Gallium Nitride Has Replaced Silicon for AI Power Needs
As AI workloads demand gigawatt-scale power, data centers are shifting to gallium nitride to maintain efficiency.
Updated on Sept. 21, 2026 in Data Centers
Live Poll
Do you expect the expansion of AI infrastructure to increase your household's electricity costs?
Data centers have begun adopting gallium nitride (GaN) power architectures to support the surging energy demands of modern AI accelerators. This shift addresses the performance limits of silicon-based power components that can no longer meet the high current and power density requirements of future processors.
Why it matters
AI computing has seen a fourfold increase in power consumption across three GPU generations, forcing data centers to rethink power distribution. GaN technology now provides a 5 percent increase in total efficiency from AC input to the GPU core, a critical improvement as facilities scale toward gigawatt-level power usage.
EPC seventh-generation GaN devices currently switch at 3 MHz in synchronous buck converters, offering higher performance than silicon MOSFETs which improve by roughly 20 percent per generation. These GaN components support 800 V DC bus voltage distribution directly into server racks to optimize space and efficiency.
The players
EPC
A developer of gallium nitride-based power management technology with eight years of deployment history on AI data center cards.
The details
GaN devices function by using a semiconductor material that allows for significantly faster switching frequencies than traditional silicon metal-oxide-semiconductor field-effect transistors (MOSFETs — semiconductor devices used to switch or amplify electronic signals). By integrating this power conversion directly onto the server board, data centers can eliminate bulky power supply drawers. This architecture reduces energy loss, enabling the cooling of future GPUs projected to dissipate 5 kW of power within the next two years.
Timeline
July 2026: Article originally published in an e-guide.
September 2026: Article published on the EE Times website.
Within 2 years: Individual GPUs projected to dissipate 5 kW.
The Tech Race
As silicon MOSFET technology nears its practical scaling limit, the move to GaN represents a critical race for power density. This transition is essential for manufacturers to keep pace with future GPU requirements that may eventually demand primary distribution buses of 1,200 V or higher.
This transition primarily affects the physical design of server hardware, allowing for more compact and efficient rack layouts. Future AI-focused data centers will likely utilize these 800 V architectures to manage the massive power density required by upcoming, high-draw GPU hardware.
The takeaway
The move to gallium nitride is a necessary response to the compounding energy needs of modern AI, replacing aging silicon standards. Watch for the industry's progress toward 10 MHz switching frequencies, which will be a key performance metric for the next generation of power components.
Further reading
For more on how infrastructure is evolving to support new hardware, visit Data Centers.
Live Poll
Do you expect the expansion of AI infrastructure to increase your household's electricity costs?
