Providing power to AI data centers was a key theme in 2025. But distributing power within them may be just as important for 2026+. More power-dense AI chips require upgrading in-rack buses from 54V to 800V. This requires specialized power distribution boards further downstream. This 14-page report explores the challenges and opportunities of future AI power delivery architectures.
A new AI data center being built in 2026 might well be following NVIDIA’s GB300 NVL72 configuration. To an energy analyst, the truly astounding feature of cutting-edge AI data-centers is not just the handling of information. It is also the handling of power. Each rack’s 72 x Blackwell GPUs demand 0.7-1.4kW of high-quality DC, with no more than 1V across each compute circuit. Today’s cutting-edge power delivery architectures are thus described, from kV-scale facility input to tensor core level, on pages 2-5.
A key hotspot is the bus, the large metal structure, spanning the back side of the rack, conveying power to all of the individual compute trays. Today’s standard 54V bus cannot support racks with more than 200kW power density due to the copper losses. They cannot support future chip sets like Rubin and Feynman. Some companies are trying to solve this by making the buses bigger, or by liquid-cooling them. We argue this is a losing strategy, on pages 6-7.
The only viable solution to enable the scale-up of AI – future chips, rising power density – is to switch to an 800V DC architecture. This also has great advantages for maintenance costs and integrating energy storage, as discussed on page 8.
However the key challenge for future 800V AI power delivery architectures lies in the Power Distribution Boards, in each compute tray, which must now down-convert 800V inputs into lower voltage levels that can be used by the chips. LLC transformers based on wide-bandgap semiconductors (GaN) need to replace simple buck converters. This might sound like Greek, but the note explains what all of this means, and why it matters, on pages 9-11.
Only a handful of companies can likely provide an end-to-end solution for these power distribution boards. As illustrated, for example, on page 12. Fascinatingly, these solutions derive from some of the power semis that were previously being developed for electric vehicles, per page 13.
We are optimistic on the long-term outlook for copper, due to rising demand across electrification, robotics, and emerging world air conditioning! 2025 was also a year of terrible supply disruptions, propelling copper prices to $11/kg, and we see marginal cost rising towards $12-14/kg to incentivize sufficient new mine development. The implications of changing AI power delivery architectures on copper demand are discussed on page 14.