Equinix Liquid Cooling in 60 Facilities for 100 kW AI Workloads
- Analysis Cooling & Heat Recovery Data Centers Efficiency & Optimization
Marin Holt
- 0
- 6 minutes read
- Equinix offers liquid-ready cabinets in 60+ facilities; HK1 pilot achieves PUE of 1.2 vs 1.4-1.5 air cooling
- Liquid cooling retrofit: $5K-$15K/rack, 18-30 month payback; 10 MW facility saves $1.5M/year at PUE 1.1
- Industry average PUE plateaued at 1.55-1.59 since 2020; Equinix achieved 28% improvement over 5 years
- AI workloads push rack density from 5-10 kW to 100 kW; GPUs generate up to 1.2 kW heat each
- AI data center market: $236B (2025) to $933B (2030) per MarketsandMarkets
Your browser does not support audio.
Equinix now offers liquid cooling in over 60 global data centers. Equinix’s liquid cooling expansion supports AI workloads demanding 100 kW per cabinet, a tenfold increase from the typical 5-10 kW (Equinix Blog, 2025-11-12). A pilot at Equinix’s HK1 data center in Hong Kong, a collaboration with Dell Technologies and Schneider Electric, recorded a PUE of 1.2. This compares to 1.4-1.5 for air-cooled systems and projects 2,000 MWh annual savings per facility at scale (Equinix Blog, 2025-11-12).
Cooling Technology and Retrofit Economics
The transition to direct-to-chip liquid cooling addresses a fundamental thermal limitation: GPU accelerators used in AI training and inference generate sustained heat loads that exceed the capacity of traditional air cooling systems. Liquid cooling retrofits cost $5,000 to $15,000 per rack depending on density requirements (RCR Wireless, 2025-09-16), with payback periods of 18 to 30 months through reduced electricity consumption. A 10 MW facility reducing PUE from 1.4 to 1.1 saves approximately $1.5 million per year in cooling energy costs at $0.12/kWh (Johnson Controls, 2026-01-15). Equinix’s xScale facilities in Tokyo and Frankfurt are designed for rapid liquid cooling adoption (RCR Wireless, 2025-09-16).
Industry PUE Trends
The industry average PUE has plateaued at approximately 1.55-1.59 since 2020, according to Uptime Institute analysis. Hyperscale operators report PUE between 1.04 and 2.0, with the wide range reflecting the mix of legacy and purpose-built facilities. Equinix has achieved a 28% efficiency improvement across its portfolio over five years through ISO 50001-aligned operations, ASHRAE A1 Allowable temperature standards, and site-level optimization targets (Equinix Blog, 2025-11-12). The company’s AM3 facility in Amsterdam uses aquifer thermal energy storage and free-air cooling to achieve a design PUE of 1.19 (Equinix Blog, 2025-11-12).
Critical Analysis
At 100 kW per cabinet (10-20x the traditional 5-10 kW baseline), GPU server racks produce dense switch-mode power supply currents injecting characteristic 5th and 7th order harmonics at the 480 V distribution bus. Boosting rack power density from 5-10 kW to 100 kW across 60 facilities will increase aggregate metered demand tenfold to twentyfold. This occurs without proportional infrastructure expansion.
The Gap
The rapid increase in power density per rack, driven by AI, presents a significant challenge to existing power distribution infrastructure. While liquid cooling addresses thermal loads, the electrical capacity and harmonic mitigation at the rack and facility level require substantial upgrades, which are not always explicitly detailed in efficiency projections.
5-Year Projection
Supply chain bottlenecks for Direct-to-Chip Liquid Cooling are projected to be severe over the next five years. This will drive developers to alternative topologies and domestic manufacturing (RCR Wireless, 2025-09-16).
Critical Perspective
The article forecasts 2,000 MWh annual savings per facility at scale. This projection assumes a PUE reduction from 1.5 to 1.2, a smaller improvement than the 1.04 PUE achieved by Google’s data centers. The 2008 Energy Independence and Security Act mandated efficiency standards that did not anticipate the harmonic loads of 100 kW AI racks. Will the existing 480V distribution infrastructure handle the harmonic distortion from 100 kW AI cabinets without significant upstream filtering?
