From megawatt-scale UPS to LiquiX liquid cooling — a full-stack answer to the AI data center's power and thermal crisis.

AI is rewriting the rules of data center design faster than the industry can keep up. As per-rack power densities push toward the megawatt range, conventional air-cooling systems are approaching their physical limits, and legacy power architectures are struggling to scale. The next generation of AI Data Centers (AIDCs) demands something fundamentally different: high-capacity, high-reliability modular UPS systems capable of handling extreme power loads, paired with cooling solutions that can dissipate heat at the chip level. Layer on top of this the Chinese government's mandatory PUE targets — and 2026 marks not an upgrade cycle, but a complete reinvention of how data centers are powered and cooled.

At COMPUTEX 2026, NVIDIA CEO Jensen Huang put it plainly: the modern data center is an "AI factory." Over the next decade, AI will become as foundational to society as electricity and the internet. But keeping that factory running at scale requires power delivery and thermal management to evolve from supporting roles into core system architecture. KSTAR is responding with a comprehensive power-plus-cooling product portfolio — a full-stack solution engineered from the grid down to the chip.

The Dual Crisis: Power Scarcity and Thermal Overload

The AI compute boom has triggered two interconnected infrastructure crises — and both are intensifying simultaneously.

Thermal density is spiking. GPU thermal design power (TDP) has already climbed from tens of watts to 700–1,000W per chip. NVIDIA's NVL72 system exceeds 132 kW per rack today — and projections suggest that Vera Rubin-generation racks, expected after 2027, could surpass 600 kW. Air cooling simply cannot keep pace.

Power delivery is hitting a wall. At megawatt rack densities, conventional 50V low-voltage distribution architectures generate current loads exceeding 20,000A. Busbars become impractically heavy and expensive; the physics of copper distribution at this scale make the traditional approach untenable.

Regulatory pressure is tightening. China's National Development and Reform Commission (NDRC) mandates a national average PUE of ≤1.5 for all data centers by end-2025, with hub nodes required to achieve ≤1.2. For operators building or expanding AIDC, liquid cooling is no longer an optional premium feature — it is a compliance prerequisite.

Power Architecture Evolution: From Distributed Redundancy to Integrated Intelligence

AI workloads place a unique set of demands on power infrastructure: instantaneous load spikes, sustained high draw, and zero tolerance for interruption. Traditional "distributed redundancy" architectures — designed for a different era — were built with generous margins but carry a heavy price: large physical footprints, slow deployment timelines, and significant efficiency losses.

As AI training and inference demand accelerates across Chinese and global markets, the industry is consolidating around a new model: high-capacity, integrated, digitally managed power systems. KSTAR's answer is a vertically integrated power stack built around its full-stack R&D capability — spanning megawatt-scale modular UPS, wind-liquid hybrid micro-module prefabricated units, and lead-acid/smart lithium battery energy storage. Together, these products form a unified, scalable power foundation purpose-built for high-density AI compute environments.

Cooling Evolution: From Airflow Management to Chip-Level Thermal Control

When per-rack power consumption breaks through 20 kW — and approaches 100 kW and beyond — air-based cooling becomes a bottleneck, not a solution. Liquid cooling is making the transition from edge case to standard practice.

LiquiX: KSTARs AI-Native Liquid Cooling Platform

KSTAR's LiquiX liquid cooling solution is engineered to cover the full range of AI data center cooling requirements — from 10 kW to 2 MW of cooling capacity — within a single, end-to-end platform. Key design principles include:

• Full-link architecture: Primary and secondary loop piping is fully prefabricated, enabling faster deployment and higher system reliability.

• 5× cooling density vs. air: LiquiX delivers more than five times the cooling capacity per unit floor area compared to conventional air-cooling systems, enabling dramatic rack density increases.

• Flexible heat exchange: Both air-to-liquid and liquid-to-liquid heat exchange configurations are supported, adapting to different facility constraints.

• Multi-format deployment: LiquiX supports IDB, IDM, and IDU data center formats, with multiple CDU (Coolant Distribution Unit) deployment configurations — accommodating greenfield builds, brownfield retrofits, and co-location scenarios alike.

A Closed-Loop System: From Grid to Chip

What distinguishes KSTAR's approach is the breadth and coherence of its product ecosystem. Rather than delivering point solutions, KSTAR has built a fully integrated power-plus-cooling stack:

• Power side: Megawatt-scale UPS → Integrated power modules → Lead-acid and high-density lithium battery cabinets

• Cooling side: LiquiX liquid CDU → Wind-liquid hybrid micro-module

Think of it as a modular energy architecture — building blocks that can be configured rapidly to meet specific density requirements, and assembled into data centers achieving PUE as low as 1.21.

The race to power AI infrastructure is ultimately a contest across power electronics, thermodynamics, and system integration. With more than 30 years of engineering expertise and deep domain knowledge of AIDC operational requirements, KSTAR has developed a complete, reliable, and future-ready solution for the AI era's compute infrastructure.