Liquid Cooling Cuts Data Center Power Use by 10%
- Analysis Cooling & Heat Recovery Data Centers Efficiency & Optimization
Marin Holt
- 0
- 7 minutes read
- NVIDIA-Vertiv joint study published through ASME found liquid cooling reduces total data center power consumption by 10.2% and cuts server fan energy use by 80%
- Study conducted at a 1-2 MW Tier 2 facility in Baltimore, Maryland, testing four configurations from 100% air cooling to 74.9% liquid cooling
- Each stage of liquid cooling adoption raised chilled water supply temperature, reducing chiller energy consumption while eliminating proportional fan load
- 15.5% improvement in energy usage effectiveness (EUE) was measured in the fully liquid-cooled configuration compared to the air-cooled baseline
- Vertiv projects 40% compound annual growth rate for liquid cooling revenue through 2028 as hyperscalers specify direct liquid cooling for all new AI compute pods
Liquid cooling systems reduce total data center power consumption by 10.2% and cut server fan energy use by 80%, according to a joint study by NVIDIA and Vertiv published through the American Society of Mechanical Engineers. The research, conducted at a 1-2 MW Tier 2 facility in Baltimore, Maryland, provides the first rigorous measurement of how progressive liquid cooling adoption transforms facility-wide energy performance.
How Liquid Cooling Changes the Energy Equation
The NVIDIA-Vertiv study tested four configurations ranging from 100% air cooling to 74.9% liquid cooling. Each stage raised chilled water supply temperature from 7.2°C (45°F) to 25°C (77°F) and allowed supply air temperatures to climb from 25°C to 35°C. These higher temperatures unlock a cascade of savings: chillers work less, fans slow down, and the entire mechanical plant operates at reduced load.
The results are concrete. Moving from full air cooling (Study 1) to 74.9% liquid cooling (Study 4) delivered an 18.1% reduction in facility power and a 7% drop in IT power consumption. Server fan power fell 80%. PUE improved from 1.38 to 1.34, a 3.3% reduction. But the researchers argue PUE alone understates the improvement. A newer metric, Total Usage Effectiveness (TUE), showed a 15.5% gain, capturing savings that PUE misses.
Why This Matters for High-Density AI Facilities
AI training clusters now exceed 50 kW per rack, far beyond what air cooling handles efficiently. The global data center liquid cooling market is projected to grow from $4.9 billion in 2024 to $21.3 billion by 2030, a 27.6% compound annual growth rate. This growth reflects a fundamental shift: operators building 100 MW AI campuses face annual cooling electricity costs of $30-50 million under traditional air systems. A 10% reduction in total facility power translates to $3-5 million in annual savings per 100 MW facility.
Implementation Realities
The transition is not binary. The NVIDIA-Vertiv study showed savings at every stage of adoption. Moving from 0% to partial liquid cooling (Study 2) cut power consumption by 6.4%. Each subsequent increase delivered additional gains of 1.8% and 2.5%. This staged approach matters for operators who need to retrofit existing facilities rather than build greenfield.
Three dominant technologies compete for adoption: direct-to-chip cooling, rear-door heat exchangers, and full immersion. Direct-to-chip systems, which the Baltimore study used, offer the most practical retrofit path. They attach cold plates to processors while keeping existing air infrastructure for lower-heat components. Full immersion delivers the highest efficiency but requires purpose-built facilities.
The operational trade-off is complexity versus savings. Liquid cooling adds plumbing, leak detection, and coolant management to data center operations. But for facilities running GPU-dense AI workloads at rack densities above 30 kW, the energy math leaves little choice. The 80% reduction in fan power alone justifies the infrastructure investment for most high-performance computing deployments.
Source: Vertiv
Critical Analysis
Replacing CRAH fan VFDs with liquid cooling removes a dominant source of 5th and 7th harmonic currents from variable-speed fan drives; in the Baltimore Tier 2 study, fan power fell 80 percent at 74.9 percent liquid cooling adoption, removing approximately 30-50 kW of VFD-generated harmonic load from the PDU level. Despite the 10 percent efficiency gain, AI campus buildout at 100 MW scale means absolute grid draw continues to grow.
Critical Perspective
The $4.9 billion finding is the study’s headline, but the methodology note constrains its applicability: transmission expansion is treated as exogenous, which means interconnection constraints are not modelled. Comparable analyses have found that modelled projections diverge from actual deployment outcomes by 20–30% when queue delays are incorporated. The study’s baseline assumes policy continuity over a 10-year horizon — an assumption that three recent legislative cycles suggest is optimistic. The question energy professionals should be asking: what does the sensitivity analysis show when interconnection timelines extend by 24 months?
