Navitas has announced a 24 kW power-delivery board developed with NVIDIA for AI data centres, designed to convert 800 volts directly to 12 volts in one stage. The company says that removes a conversion step and targets 97% efficiency at full load, a useful attempt to tackle the less glamorous question of how to feed increasingly power-hungry computing.
NVIDIA Watch analysis
What happened
The board is designed for NVIDIA’s MGX ecosystem and is planned for display at the OCP Global Summit. Navitas says it can deliver 24 kW, with power density above 2,000 watts per cubic inch, and fits NVIDIA’s 400 mm by 60 mm form factor. It is intended to supply an entire compute tray.
The single-stage design skips the existing 50-to-12-volt intermediate-bus conversion. Navitas says the simpler path can reduce power losses, board space and delivery complexity. These are company-stated specifications and a target, not independent test results.
Why it matters
Power conversion is part of the AI infrastructure bill, not a footnote to the GPU. Cutting a stage could help operators deliver power more efficiently and make room for the hardware doing the computing. The practical test will be how the board performs in a complete system, not how neat the diagram looks.
Our read
This is a specific engineering move against a real constraint: getting more power to AI computing hardware without letting conversion losses and bulky circuitry take an ever-larger bite. The 97% figure is a target, so it deserves a measured eyebrow rather than a standing ovation. Still, a direct 800-to-12-volt design is more concrete than another promise to make AI infrastructure “more efficient”.
What to watch
- Whether the OCP showcase provides measured performance under real workloads.
- When the board will be available for deployment, and in what systems.
- Whether the target efficiency holds across operating conditions, not just at full load.
Discussion spark: If a simpler power path can make AI data centres more efficient, should operators prioritise that gain even before the board’s real-world performance is proven?
Sources and evidence
Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.