Liquid Cooling Becomes Key to Improving AI Data Center Compute Efficiency
The Green Grid uses power usage effectiveness, or PUE, to measure the ratio of a data center’s total electricity consumption to the power used by its IT equipment. Generative AI is rapidly driving up the power density of GPU racks, pushing conventional air cooling closer to its thermal limits. If too much electricity is consumed by cooling, additional power cannot be converted proportionately into computing capacity, making liquid cooling a core component of AI data centers rather than an optional feature.
As of July 20, 2026, the latest reporting showed the industry’s focus shifting from securing more electricity to conserving it. Liquid cooling’s greater heat-transfer efficiency can support high-power racks while reducing cooling energy use, improving PUE and increasing effective computing capacity. The report identified no specific companies and disclosed no construction costs or project launch dates. For now, the key development is that cooling efficiency has become a direct constraint on expanding AI computing capacity.
All Coverage
1 original reportsThe Backstory
The history behind this eventThree Keys to Reliable Liquid Cooling for AI Data Centers
AI training and inference are driving up server heat density, increasingly pushing traditional air cooling beyond its limits. Power consumption per NVIDIA GPU has risen from 300W for the V100 and 400W for the A100 to 700W for the H100/H200 and 1,200W for the Blackwell GB200. Liquid cooling has consequently evolved from a supporting system into core infrastructure essential to service continuity and operating-cost management.
In an interview published by TechNews on June 23, 2026, Hung-Fu Lee, market development representative at Swagelok Taiwan, identified three priorities: precisely controlling flow and pressure; ensuring coolant compatibility with fittings and sealing materials; and maintaining long-term seal reliability. Power per AI and HPC rack has increased from the traditional 4–10kW to 30–100kW or more. The report did not disclose project investment figures.
AI Data Centers Confront Power and Cooling Constraints as Infrastructure Overhaul Accelerates
As generative AI models grow and GPU power consumption continues to rise, the chief data center bottleneck is shifting from computing capacity to power availability and cooling efficiency. Taiwan's Market Intelligence & Consulting Institute (MIC) said power demand from a single rack could reach hundreds of kilowatts, beyond what conventional power delivery and air-cooling systems can support. This is directly affecting the pace of data center construction and operating costs.
MIC's latest analysis shows the industry is accelerating the adoption of high-voltage direct current power systems (HVDC), battery backup units (BBUs) and liquid cooling, while evaluating new energy sources such as nuclear power to improve supply reliability and energy efficiency. The information did not disclose the report's publication date, investment amounts or deployment timelines, but demand for racks drawing hundreds of kilowatts has become a key threshold for infrastructure transformation.
Subscribe to Mark Radar Weekly
Every Friday, the week's strongest signals in your inbox. Unsubscribe anytime.