Nvidia Kyber Ultra NVL576 accelerator rack to consume 600 kW
The new Nvidia Kyber Ultra NVL576 accelerator rack will consume a massive 600 kW of power, posing major challenges for modern data centers.

The rapid expansion of the artificial intelligence market is increasingly hitting power grid limitations. Starting next year, this energy bottleneck could become an insurmountable barrier for the industry as next-generation hardware demands unprecedented electricity supplies.
Qisqacha: - The Nvidia Kyber Ultra NVL576 rack consumes 600 kW of power. - The system integrates 576 Rubin Ultra accelerators via CPO technology. - Performance reaches 15 exaFLOPS for inference and 5 exaFLOPS for training.
Architecture and Rubin Ultra Performance
Nvidia recently announced the Kyber Ultra NVL576 rack, featuring 576 Rubin Ultra accelerators linked together using Customer Performance Optic technology. Prior to this, the largest announced system was the Kyber Ultra NVL144 containing 144 accelerators.
According to official specifications, the Rubin Ultra NVL576 delivers a staggering 15 exaFLOPS of 4th-order floating-point performance for data inference and 5 exaFLOPS of 8th-order performance for data training. This is 14 times higher than the capability of the Blackwell GB300 NVL72.
Read alsoKing Slide controls up to 90 percent of AI server rail marketPower Consumption and Infrastructure Crisis
However, this computing monster requires 600 kW of power, which is equivalent to the consumption of roughly 500 average American homes. For comparison, a Blackwell accelerator rack required about 120 kW, while the previous Hopper generation consumed only 40 kW.
The challenge goes beyond the aggregate power usage of data centers. Retrofitting existing facilities to supply 600 kW directly to individual racks is practically impossible. Consequently, appropriate data centers will need to be built entirely from scratch, requiring massive additional investments.
Implications for the AI Industry
The introduction of such power-hungry hardware shifts the economics of data center construction. Technology companies and investors will have to allocate unprecedented budgets not only for purchasing advanced Nvidia accelerators but also for building entirely new power supply infrastructures.
Source: iXBT.com



