Johnson Controls launches its AI Factory Reference Design at the Data Center Expo, featuring zero-water cooling for gigawatt-scale AI data centers.
What a zero-water AI factory design actually targets
Johnson Controls has introduced an AI Factory Reference Design at the Data Center Expo, aimed at gigawatt-scale AI facilities that cool without consuming water. That framing matters because AI training and inference clusters pack heat into dense racks, and conventional evaporative cooling trades electrical efficiency for continuous water use. In drought-prone regions, or anywhere water rights are constrained, that tradeoff can block a project long before power or fiber become the bottleneck.
A reference design is not a turnkey building. It is a documented system architecture that shows how mechanical, electrical, and controls pieces fit together so operators, engineers, and builders can adapt the pattern to a specific site. For AI campuses, the useful parts are the cooling topology, redundancy assumptions, and control logic that keep liquid and air loops stable as load swings with batch jobs and model rollouts.
Why water-free cooling is hard at gigawatt scale
Zero-water cooling typically means rejecting heat to outdoor air with dry coolers, refrigerant circuits, or closed liquid loops that do not evaporate process water. Those approaches avoid makeup water and wastewater treatment, but they push more work onto fans, compressors, and heat exchangers. When ambient temperatures rise, free-cooling windows shrink and the plant draws more power just to stay within chip thermal limits.
At gigawatt scale, small efficiency gaps multiply. A design that works for a single hall may not scale if pump head, header sizing, and partial-load behavior are treated as afterthoughts. Operators should expect the reference design to spell out how the plant behaves when only part of the campus is occupied, when one cooling train is offline, and when outdoor conditions leave little margin for dry rejection.
- Confirm the design’s design-day assumptions for outdoor temperature and humidity, not only average conditions.
- Map heat density by hall so liquid distribution matches rack and cold-plate requirements.
- Plan power capacity for peak cooling load, not only for IT nameplate load.
- Define fail-over paths that preserve chip temperatures without relying on temporary water use.
How teams should evaluate the reference design
Treat the Johnson Controls package as a baseline to stress-test, not a checklist to stamp. Ask how the cooling plant stages equipment as AI load ramps, how sensors and controls detect imbalance before temperatures climb, and how maintenance is done without draining or opening the entire liquid path. Integration points with building management systems, power distribution, and fire protection are as important as the chillers or dry coolers themselves.
Also separate marketing language from site physics. Zero-water does not mean zero energy or zero risk. It means the project’s water ledger stays near zero while heat still has to leave the building. Compare total cost of ownership across climate zones: capital for larger dry rejection gear, operating cost for fans and compressors, and the avoided cost of water infrastructure, permitting delays, and public opposition.
Practical next steps for operators and builders
If you are planning an AI campus, use the reference design to align stakeholders early. Mechanical engineers can size loops and redundancy; electrical teams can reserve feeders for cooling peaks; sustainability leads can document water and carbon claims with clear system boundaries. Commissioning plans should include thermal step tests that mimic AI workload swings, not only steady-state setpoints.
For brownfield expansions, check whether existing halls can adopt closed-loop or dry-rejection modules without reopening water-intensive central plants. For greenfield sites, site selection should weight ambient design conditions as heavily as power availability. The value of this kind of announcement is a shared architecture language: zero-water cooling for gigawatt-class AI load, expressed as an implementable plant concept rather than a single product SKU.