Expert brief
Liquid cooling: the thermal revolution for H100/B200 GPUs
Updated 2026-06-06•By Antoine Cazals — Voltarione•Read ~3 min
AI racks reach 100–130 kW vs 8–15 kW air. Direct-to-chip and immersion become mandatory. PUE gain: 0.1 to 0.25.
Sources & references
Data and methodologies based on:
Read next
- Battery PPA / BESS tolling: monetising storageA BESS tolling agreement leases a battery's capacity (MW + MWh) to an offtaker for a fixed fee.
- Green hydrogen PPA: powering an electrolyserUnder RFNBO (Renewable Fuels of Non-Biological Origin), green H2 requires an additional PPA + hourly matching + geographic correlation with the electrolyser..
- AI data centers: 100 kW/rack, where to find the power?AI racks (NVIDIA H100/B200) reach 100 kW density, vs 8 kW historically.
- Grid connection wait: 6 months to 8 years in EuropeTSO/DSO connection lead times are stretching: 6 mo in Norway, 14–18 mo TenneT DE, 36 mo National Grid UK, up to 96 mo EirGrid IE for >40 MW..
- PPA & energy glossaryAll key definitions on one page.