US Liquid Cooling editorial · Reviewed · 4 min read
AI servers can be cooled with surprisingly warm water when the hardware is designed for it. A higher permitted coolant supply temperature can make outdoor dry cooling practical for more hours of the year. It does not mean every data center can remove its chillers, or that every liquid-cooled server accepts 45°C water.
The useful question is where that temperature is specified, and whether the complete heat-rejection path can meet it under local operating conditions.
Why is 45°C part of the current discussion?
In its Vera Rubin platform technical explanation, NVIDIA describes warm-water, single-phase direct liquid cooling with a 45°C supply temperature. Its June 2026 cooling article connects that architecture with possible chiller-less operation in favorable climates.
Those are platform and facility-design statements. They are not permission to raise the supply temperature of an existing installation to 45°C without checking its equipment requirements.
The distinction matters for both US and European projects: a headline about a new AI platform is a reason to revisit design assumptions, not a replacement for a site-specific specification.
How can warm water cool a much hotter chip?
Heat moves from a hotter surface toward a cooler one. The coolant does not have to feel cold to a person to accept heat from an operating processor. The cold plate, interface materials, flow and permitted chip temperature determine the usable thermal path.
The server manufacturer defines the acceptable inlet conditions. A limit might depend on configuration, load or other environmental parameters. Obtain the conditions for the actual server and cooling assembly, including what happens during a fault or a brief temperature excursion.
Avoid treating “warm-water capable” as a complete technical specification. It should lead to a documented operating envelope.
Follow the temperatures from outside air to the rack
Consider a simplified dry-cooling system with a facility loop and a separate technology loop connected by a CDU. A first-pass temperature budget might be written as:
Technology supply temperature ≈ outdoor dry-bulb temperature + dry-cooler approach + CDU approach, with additional allowance for distribution and operating effects.
For an illustrative case, 32°C outdoor air plus an assumed 5°C dry-cooler approach and 3°C CDU approach gives about 40°C technology supply. At 40°C outdoor air, the same assumed differences give about 48°C. These are explanatory arithmetic examples, not equipment selections or guaranteed operating points.
Actual approaches vary with duty, flow, equipment size and fan operation. The example shows why the climate and heat exchangers matter as much as the advertised server inlet limit.
Chiller-less does not mean infrastructure-free
A dry cooler still needs heat-exchanger area, fans, controls and suitable fluid circulation. The facility still needs a resilience strategy and a response when conditions exceed the intended envelope.
Ask whether “chiller-less” means all-year design operation, normal operation with emergency refrigeration, or only a predicted reduction in compressor runtime. Those are materially different proposals.
DCX's 8 MW CDU announcement is one example of suppliers responding to 45°C AI cooling requirements. Its stated capacity and approach figures belong to specified product conditions, not to every installation using warm water.
What about water consumption?
Dry heat rejection and evaporative heat rejection have different site-water implications. A warm-water server may enable a low-water design, but the presence of a closed server loop alone does not establish the facility's total water use.
Identify the actual heat-rejection equipment, any adiabatic operating mode and the reporting boundary. Keep operational water consumption separate from the initial fill and from water associated with electricity generation.
Our water-use explanation covers those boundaries. This article addresses the specific temperature condition that can change the heat-rejection options.
What should a project team request?
Request the server inlet envelope, liquid-capture fraction, design heat loads, CDU selection conditions, hourly climate assessment, worst-case heat-rejection performance and fault sequence. Ask for a clearly labeled temperature diagram with normal and degraded operating cases.
Then compare the proposed system with your current plant. The opportunity is real when supported equipment, climate and infrastructure align. The headline temperature is only one input to that decision; our heat-rejection planning guide helps organize the rest.
Sources & further reading
- Inside the NVIDIA Vera Rubin Platform: Six New Chips, One AI Supercomputer | NVIDIA Technical Blog ↗
- Hotter Than a Hot Tub: The 45°C Breakthrough to Cool AI’s Biggest Machines | NVIDIA Blog ↗
- 8 MW CDU announcement ↗
Explore the company directory → · All liquid-cooling guides →