As AI processors demand more power, can eliminating a conversion stage improve efficiency and free up space inside high-performance servers?
Navitas Semiconductor has developed a 24kW power delivery board in collaboration with NVIDIA, designed to convert 800V direct current (DC) to 12V DC in a single stage for next-generation AI datacentres.
The board targets 97% efficiency at full load and power density exceeding 2,000W/in³. By removing the intermediate 50V-to-12V conversion stage, it aims to reduce power losses, board space and power-delivery complexity while supplying an entire compute tray.
Designed in NVIDIA’s 400mm × 60mm form factor, the board can be integrated close to graphics processing units (GPUs) and other compute processors, helping minimise delivery losses as AI workloads increase power demands. Its architecture provides 12V directly to the voltage regulator module (VRM) that supplies power to the processors.
The design uses Navitas’ gallium nitride (GaN)Fast technology, incorporating 16 NV6034 650V, 18mΩ GaN field-effect transistors (FETs) on the primary side. These are arranged in a stacked half-bridge topology and use dual-cooled DFN 8 × 8 packaging.
The development extends the companies’ earlier 20kW, 800V-to-6V power delivery board, offering another intermediate-voltage option for AI server power architectures. Such designs are intended to help infrastructure manufacturers address rising compute power requirements while improving power density and simplifying conversion stages.
The 24kW board is planned for display in NVIDIA’s AI Factory MGX ecosystem partner area at the OCP Global Summit 2026. Its stated efficiency and power-density figures are targets, rather than confirmed operating results.
“The next era of AI will be defined not only by how much compute we can create, but by our ability to deliver the power needed to unleash its full potential,” says Chris Allexandre, President and CEO of Navitas Semiconductor.
Click here for the official announcement.
