New gallium-nitride reference designs target AI rack power efficiency as data center energy demands climb The race to power AI data centers just got a new contestant, and it’s not a hyperscaler. Power Integrations, the semiconductor company best known for making compact power conversion chips, pulled back the curtain on two ultra-slim auxiliary power supply reference designs at COMPUTEX in Taipei on June 1. The designs are built specifically for Nvidia’s 800 VDC Kyber architecture, the liquid-cooled blade-rack system that underpins the next generation of megawatt-scale AI infrastructure.
What Power Integrations actually built #
The two reference designs are built on Power Integrations’ 1700 V PowiGaN gallium-nitride technology, a platform the company has been developing for high-voltage industrial and data center applications.
The first is a 15 W single-output model measuring 30 mm by 30 mm by 7 mm. The second is a more capable 35 W six-rail isolated design at 80 mm by 60 mm by 8 mm. Both designs achieve at least 88% efficiency across line and load conditions. They also deliver approximately 30% space savings on power distribution boards, along with a roughly 30% reduction in bill-of-materials costs on main power distribution boards.
Power Integrations has been working alongside Nvidia on 800 VDC architecture since at least 2025. The collaboration was referenced in a white paper on 1250 V and 1700 V PowiGaN technology, released at the OCP Global Summit on October 13, 2025.
Why 800 VDC is the number that matters #
The broader context here is an industry-wide migration away from legacy 54 VDC power distribution toward 800 VDC systems. Higher voltage means less current for the same amount of power, which translates directly into thinner copper cables, fewer conversion stages, and lower resistive energy losses across the entire distribution chain.
The old 54 VDC standard made sense when racks consumed tens of kilowatts. Modern AI training clusters, particularly those running Nvidia’s latest GPU configurations, can demand power at megawatt scale per rack. At that level, 54 VDC becomes a liability.
Power Integrations’ gallium-nitride transistors are the enabling technology here. Compared to traditional silicon-based power devices, GaN can switch at higher frequencies and handle higher voltages in a smaller physical footprint, which is exactly what the Kyber architecture demands from its auxiliary power supply components.
Market implications for investors tracking AI infrastructure #
Power Integrations trades on Nasdaq under the ticker POWI. The Nvidia collaboration is a meaningful signal for investors. Design wins with Nvidia on a reference architecture carry long lifecycle implications. Once a power design is validated and integrated into a reference platform, the switching costs for alternatives are high. Other system integrators building on the Kyber architecture will have a strong incentive to use components that are already validated against the reference design.
For data center operators, the efficiency gains translate directly to operating cost reductions. At scale, an 88%-plus efficient auxiliary supply versus a less efficient alternative means real savings on electricity, cooling, and physical space. There is also an indirect angle for crypto infrastructure. Mining facilities, particularly those running high-density GPU or ASIC configurations, face the same fundamental challenge as AI data centers: delivering enormous amounts of power into dense compute hardware as efficiently as possible. Power Integrations made no direct references to crypto or blockchain applications in the announcement.
What sets this announcement apart from generic GaN power device offerings is the specificity of the Nvidia integration. A reference design validated against Nvidia’s Kyber architecture at COMPUTEX is a credentialing moment that is difficult for competitors to replicate quickly.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our