# AI GPUs generate so much heat that liquid cooling is becoming standard in data centers

**Date:** 2025-12-18  
**Tags:** DataCenter, GPU, Cooling  
**URL:** https://kelexine.is-a.dev/til/liquid-cooling-gpu

---

TIL: AI GPUs generate so much heat that liquid cooling is becoming standard in data centers. Two approaches: direct-to-chip (liquid contacts heat spreader) and immersion (entire server submerged in non-conductive fluid). Immersion cooling can cut cooling energy by 90% compared to air cooling.


```text
Cooling Comparison for AI Racks:

Air Cooling:
- Max density: 20-30 kW/rack
- PUE overhead: 0.4-0.5

Direct-to-Chip Liquid:
- Max density: 50-80 kW/rack  
- PUE overhead: 0.1-0.2

Immersion Cooling:
- Max density: 100+ kW/rack
- PUE overhead: 0.03-0.05
```




---

*This content is available at [kelexine.is-a.dev/til/liquid-cooling-gpu](https://kelexine.is-a.dev/til/liquid-cooling-gpu)*