Retrofitting data centres for liquid cooling hits pipework and floor loads

Liquid cooling retrofits of existing data centres are feasible but rarely simple, according to a Data Center Knowledge analysis published on 2 November. The first obstacle is pipework. Supply and return lines must run close to servers. Power and network cabling already fill the space in those halls.

Fixing that can mean repositioning racks, reworking cable pathways or building new coolant routes. It can also mean expanding white space outright. Weight is the second problem. Data Center Knowledge reports that fully loaded racks in AI-heavy environments can exceed 3,000 pounds. Raised floors are common in older air-cooled halls. They may not carry the load without reinforcement. Downtime is the third problem. Structural and mechanical work can take months.

Greenfield builds avoid all of it. Designers account for piping, rack loads, redundancy and heat rejection from the start.

Once liquid reaches the chips, the operating model changes. A second Data Center Knowledge piece, also from 2 November, lists what operators must now instrument. The list includes liquid temperature, pressure, flow rate and level. It also includes leak detection and coolant health measures such as pH, turbidity, conductivity and refractive index. Dew point becomes a live risk. Piping colder than the surrounding air can condense water near equipment. A common practice is holding coolant supply at least 2°C above the measured ambient dew point.

The hardware is arriving regardless. At OCP Global Summit 2026 in San Jose, ASRock Rack showed a full NVIDIA Vera Rubin NVL72 rack, according to ServeTheHome's 2 November report. The XGMRACK-VR/NVL72 sits in a 48RU enclosure with a 5000A busbar and a liquid manifold. Inside are 18 compute trays and nine NVLink 6 switch trays, for 36 Vera CPUs and 72 Rubin GPUs. Its four power shelves are rated at 110kW each.

Smaller systems follow the same route. ServeTheHome also described a 2U16X-GNR2/DLC Rubin machine. It fits an eight-GPU HGX Rubin NVL8 board into 2U with no fans at all.

Power delivery is being reworked alongside the cooling loop. On 3 November Tom's Hardware reported that Envision Energy has completed what it calls the first commercial deployment of an integrated 800-volt DC system. The site is its Galaxy AI data center in Ulanqab, Inner Mongolia. Envision has tested it since August. The Dubhe system converts 10kV medium-voltage AC directly to an 800V DC bus. Envision says the campus will scale to 3GW and eventually support up to one million accelerators.

In the UK the binding constraint is the grid. The Guardian reported on 3 November that Drax, Brookfield and Radiant plan a datacentre of more than 1 gigawatt at the Drax site in North Yorkshire under a £20bn deal. The plan includes up to 100MW of compute from the first quarter of 2028. National Grid agreed to buy 29.5 hectares at Dorman Point on Teesside, the Guardian reported on 1 November. The Guardian said the site could reach 3.2GW of capacity.

Water is the other front. Data Center Knowledge reported on 23 October that evaporative cooling can use roughly 20,000 to 50,000 gallons a day per MW of IT load. Matt Green, chief executive of HVAC firm Brucker, told the publication an open-loop system can cost 60% to 70% less upfront than a closed-loop alternative. Brendan Mark of Capture H2O argued water-cooled systems are about 25% to 40% more efficient than closed-loop air.

Two smaller items point the same way. CATL signed a memorandum of understanding with Germany's Intilion on 28 October to deploy 3 GWh of sodium-ion storage in Europe over five years, CnEVPost reported. Tutor Intelligence said on 2 November it had raised $75 million in Series B funding, The Robot Report reported.