traditional silicon-based power architectures are approaching their performance limits. This is driving a fundamental transformation in the field of AI power conversion, from traditional low-frequency architectures to high-frequency, high-density power systems.

Gallium nitride (GaN), especially low-voltage GaN, is becoming the key technology for solving such challenges. Low-voltage GaN power devices combine low on-resistance, high-speed switching characteristics and high power density, making them very well suited to the final-stage power supply link of AI systems. But this technological iteration is not merely about replacing silicon devices with GaN; it also requires simultaneous technological breakthroughs in magnetic components, capacitors, packaging, sensing, thermal management and system architecture.

As AI data centers shift to 800 V DC power distribution, power conversion is also becoming increasingly distributed, and coordination between GaN devices and the entire power conversion ecosystem is becoming crucial. The future of power electronics technology increasingly depends on whether these technologies can work together.

From Low-Voltage GaN to High-Density AI Power Supplies

For many years, GaN power devices have mainly been used in scenarios where high switching frequency and high efficiency can offset their higher cost and greater design complexity. But this trade-off logic is changing rapidly, especially in the low-voltage field.

EPC co-founder and CEO Alex Lidow said the technology has reached a critical juncture: in the 40 V and below voltage range, GaN transistors have surpassed silicon MOSFETs in multiple performance metrics. At the same time, manufacturing economics continue to improve, allowing the technology to reach a much larger application market. "Just in the past year, low-voltage (40 V and below) GaN transistors have overtaken MOSFETs in all performance metrics."

This is significant because low-voltage power conversion accounts for a high share of the entire power electronics market. At a fixed power level, the lower the voltage, the greater the current, and the more critical conduction losses and power density become. For AI data centers, low-voltage scenarios are particularly important: the final-stage power conversion circuit needs to deliver extremely large currents to GPUs and accelerators operating at very low voltages.

The power consumption of a single AI processor has already risen sharply. "What we are seeing is processor power consumption rising from hundreds of watts to 2 kilowatts or even 3 kilowatts," said Ian Chan, chief technology officer of Cyntec. This brings two coexisting design constraints: conversion efficiency and installation space.

Power converters need to carry extremely large currents in an increasingly small volume. Higher switching frequencies can reduce the size of passive components such as inductors, transformers and capacitors, easing the pressure. But higher frequency also amplifies the shortcomings of other links in the power ecosystem. "In AI data centers, the space for each compute tray is very limited. GaN stands out because it supports higher switching frequencies, reduces device size, and achieves the high power density needed to support large currents."

Based on his experience developing GaN power modules, Chan gave an example at the system design level. A power module targeting about 1 kW and 48 V to 12 V conversion can use just 4 GaN devices, whereas a silicon superjunction MOSFET solution requires 8. He also mentioned that the GaN solution can operate above 1 MHz, while the silicon solution is only about 500–600 kHz, further reducing the size of magnetic components and other passive components.

This means the design logic of power converters is undergoing a fundamental change. Semiconductor device performance alone is no longer enough. GaN transistors support higher switching frequencies, creating the conditions for miniaturizing magnetic components, but the magnetic components themselves must also maintain high efficiency at the same frequency. Magnetic components, capacitors, current sensing, gate drivers, controllers, PCB layout and thermal management all become inseparable parts of the design.

This is also the significance of the cooperation between EPC and Cyntec. The two companies are at different links in the same power conversion ecosystem: one provides GaN power devices, while the other focuses on high-frequency magnetic components and passive component solutions.

Therefore, GaN cannot simply be regarded as a direct replacement for silicon devices. Only by fully leveraging GaN's high switching speed and low loss characteristics and optimizing the entire converter globally — covering semiconductors, magnetic components, packaging and PCB layout — can its value be realized. Engineers need to consider the entire power conversion loop, including parasitic inductance, magnetic losses, gate drive and heat dissipation paths, to unlock the full potential of GaN.

From 800 V DC to the Chip: Integration Becomes the Next Breakthrough

The next major challenge comes at the architecture level. Rack power consumption continues to rise, and AI data centers are beginning to shift to 800 V DC power distribution. Compared with long-distance low-voltage high-current transmission, high-voltage distribution can greatly reduce conduction losses and achieve more efficient power transmission.

Shifting to 800 V DC is not just a change in bus voltage, but a large-scale restructuring of AI data center power architecture: power conversion links are distributed across different levels such as the computer room, rack and processor.

In this architecture, AC power is converted to an 800 V DC bus near the rack; subsequent multi-stage conversion then steps the voltage down progressively to supply GPUs and other accelerators.

Lidow describes this as a physical split of the power supply function: the side-mounted power unit performs the initial conversion and establishes the 800 V DC bus, while distributed conversion modules closer to the load handle the final voltage conversion.

The input-series output-parallel (ISOP) topology is a widely watched approach. Low-voltage DC-DC modules with inputs in series and outputs in parallel can both withstand high-voltage input and retain the advantages of low-voltage power stages.

EPC99040 800V to 6V

"The 800 V primary side will use a multi-stage ISOP topology. We can stack low-voltage DC-DC converters with inputs in series and outputs in parallel to achieve very high conversion efficiency," Lidow said.

This architecture also explains why low-voltage GaN still has value after the industry moves to an 800 V DC bus. The high-voltage distribution layer and the low-voltage conversion stage at the load end solve different problems and can accommodate different semiconductor technologies.

EPC91123 800V to 12.5V

The challenge moves from the rack down to the processor

Modern GPUs and AI accelerators need to draw enormous currents at extremely low operating voltages. If such large currents were transmitted directly over conventional PCB traces, they would produce huge conduction and distribution losses. As a result, the final-stage power conversion circuit is moving ever closer to the processor package.

Chan said this trend is forcing a disruptive change in power module form factors. Cyntec's customers require converters only a few millimeters tall, compared with about 9 mm for previous-generation products.

Under such size constraints, power converters must integrate inductors, capacitors, semiconductor chips, sensing components and other devices within an extremely tight vertical space. "In the past the Z-height was 9 mm; now customers want it below 3 mm. Below 3 mm, we have to embed and integrate capacitors, inductors and chips all together."

This is not just a packaging challenge but a system-level optimization problem. Passive components must maintain acceptable losses at ever-rising frequencies; semiconductor devices must simultaneously achieve extremely low on-resistance and high-speed switching; packaging must suppress parasitic inductance and plan heat dissipation paths; and converters must respond quickly to the severe transient currents generated by AI processors.

The requirements at the point of load converge: high-frequency switching, low conduction loss, low parasitics, high current capability, thermal performance and extremely fast transient response must all be achieved within the same structure.

Low-voltage GaN is highly attractive in this scenario. Devices rated at around 18 V can achieve extremely low on-resistance and are expected to outperform silicon MOSFETs in the same setting. Combined with high switching frequency and excellent voltage withstand capability, they can reduce both conduction and switching losses while shrinking passive components.

But the real opportunity lies not in improving a single parameter, but in the coordination among semiconductors, magnetic components, capacitors, packaging, sensing, control and thermal design.

For power engineers, next-generation AI power conversion increasingly requires a co-design approach rather than optimizing components individually. This is also why the new generation of AI power systems will move toward integrated platforms rather than piles of discrete devices. This logic applies beyond data centers. AI-equipped robots, humanoid robots, autonomous vehicles, drones and other devices need both powerful computing and strict control of weight, size, efficiency and heat dissipation.

Lidow believes AI is the most immediate driver today, followed closely by what he calls "AI kinetics" — physical machines that rely on AI to move, interact and operate autonomously. GaN combines high frequency, low loss and easy miniaturization, giving it the opportunity to be widely deployed in these emerging platforms.

The lesson for power electronics engineers is that the traditional boundaries between semiconductor design, magnetic components, packaging and system architecture are blurring.

Future power engineers will need to master the entire technology chain

As Lidow said, advanced magnetic components may be one of the largest incremental opportunities for improving system efficiency in the future. Chan reaches a similar conclusion from another angle: engineers need cross-disciplinary knowledge covering power devices, passive components, packaging and system requirements.

The evolution of AI power infrastructure is not just a voltage upgrade or a power increase, but a move toward highly integrated power systems in which every component is optimized as part of a whole. GaN is an important enabler of this transformation, especially for low-voltage high-frequency power conversion scenarios. But its full value can only be unlocked by co-designing semiconductor devices with magnetic components, capacitors, packaging and control architecture.

As AI computing demand continues to rise, the next competitive advantage in power electronics may not come from a single breakthrough component, but from the ability to fuse multiple technologies into a simpler, more efficient, higher-density power architecture. GaN is a key enabler of this transformation, but its true impact depends on what complete solutions engineers build around it.

This article is reproduced from media reports or online platforms and represents the author's personal stance or views. Our republication is for sharing only and does not mean we endorse or agree with it. If the source is incorrectly attributed or your lawful rights are infringed, please contact customer service promptly, and we, as a neutral platform service provider, will promptly correct, delete or handle it in accordance with the law. Comments: No user comments yet. Latest News More Related Recommendations

Intel appears at Tencent Global Digital Ecosystem Summit, accelerating agent deployment through multi-dimensional innovation 2026-10-29 22:43

Physics-based verification, the cornerstone of trustworthy agentic EDA 2026-10-29 18:24

Revisiting chip design shift-left 2026-10-29 18:22