NVIDIA’s Rubin Ultra AI GPU Sparks Surge in Power-Hungry AI Servers
NVIDIA’s Rubin Ultra AI GPU Redefines AI Server Power Needs – Over 1,000 kW Predicted by 2029.
The new Nvidia Rubin Ultra AI GPU and HBM4 memory will bring a new level of raw power to future AI systems. Experts believe that by 2028 or 2029, AI computers could consume more than 1,000 kW of electric power, while now it only consumes 130-250 kW.
Next-Gen AI Servers: Blackwell Ultra & Rubin
The modern AI servers deliver highly competitive rack densities of up to 250 kW, even though new generation platforms such as NVIDIA’s Blackwell Ultra and Rubin AI servers will set new standards. They have estimated that between the years 2026 and 2027, they will need between 250-900 kW per rack in the servers that will be able to take up to 576 GPUs in a single configuration.
This path indicates a move toward ultra-dense racks for unique and different workloads and next-generation applications. Vertiv, the market leader in this industry testifies to this trend as it predicts future peak rack densities to exceed 1 mW thanks to the innovations made by NVIDIA.
AI’s Growing Appetite for Power
Such a change in consumption is an indication that AI is on an unrelenting expansion trajectory. Bulk power density capability in 2020 has only been around 8.2 kW per rack. Today, the highest power-consuming server is NVIDIA’s GB200 NVL72 AI cabinet that comes with up to 132 kW which has actually been informed by the growing demand for AI-based computational solutions.
NVIDIA Rubin Ultra AI GPU claims performance like no other but makes vital directions about data center efficiency and sustainability. The current growing and transforming AI industries still require infrastructure changes to cater to the power-demanding capabilities of AI.
