STMicroelectronics has expanded its 800 VDC power conversion portfolio with two new architectures designed specifically for next-generation AI data centres, developed in close collaboration with NVIDIA and showcased during NVIDIA GTC 2026. The new 800 VDC-to-12V and 800 VDC-to-6V solutions complement ST’s existing 800 VDC-to-50V stage, completing a comprehensive ecosystem for gigawatt-scale AI compute infrastructure that prioritises efficiency, scalability and sustainability.
The announcement underscores how power architecture is becoming as critical as compute itself in the race to build hyperscale AI facilities. As AI training clusters and inference farms demand ever-higher densities, traditional 48V or 54V distribution systems are hitting physical limits around efficiency and thermal management. ST’s 800 VDC approach reduces conversion losses, minimises copper usage and supports the diverse form factors emerging across GPU generations and server configurations.
Power Innovation Meets AI Infrastructure Demands
Marco Cassis, President of Analog, Power & Discrete, MEMS and Sensors at STMicroelectronics, explained that AI infrastructure scale requires system-level innovation tailored to different server form factors. “With these new converters for 800 VDC power distribution, ST brings a complete set of solutions to support the deployment of gigawatt-scale compute infrastructure with more efficient, scalable, and sustainable power architectures,” Cassis said.
The new 800 VDC-to-12V converter eliminates the traditional 54V intermediate stage, streamlining the path from rack-level power shelves to the voltage domains that feed advanced AI accelerators. This direct conversion improves overall rack efficiency, reduces material costs and simplifies integration for future GPU generations.
Meanwhile, the 800 VDC-to-6V architecture targets system builders who need power stages positioned close to the GPU itself. By minimising IR drop and improving transient response under fast load changes, it addresses one of the most demanding requirements in large-scale training clusters where GPUs cycle rapidly between low and peak power states.
A Complete Portfolio for Diverse AI Server Architectures
ST’s expanded portfolio now covers the full spectrum of intermediate DC buses — 50V, 12V and 6V — that will coexist in future AI data centres depending on rack density, GPU configuration and cooling strategies. The company combines technologies across silicon, SiC and GaN power semiconductors, analog and mixed-signal ICs, and microcontrollers to achieve peak efficiency and power density.
This builds on ST’s October 2025 prototype of a compact GaN-based LLC converter operating directly from 800V at 1 MHz, delivering over 98% efficiency and exceeding 2,600 W/in³ power density in a smartphone-sized footprint at 50V output. Those benchmarks set the stage for the new stages, which include high-density power delivery boards optimised for the specific challenges of 12V and 6V conversion.
Why 800 VDC Matters for AI at Scale
The shift to 800 VDC distribution reflects fundamental changes in data centre design driven by AI workloads. Higher voltage reduces current and therefore resistive losses in cabling, while enabling power shelves that can support entire racks of next-generation accelerators. For OEMs and hyperscalers, the ability to mix and match conversion stages gives flexibility to optimise for different workloads — from dense inference racks to massive training superclusters.
As NVIDIA continues to define the architecture of AI infrastructure through its reference designs, partnerships like this one with STMicroelectronics become critical enablers. The result is a power ecosystem that keeps pace with compute roadmaps, ensuring that energy efficiency keeps up with the relentless demand for more FLOPS.
