From Grid to Core: Powering AI

AI compute is scaling fast, and the power infrastructure behind it needs to keep up. As rack power climbs, moving all that energy efficiently brings challenges with conversion losses, current, thermal management and system complexity.

At this year’s 2026 OCP Global Summit, we’re demonstrating onsemi technologies and showing how we’re tackling the biggest power challenges behind scaling AI.

AI Scalability Requires Power

The evolution toward higher-density AI infrastructure changes the role of the power system. Instead of treating power conversion as supporting infrastructure, designers need to consider the complete journey energy takes from the grid to the compute core.

onsemi’s grid-to-core approach looks at that entire journey. Our silicon, silicon carbide (SiC) and gallium nitride (GaN) technologies address challenges across high-voltage conversion and distribution, intermediate bus conversion and core power delivery.

Conversion from the 800 VDC Bus to 50 V

• EliteSiC™ JFET High Voltage Intermediate Bus Converter
  This compact SiC based power solution efficiently converts up to 6kW of power from 800V to 50V, ideal for 800V AI data centers, energy storage and backup systems. Advanced cascode SiC JFET technology enables switching frequency beyond 700kHz, isolation and top cooling reduce system complexity and make this half-brick robust, safe and fast to implement. Achieving above 98% peak efficiency and 2.5 kW/in³ power density in real-world conditions, this solution enables denser, more efficient, more scalable data center power.
• GaN-based LLC High Voltage Intermediate Bus Converter
  This GaNEXUS based solution efficiently converts up to 12kW of power from 800V to 50V, also for AI data center applications, energy storage and backup systems. The stacked LLC architecture pushes the envelope of high voltage GaN performance with an efficiency of 98.0% at full load and a peak of 98.8% at 600kHz switching frequency. The wide input voltage range of 700V to 900V, two output channels, and top side cooling make this IBC solution flexible, compact and ready for deployment.

Together, these approaches demonstrate how different wide-bandgap technologies both address the challenge of stepping down high-voltage DC distribution while meeting the power density demands of increasingly constrained rack environments.

Moving 800 V Closer to the GPU

• 800 V-to-6 V Vertical GaN High Voltage DC Intermediate Bus Converter
  The high voltage drop from 800V to 6V represents the most difficult conversion in the data center rack. onsemi’s 1200V vertical GaN based solution converts a nominal 20kW at a projected efficiency of 96%. The vertical 1200V GaN architecture eliminates stacking of FETs and employs a unique half-turn matrix transformer to reduce solution size. Further, the ultralow 20mOhm on resistance of the vertical GaN enhances system efficiency and reduces losses. The entire solution achieves a power density of 2.6kW/ in³ and a peak power output of 24.5kW.

The design illustrates an important part of the grid-to-core strategy: reduce the distance, complexity and number of stages between high-voltage distribution and low-voltage rails feeding AI processors. Using 1200 V vertical GaN allows the design to use a non-stacked primary, eliminating the voltage-balancing complexity associated with a stacked 750 V lateral GaN architecture. The half-turn matrix transformer also reduces core size by 20% compared with the reference transformer architecture.

The 800 V-to-6 V demonstrations, including the five-phase power stage and T-shaped board being shown at OCP, bring this concept physically closer to the compute tray and the GPU.

Protecting the 800 V Power Path

• 800 V DC Hot Swap
  This single stage hot swap solution combines a JFET with a hot swap controller to manage startup, faults and live insertion on an 800 V bus, providing critical protection from overcurrent and short-circuit faults. The compact design enhances system reliability and control with capacitive-load charging and options for external enable or automatic startup.

Testing documented for the design includes sustained 20 A conduction for 30 minutes, dual-FET operation, output-short testing and high-dv/dt testing.

This capability complements high-efficiency DC/DC conversion. A scalable 800 V architecture needs both efficient conversion and controlled distribution, including mechanisms that allow boards to energize safely and remain protected when faults occur.

A System-Level Approach to AI Power

These solutions highlight onsemi’s array of system level solutions for 800 V AI data centers, which open the door to the more efficient, next generation approach to AI power. Our work doesn’t stop at the bus, however, as power still requires conversion, protection and efficient delivery throughout the rack and all the way to the processor.

onsemi is helping turn the 800 V vision into real hardware and tackling power challenges across the entire path. From the grid to the core, we’re building the power technology AI needs to keep scaling.