AI Briefing0.003–0.005kWh per AI Token: Direct Liquid Cooling (DLC) and PUE 1.1 Design for 80kW per Rack
·2026.10.21 09:57
Key point
Explains the energy required for AI token generation and the principles of Direct Liquid Cooling (DLC) and PUE 1.1 design for handling 80kW power density per rack.
Details
- AI token generation involves energy consumption, with the total system power consumption estimated at approximately 0.003–0.005kWh per token for the latest hyperscale models
- The power intake capacity of medium-to-large data centers in the Seoul metropolitan area is in the range of 22–30MW, where 30MW corresponds to the power consumption of a city with a population of 100,000
- The power density per rack in GPU-based AI data centers (AIDC) is rising to 80kW, leading to the limits of air cooling and the cooling paradox
- Direct Liquid Cooling (DLC) has a heat transfer capability more than 4,000 times higher than air, enabling a PUE of 1.1 and allowing over 90% of total power to be focused on AI computation
- A high-efficiency system that uses only 3MW of a 30MW power intake capacity for cooling and allocates 27MW to token production is a core element of infrastructure competitiveness
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Data Centers in the AI Era: How to Cool the Scorching Heat of GPUs
kt cloud Unveils Goal to Secure 1GW AIDC Capacity and Composite AI Strategy at 2026 Summit[Operations Guide] How kt cloud Data Center Chillers Work and How to Optimize Cooling Efficiency