Innovations in Data Center Compute and Power Supply for Large-Scale AI Sustainability

AI data centers already demand 100~200MW of power, with a single site consuming power on the scale of a small city. GPUs consume 600~700W per card, pushing rack power beyond 40kW.

As a result, operators struggle with peak demand charges, and urban power grids bear additional burden. The bottleneck for AI scaling is no longer compute, but power.

The problem lies in the bursty nature of AI inference. Power surges occur within milliseconds, but typical lithium-ion battery-based data center storage carries fire risks when placed next to servers, and has limits in responding instantly.

Rebellions and Standard Energy have partnered to bridge this gap. Rebellions' NPU delivers about 5x higher power efficiency than GPUs, and Standard Energy's Vanadium Ion Battery(VIB) ESS is non-flammable and water-based, enabling placement right next to servers.

• 2.4~3ms response time meets 4ms UPS requirements
• 20,000+ cycles and about 96% efficiency
• About 50% reduction in peak power
• Enables safe ESS placement adjacent to servers

In field validation, a 2,000kW peak was reduced to 1,000kW, and in a Texas case, the monthly electricity bill dropped from $30,000 to $19,200. Peak shaving alone delivered about 30% in savings, and when combined with NPU efficiency, total AI inference costs were reduced by up to 79%.

The next step is combining AI control software with ESS power management systems to predictively handle demand surges. Further ahead, there's a vision of matching AI workloads with intermittent renewable energy sources like solar and wind, turning data centers from a grid burden into a stabilizing device.

This partnership presents more than a single product—it offers a global expansion model combining AI and energy infrastructure. At the same time, it shows that Korean deep tech can set a new standard for AI power infrastructure.