Innoscience outlined its roadmap for enabling next-generation AI data center power delivery through an end-to-end “All-GaN” architecture aligned with the evolving NVIDIA MGX ecosystem and emerging 800 VDC power distribution standards.

As AI factories scale toward megawatt-level rack power, efficient power conversion has become a critical bottleneck. Innoscience positions gallium nitride (GaN) technology as a key enabler of future AI infrastructure, citing its ability to deliver higher switching frequencies, lower losses, improved thermal performance, and greater power density than conventional silicon solutions.

The company’s roadmap spans the entire AI power delivery chain, from 800 VDC rack distribution down to GPU core voltages. At the front-end conversion stage, Innoscience demonstrated a 12 kW 800 V-to-48 V all-GaN LLC converter utilizing 650 V and 100 V GaN devices, achieving approximately 99% peak efficiency and 98.2% full-load efficiency while operating at 1 MHz. The company has also introduced 150 V GaN devices that reduce secondary-side synchronous rectifier requirements by 50%, simplifying system design and improving power density.

Beyond the 48 V architecture, Innoscience is extending its portfolio to support emerging 800 V-to-12 V and 800 V-to-6 V conversion topologies that are gaining traction in next-generation AI server designs. These architectures aim to reduce conversion stages, lower distribution losses, and move power conversion closer to GPUs.

For intermediate bus conversion, the company highlighted its 100 V GaN portfolio for 48 V-to-12 V multiphase buck converters, targeting higher efficiency and power density in AI servers where even marginal efficiency gains can significantly reduce cooling and operating costs at data center scale.

At the point-of-load level, Innoscience is developing 15 V DrGaN technology for vertical power delivery (VPD) architectures. Operating at switching frequencies between 3 MHz and 5 MHz, these solutions are designed to support future GPU power requirements by reducing passive component size, improving transient response, and enabling power stages to be located closer to AI accelerators.

Original – Innoscience Technology