Liquid cooling is fast becoming the default for artificial intelligence (AI) data centers, as the chips inside each server rack generate more heat than air can carry away. The shift has accelerated sharply over the past year.
AI training racks draw around 70 kilowatts (kW) on average today. Coming generations of Nvidia hardware could push that to 500kW, a load that conventional fans, floors and electrical distribution were never designed for.
“Liquid cooling is the only solution for these rack densities,” Matthew Baynes, vice president of Schneider Electric’s Secure Power and Data Centre division for the UK and Ireland, told TechJournal.uk in an interview. “Mechanical cooling technology design has rapidly moved from air to liquid in the past 12 to 18 months.”
“The next evolution of chip could be up to 500 kilowatts per rack. A traditional data center just cannot cater for that type of design or type of rack load,” he said.
A rack’s kilowatt rating measures the power it draws continuously, not its hourly use. A 500kW rack running around the clock would consume about 4.4 million kilowatt-hours (kWh) a year, as much electricity as roughly 1,750 UK homes, based on the typical household benchmark of 2,500kWh a year set by the energy regulator Ofgem.
By comparison, the average enterprise server rack draws about 9kW, or roughly 79,000kWh a year if run nonstop, enough for about 32 homes. An AI rack packs dozens of graphics processing units (GPUs) wired to work as one machine. Nvidia’s current GB300 NVL72 holds 72 GPUs and draws about 140kW.
Its Rubin Ultra rack, due in 2027, is designed to draw around 600kW.
A GB300 rack is no bigger than a standard server cabinet, measuring 600 millimeters (mm) wide, 1,068mm deep and 2,236mm tall. Loaded, it weighs about 1,580 kilograms (kg), and Nvidia says its Vera Rubin rack weighs about 1,800kg.
Clusters of such racks train large AI models on vast datasets. The same hardware also runs finished models to answer users’ requests, a task known as inference. In an industry benchmark, a single GB300 rack generated about 674,000 tokens, or word fragments, per second on DeepSeek’s R1 model.
Baynes said Nvidia’s move from GB300 to Vera Rubin pushes densities to 200kW and 250kW, with the generation after that possibly reaching 500kW.
“Their pace of research and development (R&D) change is nothing like the industry’s seen before. Every six months, what we talked about in terms of rack density changes by a factor of whatever it might be,” he said.
JLL, the commercial real estate firm, expects average rack density to triple to 45kW by 2030, with 80% of new facilities adopting liquid cooling.
“You may have seen Schneider with a strategic acquisition of Motivair, an American liquid cooling company, to position ourselves strongly in the world of liquid cooling,” Baynes said.
Schneider agreed in October 2024 to buy a 75% stake in Motivair for $850 million in cash and plans to acquire the remaining 25% in 2028. Founded in 1988 in Buffalo, New York, Motivair makes coolant distribution units (CDUs), rear door heat exchangers, cold plates and chillers.
On September 23, Schneider launched the WCDU, the largest CDU in the Motivair range. A single unit can cool up to 3.5 megawatts (MW).
Up to 20 units can run together to serve data halls of 30MW to 40MW. The WCDU sits in the technical corridor outside the server floor and combines liquid cooling with fan-wall air cooling, so operators can shift the balance between the two without redesigning the hall. It ships in select regions in October.
Months to weeks
Baynes took the UK and Ireland role in January after more than 20 years at Schneider.
He said a newer group of customers, AI labs and neoclouds that rent out GPUs as a cloud service, wants speed above all. A traditional data center takes 18 to 24 months from design to ready for service, but these operators want to be running within six to 12 weeks.
“They are challenging the industry to move data center builds from months to weeks,” he said. “They’ve got [GPUs] on order with Nvidia. They need to deploy them and convert that into revenue as soon as possible. These GPUs are expensive. You need deep pockets.”
He said a lot of Schneider's recent demand has come from AI labs and neoclouds in Europe and beyond, and the company is in discussions with them in the UK and Ireland.
Britain’s problem is that most of this AI training work is heading elsewhere.
“The cost of energy in the UK is around four times the cost in the Nordics, which means the UK struggles to attract that type of IT (information technology) load,” Baynes said. “We’ve struggled to attract the large-scale AI training and AI lab load market in the UK because, firstly, we are challenged with having access to energy, but of course our cost is prohibitive too.”
He said AI training campuses range from 50MW to one gigawatt (GW), and much of that training need not stay within national borders. It therefore follows cheap power to:
The Nordics, with abundant hydroelectricity
Iberia, with solar power
France, with nuclear capacity and national AI ambitions
“In the UK, we have the AI Growth Zones, and there’s an element of power capacity there, and a slight incentive on cost of energy. It still doesn’t get close to the cost of the Nordics or Iberia,” he said.
He said UK power costs about three times Iberia’s, and discounts of 10% to 15% in the growth zones still do not make the economics of training work.
Queuing for the grid
Access to power is the other obstacle. Baynes said a new UK grid application now faces a wait of seven to 10-plus years, and press reports suggest some new connections will not be available until 2040.
He said established operators that have filed requests year after year will see connections arrive steadily, while newcomers wait longest. Reform from Prime Minister Andy Burnham’s government, which took office in July, could speed up the queue.
“If you’re stepping in now as a new logo and requesting a connection outside of the [AI Growth Zones], it could be quite lengthy. That’s one of the biggest challenges to hyper-scale data center growth in the UK, but steady growth will be maintained,” he said.
As AI shifts toward inference, Baynes expects that steady growth to come from enterprise and government workloads that need to stay in the country.
“When AI becomes more enterprise and more sovereign, we’ll see more demand come into the UK, and we are seeing interest in neoclouds landing in the UK and building some sites in the UK,” he said. “The cost of energy would not be number one in that regard.”
JLL expects inference to overtake training as the dominant AI workload in 2027, and AI to account for half of all data center workloads by 2030.
Operators are also looking for ways to get around the grid altogether.
Baynes pointed to Era4, a UK developer that uses waste to generate power, though he said such partnerships are not available to everyone. Era4 operates 44 sites and is building AI data centers on landfill, including one in Canterbury linked by private wire to a landfill-gas plant.
He also cited on-site microgrids such as Pure DC’s 110MW system in Dublin. He said microgrids can cut energy costs compared with the national grid, though they require upfront capital spending that can be restructured as an operating cost. He called them a viable option.
Small modular reactors (SMRs), such as those Rolls-Royce is developing, are further off, he said.
“In the UK, we’re a nuclear country, so there’s a possibility for us to do it. But there are quite a few hurdles the industry needs to go over before maybe we’ll see a live SMR in the UK,” he said.
Schneider ships the WCDU in October, but its UK customers still need the new government to shorten the power queue.