The design in data centres has not evolved a lot until recent years but technology has also evolved to improve reliability, cost and power efficiency.

Now, racks with one megawatt of power are being discussed, something unimaginable just a few years ago. These racks require new ways to deliver power, cool servers and maintain uptime, all while trying to stay flexible for other workloads like cloud and co-location.

At a recent panel, at Datacloud Global Congress 2025, industry experts discussed how rising demands from AI and high-performance computing (HPC) are forcing them to rethink design, reliability, power usage and even sustainability.

Speakers:

Anne Dooley, CEO of Wint Up Technologies, said the time allowed to build data centres keeps shrinking.

“The schedule [has] been shortened by the focus on paralleling offsite and onsite construction,” she said, explaining that using pre-built parts like plant rooms and generators helps speed things up.

But, AI adds new problems, she claims. “Now you see the GPUs coming into the data halls… air cooling is no longer sufficient,” she said. Facilities are starting to use a mix of air and liquid cooling, and layouts have to change to make room for more technical equipment.

Dagi Berhane, who leads Salesforce’s data centre architecture, said today’s infrastructure decisions must prepare for undefined futures.

“Architecting against the unknown should be something that we should focus on as a takeaway,” he said.

Criticising oversimplifications in industry buzzwords, he added, “I’m sure you hear it all the time- an AI data centre. There is no such thing.” Instead, AI must be seen as “a vehicle,” one of many technologies fueling compute demands.

The challenge is not just technical- it’s operational and strategic. “Imagine being hit with M&As. We are extremely notorious when it comes to things like that.”

Tate Cantrell, CTO of Verne Global, urged the industry to rethink how it collaborates with manufacturers and suppliers.

“Don’t fight the supply chain- it’s your friend,” Cantrell said. Additionally, he pointed out that today’s vendors often have better insights into chip manufacturers than the data centres themselves do, due to close integrations.

That collaboration, combined with a shift toward modular construction and repeatable reference designs, is vital.

“Trying to eliminate, identify where the constraints are. If you could eliminate most of that work and move it to an offsite location, you’re then relying on a team that day in and day out is working together,” he said. “That’s how we as an industry can become more predictable.”

Different companies are building their cooling systems and equipment, but that can slow things down.

Vedran Brzić, head of infrastructure for Vertiv in EMEA, said everyone should work together to build and test complete systems.

“It wasn’t a straightforward process, all the products in the chain were developed and tested to a specific purpose,” he said. “Once you do the design, prototype and testing, then you can be sure the overall system is going to work.”

He believes vendors need to create “a portfolio of systems or modules that are compatible with a variety of customers”.

Antonio Coccia, VP data centre solution of Riello UPS, said the industry is moving away from the old Tier III and IV reliability models, which require costly extra equipment.

“If you went down the road of the Tier III or Tier IV configuration, you would immediately have to scale up your equipment. This is a lot of money,” he reveals.

Instead, companies are using “shared-redundant” systems that save costs while still offering good uptime. He also said backup batteries, which used to be built for 15 minutes of power, now only last two to three minutes.

“Design requirements are becoming very, very challenging,” Cantrell agreed. “Extending the capability of battery storage is something that we all should be looking into,” he said.

AI servers are so expensive, sometimes up to $250 million per rack, that mistakes are no longer an option.

“We can’t afford to learn from a very expensive infrastructure,” Berhane said. “We have to test our blast radius, failure domains, even the deployment methods, in a very realistic digital twin exercise.”

He also warned against relying on outdated efficiency metrics like PUE. “Are we really still talking about PUEs in the AI market?” he asked. “We need to have realistic expectations.”

With AI demanding so much power, many companies are struggling to stay green. Cantrell said this has forced some tough decisions.

“It’s clear we’ve pulled back on our carbon objectives,” he admitted. “And we’ve done it because we have to.” He added, “We can’t just build infrastructure. We need to build credibility as an industry.”

However, according to Brzić, there are two major issues shaping data centre planning: sustainability laws and data sovereignty. “Data sovereignty is probably non-negotiable,” he said. “It’s a matter of national security at some point.”

Dooley added that power availability is now the top concern for choosing a location. “Operators are going out of the bigger urban markets, developing in Spain, Portugal, and the Nordics,” she said.

As racks grow more powerful, the whole electrical setup must change. “Today’s electrical system design will not exist anymore,” Coccia warned. Meanwhile, he explained that standard circuit breakers can’t handle the current needed for multiple high-power racks.

“With three racks at 1,000 kilowatts, we are already hitting the limit,” he said. Brzić agreed: “It’s a complete shift in the way we distribute power.”

Berhane added that while it’s possible in new buildings, retrofitting AI into older data centres is “when things become interesting.”

He also stressed that collaboration is essential. “Talk to each other- or else it’s just going to be useless”, he concluded.