A standard data center sounds like a jet engine and consumes almost as much energy for cooling as it does for computing. Nvidia decided to change both parameters at once — and the starting point was an unexpected idea: cool chips with liquid that is itself hot.
45 degrees — this is a feature, not a bug
Jacuzzis heat up to 38–40 °C — so hot that most people can only stay in the water for about 15 minutes. The liquid in the cooling system of Rubin servers can be even hotter — up to 45 °C. And this higher temperature is what makes the system more energy efficient.
The logic is counterintuitive, but simple: cooling equipment consumes more energy the lower the temperature it needs to maintain. Cooling traditionally consumed up to 40% of a data center's electricity consumption. Raising the coolant temperature by just one degree reduces cooling costs by approximately 4%.
The coolant — a mixture of 75% water and 25% propylene glycol — circulates through cold plates directly on the processors. Liquid enters the rack at 45 °C and exits at approximately 55 °C, absorbing heat from the chip surface — yet the processors continue to operate at full power because liquid plates keep device temperatures within acceptable limits.
First generation without a single fan
Rubin is the first Nvidia infrastructure where absolutely everything is cooled: every chip and every network component, in a closed loop, without a single fan in the system.
Unlike traditional data centers with noise levels of 85 decibels and higher, Rubin's architecture allows for nearly silent operation. In many climate zones, heat can be dissipated through external dry coolers without running mechanical chillers and fans at all.
"The temperature of the surrounding air in the data center doesn't matter — hot summer air is perfectly fine, because no server depends on cold air"
Nvidia blog on liquid cooling for AI factories
Evolution, not revolution
It's worth holding back from excessive enthusiasm. Several cooling specialists note that this is rather the next step on a path the industry has already been following. Blackwell generation systems already operated at inlet temperatures around 40 °C, and some manufacturers were designing equipment for the mid-forties before Rubin's announcement.
Liquid cooling for memory and some other components remains expensive and technically complex — so chillers won't disappear completely. The direction of movement — higher operating temperatures, more liquid, fewer chillers — is real, but this is a continuation rather than a clean break with the past.
Scale: $47 billion per gigawatt
More efficient cooling is not just an engineering detail, but a condition for the entire industry's existence at current growth rates.
- If Hopper required about 40 kW per rack, Blackwell already 120–150 kW, then Rubin is rated at 200–1000 kW depending on configuration.
- According to Foxconn estimates, a data center based on Vera Rubin will cost up to $47 billion per gigawatt of power, and annual electricity costs for such a facility will be approximately $1.3 billion.
- By 2030, global computing infrastructure could consume about 174 GW — more than twice as much as in 2024. To cover this growth, the industry must add approximately 18 GW of new electrical capacity annually.
Analysts increasingly point out that not operational costs, but equipment depreciation will determine data center economics: at an electricity price of $0.15 per kW·h, annual operating costs for one gigawatt will be $1.3 billion, while depreciation over six years — approximately $7.9 billion per year.
If liquid cooling truly removes the energy constraint for AI infrastructure growth — a practical question arises: will power generation manage to add 18 GW per year by 2030, or will the bottleneck already shift from chip heat dissipation to the network infrastructure outside the data center fence?
