AI servers of the future are expected to use over 1000 kW of power, with the rise of NVIDIA's next-gen Rubin Ultra AI GPU and its new HBM4 memory coming that will see a rise in power consumption of AI servers.
The current best AI servers and clusters with a rack density of 130 kW to 250 kW of power according to an analyst by data center firm Vertiv. But these numbers are going to skyrocket with Blackwell Ultra and Rubin AI servers using between 250 kW and 900 kW of power with up to 576 GPUs in a single rack expected in 2026-2027.
After that, NVIDIA will have its future-gen, beefed-up Rubin Ultra AI GPUs inside of new AI servers in 2028-2029+ that will use 1000 kW or more. Vertiv notes that "densification roadmaps point to extreme peak rack densities of 1mW+ resulting in estimated new build average rack density increasing across varying workloads applications and market segments".
Rewinding the clock back to just 4 years ago, in 2020 the average rack density was rated at around 8.2 kW, but with the insatiable demand of AI, that power consumption has skyrocketed to the likes of 132 kW with NVIDIA's new GB200 NVL72 AI cabinet, which is the highest power-consuming server in history.
NVIDIA's next-gen Rubin and Rubin Ultra AI GPUs are going to be quite exciting, with the new Rubin AI GPU architecture and future-gen HBM4 memory, which you can read more about here.
