Kathy Gibson reports – The sustainable and reliable supply of power is a key concern as rapidly-growing data centres consume more and more electricity.
The last 20 years have seen a massive evolution in the data centre, , which has created new challenges for power and cooling, says Giovanni Zanei, vice-president: large power conversion at Vertiv.
And this is accelerating as artificial intelligence (AI) becomes ever more pervasive.
“As rack density increases, we need to bring more power to the rack, and also provide more and better cooling,” Zanei say. “So the whole solution needs to be denser and more compact.
“This all leads to new challenges and opportunities for the critical power train, from prime power supplies to power distribution, from power conversion to power management.”
The first challenge is that the power infrastructure needs to enable a wide range of different new energy sources – not just utility power but also micro-grids produced by gas turbines, fuel cells, renewables and BESS.
And the exponential growth of AI driving the deployment of ever-higher rack densities. In fact, Zanei says we could see 300kw to 500kw racks within just three years.
“As compute-intensive AI workloads grow, compact and high-capacity scalable power systems have become essential to provide reliability without compromise,” he says.
The AI load is not stable either: it is a heavy and variable – even spiky – load, which places additional pressure on the power train to deliver consistent power to the data centre.
“What is needed here is a AI-ready UPS that supports a unique load profile, to handle AI load demands without performance or component degradation,” Zanei explains. “It acts as a shield to protect the source – either utility or generator – from seeing step loads.
“And it must be designed to support large increases in rack power density by working at higher voltages.”
At the same time, it’s important to minimise the impact of this spiky load on the batteries.
“As data centre power challenges grow, it is important to save space, while reducing installation and commissioning time,” Zanei adds.
Decarbonisation is a massive issue, particularly with data centres’ power demands growing. In many countries or regions, the grid is often constrained – or even completely unavailable.
“What’s needed is a path to grid independence,” Zanei says. “We need to start by reducing energy wastage. Vertiv systems reduce data centre PUE to the minimum possible while maintaining the highest levels of availability.
“The solutions also allow for dynamic grid support to enable cost savings and revenue generation by participating in demand management and other grid support services, where the data centre’s UPS can be used to support the grid.”
Big data centre operators are still on a learning curve when it comes to the new realities of power management in the AI era.
Francesco Mangano, head of data centre engineering at Data4, explains: “This is a transition period. Like AI, we are learning how to behave. But we need to deploy the right design architectures now to cope with the changes that are coming.”
Phillip Marangella, chief marketing and product officer at EdgeConneX, says the goal posts keep moving. “The amount of capacity we need to build, to continue to support cloud and now AI that doubles and triples requirements keeps growing.
“For years we have been blowing cold air on our racks. Now we are already working with 135kw racks, and this is going to double soon; then we will be talking about 1Mw racks – which was an entire data centre not long ago.
“We face the challenge of ensuring that by the time we build our power solutions they are not obsolete. We are going to have to work together as an industry to figure this out.”
Ana Jakovcic, strategic accounts director: EMEA at Vertiv, says the challenge is compounded by the fact that not all workloads are AI; and not all data centres were designed as AI factories.
“There will be new builds where the data centre design will be from scratch. And we also need to upgrade existing sites.”
She points out that densification creates challenges with power, but also with cooling – and this creates issue with space. “We need to develop solutions that tackle all of these issues.”
Stefan Nilsson, chief commercial officer at Conapto, says the industry needs to think carefully how build out new capacity.
“There are good and bad places to be. AI factories consume a lot of power, so they need to be designed in a careful and sustainable way.
“At the same time, we will see smaller facilities for places where we cannot source enough electricity for the big data centres.
“We need to mix and match in thoughtful and sustainable way.”
Mangano agrees, and says one of the major challenges in mixing existing and future infrastructures.
“The first thing customers want is flexibility, and either a partial or full deployment of AI. So the core needs to be able to deliver a mix of infrastructures.
“Flexibility is our biggest challenge. And we need to design now for two or three years out.”
Marangella points out the Gigawatt data centres are now being deployed around the world – and operators need to ensure their investment is protected.
“Flexibility and scalability are key. We have got to make sure these data centres are not obsolete as they are built.”
The availability of power, water and cold air may drive the decision on where to deploy these big data centres. “So we many nee new markets arise where there is large-scale power available.
“But then you also need to ensure sustainability, so there are all sorts of tensions.
“What we may see is new locations being selected because of power availability. If workloads are not latency sensitive, they could be in locations that are not approximate to customers.”
Nilsson’s company recently completed a new data centre that combines AI and traditional computing.
“Remember, we still have customers that require traditional services, and this is not going to go away. The challenge was to design for both.”
One of the innovations in this data centre was building a power infrastructure that would support the electricity grid.
“This was a way for us to get a return on investment,” Nilsson explains. “It is extremely expensive to build a data centre, and we needed to find revenue stream. We had a 20Mw facility and no customers using it yet, so we could sell battery capacity back to the grid.”
As power utilities add renewable sources, the grid becomes quite unstable, he explains.
“But it needs to be stable. There is a very small window that it operates it, and outside of that you get problems. So grid owners need battery capacity to stabilise.
“As a data centre operator we can help. We have tons of battery capacity in our facility, and might not have customer loads on the full capacity. So we can sell it to the grid.
“In Sweden we are a big player in supporting the grid – and it is welcomed by the grid operator.”