NVIDIA Raises Liquid-Cooling Threshold to 45°C for AI Servers
*NVIDIA’s latest AI systems run their cooling loops at 45 degrees Celsius, a temperature that cuts the energy required to reject heat from large clusters.*
NVIDIA now ships AI servers that tolerate cooling liquid at 45 °C. The change directly lowers the power spent on heat removal in AI factories.
Hot tubs typically run between 38 °C and 40 °C. Most users limit sessions to roughly fifteen minutes at those temperatures. NVIDIA’s specification exceeds that range by five degrees while still protecting the silicon.
The higher inlet temperature allows facility chillers and heat exchangers to operate with smaller temperature lifts. Less compressor work and fewer fans translate into lower total facility power. The company states that the 45 °C limit is the precise point at which the efficiency gain becomes material for clusters measured in tens of megawatts.
No independent benchmarks or third-party validations appear in the announcement. The claim rests on NVIDIA’s internal engineering data for its newest rack-scale designs.
Why it matters
Power used for cooling already accounts for a large fraction of an AI data center’s electricity bill. Raising the acceptable water temperature by a few degrees reduces that overhead without requiring new chip-level thermal solutions. Operators who deploy these systems at scale will see the difference in both operating cost and available grid capacity. Facilities still limited to older 25–30 °C loops will need to upgrade their cooling plants to capture the same savings.
---
Sources:
No comments yet