TaichungAir cooling
Self-built and operated
A self-built and operated site with 24 racks in 4 rows at 20 kW per rack. In-row units and a chilled-water system support the Glows.ai GPU cloud.
Data Center Systems
Data center systems built for high-density GPUs at full load
The MEP scope covers everything downstream of the main low-voltage panel: power distribution to each rack, removal of heat from the data hall, and fire suppression that does not damage equipment. Rack power density determines all three specifications. Conventional IDC design standards do not apply to high-power GPU environments.
EVIDENCE｜ Results and data
Four delivered sites using air and liquid cooling
The self-built Taichung site has 24 racks at 20 kW each and hosts the Glows.ai GPU cloud. Zhonghe combines 1.5 MW of power with 900 kW of cooling and can support B300. Fukushima was the first liquid-cooled project, with a 250 kW-class CDU.
ZhongheAir cooling
Turnkey procurement and construction
Turnkey procurement and construction with 1.5 MW of power and 900 kW of cooling. The rack environment can support B300.
YunlinAir cooling
Electrical panels and cooling
Renovation of an existing building, with complete delivery from electrical panels and transformers through racks and cooling.
FukushimaLiquid cooling
Liquid-cooled container site
A containerized design with a 250 kW-class CDU.
SOLUTION｜ The KONST approach
Power, cooling, and fire protection designed around rack density
Rack power density determines panel capacity, air or liquid cooling, and whether redundant feeds are required. A 5 kW conventional server rack and a 40 kW GPU rack need entirely different systems.
Power distribution
MP main panels and RP subpanels divide capacity by zone. Rack PDUs are sized at 5 kW or 40 kW. High-power racks use redundant A/B feeds so a single-feed failure does not shut down the entire rack.
Cooling
Around 20 kW per rack, in-row cooling with chilled water is sufficient. Above 50 kW, airflow can no longer remove the heat effectively, so liquid cooling reaches the heat source directly and a CDU distributes the flow.
Fire protection
Gas suppression avoids the secondary damage that water or powder can cause. It works with VESDA very early warning detection, and the extinguishing agent is nonconductive and leaves no residue.
FEATURE｜ Service details and specifications
Choose air or liquid cooling by rack power
At about 20 kW per rack, in-row units can deliver enough cold air for stable operation. At tens or hundreds of kW, airflow cannot carry away the heat, so coolant must reach the heat source directly. This power threshold determines the cooling design.
Air cooling: in-row units with chilled water
Air cooling is most stable at about 20 kW per rack. Units sit between rack rows and deliver cold air directly to the intake side, while hot and cold aisle containment isolates return air. An air-cooled chiller supplies the cold source.
Liquid cooling: coolant distribution unit (CDU)
At tens or hundreds of kW per rack, air can no longer remove enough heat, so coolant reaches the heat source directly. The CDU manages primary-to-secondary heat exchange and flow distribution.
FEATURE｜ Service details and specifications
Seven systems in a high-density data center
The delivery scope includes every system from secondary-side power distribution through fire protection. The table maps each equipment category to its function.
| System | Equipment | Function |
|---|---|---|
| Secondary-side distribution | PDU panels, MP/RP panels, and RFAC panels | Core data center distribution, zoned power management, and power control |
| Rack power | 40 kW/5 kW PDU power strips | Power distribution for high-power and standard-power racks |
| Cooling system | 30 kW/60 kW rack-mounted precision cooling units | Direct cooling for server heat sources |
| Water system | 60 RT air-cooled chiller, expansion tank, and chilled-water pump | External cooling plant and circulation piping |
| Rack equipment | 60×120/80×120 IT racks | House servers and network equipment |
| Airflow management | Hot and cold aisle containment | Separate hot and cold airflow to reduce recirculation and mixing |
| Fire protection | Gas fire suppression system | Nonconductive, residue-free fire suppression |
PROCESS｜ Process
Three construction stages, then integrated testing and server installation
Onsite work proceeds through water piping, cooling, and UPS installation. Each stage has its own acceptance test: pipe pressure and leak testing, hot and cold aisle containment, and UPS transfer testing. Servers are installed only after integrated testing and commissioning are complete.
Water piping
Complete and insulate the chilled-water supply and return piping, then pressure-test it for leaks.
Cooling system
Install the in-row units and chiller, connect the secondary-side piping, and tune the supply-air temperature.
UPS system
Install the UPS and battery cabinets, then test utility transfer and standby load.
FAQ｜ Frequently asked questions
Frequently asked questions
Should an AI data center use liquid or air cooling?
Choose based on power per rack. Around 20 kW, in-row cooling with chilled water can handle the load at lower engineering and maintenance cost. Above 50 kW, the airflow, noise, pressure, and space requirements of air cooling become difficult to manage, so liquid cooling is more practical. High-density B300 configurations are generally planned with liquid cooling.
Can an existing IDC facility be converted for high-power equipment?
In most cases, yes. Work is concentrated on secondary-side distribution upgrades, additional cooling, and rack layout changes. The main constraints are whether the building has spare utility capacity and whether outdoor or roof space remains for cooling equipment. If both are at their limits, the target density must be reduced.
Is hot and cold aisle containment required?
It is required in high-power environments. Without containment, cold supply air mixes with hot return air before reaching the servers. The intake temperature rises, forcing the cooling system to compensate with colder supply air, which directly increases PUE.
Where do power distribution problems occur most often?
Usually in zone allocation, not total capacity. Designers may distribute racks evenly across subpanels, only for high-power racks to cluster in a few rows during installation. One zone then approaches its limit while another uses only half its capacity. Finalizing the rack layout and distribution zones together avoids this problem.
What is the hardest part of a liquid-cooling project?
The piping. If pressure drop, flow rate, and distribution across nodes are not calculated in advance, some nodes may receive too little flow and develop local hot spots. The issue cannot be observed before power-up.
Are fire detection and suppression the same system?
No. VESDA very early warning detection belongs to the monitoring and management system. Gas release belongs to the MEP system. Both are finalized together during design.
What PUE can the facility achieve?
It depends on climate, cooling method, and load factor. Digital meters in the environmental monitoring system continuously measure operating PUE for monthly tracking. The design target is set for each project's conditions.
Is your data center ready for high-density GPUs?
Get an assessment of the required power and cooling upgrades.