On August 1, NVIDIA updated its list of partners for its next-generation 800V high-voltage direct current (HVDC) power architecture, naming Chinese GaN specialist Innoscience (2577.HK) as the only supplier from China. The company will provide full-chain GaN power solutions for NVIDIA's Kyber rack systems, marking a significant breakthrough in AI data center infrastructure. The announcement triggered a sharp rally in Innoscience's Hong Kong-listed shares, which surged nearly 64% intraday before closing up around 50%, with over HK$2.6 billion in trading volume.

NVIDIA had first announced the 800V supplier alliance during Computex in May, positioning it as a key step toward supporting 1MW GPU server racks starting in 2027. The alliance includes global power and semiconductor leaders such as Infineon, MPS, Navitas, Rohm, STMicroelectronics, and Texas Instruments, alongside power module and system providers including Delta, Flex Power, Lite-On, Eaton, and Vertiv.

Innoscience's entry into this group underscores China's growing competitiveness in next-gen power semiconductors. The company specializes in GaN-on-Silicon devices across 15V–1200V applications and operates the world's largest 8-inch GaN wafer production line, using a full IDM model spanning chip design, epitaxy, manufacturing, and testing.

Why 800V HVDC?

Traditional 54V power distribution in data centers has reached physical limits, with each rack capped around 200kW. The next-gen Kyber system aims to increase rack power density beyond 300kW and room-level density by over 10×. The 800V HVDC system enables direct conversion from 13.8kV AC to 800V DC at the power entrance, eliminating multiple intermediate conversion steps. This reduces energy loss, improves end-to-end efficiency by 5%, and shrinks rack footprint by up to 40%.

Higher voltage also reduces copper usage by 45% and cuts cooling requirements, significantly lowering total energy consumption. Maintenance costs are expected to drop by as much as 70% thanks to fewer PSU failures.

GaN: The Key Enabler for Efficient Power Conversion

As AI workloads push GPU power beyond 10kW and rack-level demands exceed 600kW, traditional silicon struggles to keep pace. Gallium Nitride (GaN), with its wide bandgap (3.4eV), higher electron mobility, and compact form factor, is ideally suited for high-frequency, high-efficiency power conversion in constrained data center spaces. Innoscience's GaN devices achieve up to 98.5% peak efficiency and enable power densities of over 92W/cm³.

IDM Strength and Strategic Role

Innoscience's IDM model gives it a critical advantage. Its self-developed 8-inch GaN-on-Si fab achieves 95% yield and 40% cost reduction over peers. Its three-tier high/mid/low voltage DC-DC conversion architecture offers a complete solution from grid to GPU. NVIDIA's selection of Innoscience reflects confidence in its technical capabilities, production scale, and system integration.

Toward Greener AI Infrastructure

This partnership marks a milestone in defining the power standard for next-gen AI infrastructure. With Microsoft, Google, and other hyperscalers exploring similar GaN-based HVDC paths, Innoscience is poised for rapid expansion. Industry estimates suggest 800V architectures could cut data center energy usage by 20–30% and reduce carbon emissions by millions of tons annually. Adoption could drive a trillion-yuan retrofit market over the next 3–5 years.

As NVIDIA's Blake Mathias puts it: “We're not just building faster GPUs. We're rebuilding the entire power stack to unlock the full potential of AI.”