Technology First Read
Power Isn’t the Only Electrical Challenge For AI Data Centers
The voracious appetite for electrical power by data centers could be sated, but in a way dripping with irony
Everyone knows that data centers are voracious consumers of electricity. In fact, the U.S. is currently scrambling to meet unprecedented levels of power demand not seen since the early days of electrification and the widespread adoption of air conditioning.
In response, utilities and businesses across the nation are considering every conceivable solution from more gas power to experimental small modular nuclear reactors to geothermal technology to even the development of the world’s first commercial nuclear fusion energy system in Virginia.
Finding power for data centers is only the beginning. The data centers themselves require significantly more electrical engineering work than any other facility, and there simply aren’t enough engineers, digital designers or contractors in the U.S. to meet current demand.
The need for new electricians alone is expected to increase 6% annually, according to a Bureau of Labor Statistics report last year, and many of those electricians will be working on the 446 new data centers planned for North America by the end of the decade.
This labor shortage is another obstacle to American ambitions to consolidate the nation's leadership in the artificial intelligence race.
The significance of this shortage cannot be overstated because of the prime position electrical systems occupy in data centers: “You can’t do anything without electrical,” one data center developer told me. “ Electrical is the biggest spend in the data center … it can be 45% of the entire spend on the project.”
As companies rush to build more power-hungry, complex data centers to meet the nation’s demand for greater computing power, artificial intelligence offers a unique solution to problems of its own creation. These buildings that house the servers, allowing artificial intelligence to “think,” need that intelligence to meet the demand for these services.
Looking for quick answers on construction and engineering topics?
Try Ask ENR, our new smart AI search tool.
Ask ENR →
Power Needs Inside the Data Center
“Electrical has never been a bigger issue,” John Diamond, the principal at Strategic Facility Advisors, tells me. Diamond has worked on power problems for years, developing everything from commercial nuclear power plants to designing, building, operating, and analyzing data centers for companies like Adobe, eBay, Google, and Equinix.
The trend of rising computer demand is best illustrated by a recent Goldman Sachs report, which highlighted the hockey-stick trajectory of global power and computing needs. In 2015, data centers provided 184 million compute instances and consumed 197 terawatt hours of electricity. By 2023, those numbers had skyrocketed to over 1 billion compute instances and 411 TwH.
As computing operations for artificial intelligence require ever more power, they also require more sophisticated electrical systems to supply it. The electricity demands for server racks and cabinets are rising from 100 kilowatts now (roughly enough electricity to power three average American homes for a day) to 300 kilowatts – and potentially 600 kilowatts in a few years, Diamond says.
These rapidly rising power demands in data centers are coupled with increased cooling infrastructure needs and a radical rethinking of the design of these incredibly complex systems. Further, if design complexity is one challenge, finding the talent to generate and execute these designs is another.
The Labor Crunch
How dire the workforce problem is for data center developers depends on who you ask — and how data centers sell their capacity. Some companies sell their services before construction, while others build first and then begin marketing their capabilities. While the former are more sensitive to time pressures than the latter, both feel the current constraints of the labor market acutely.
Experts like Sean Mulligan, who designed data center facilities for Facebook and other large technology companies, see the issue as part of a cascade of shortages confronting the industry. Given the projections for up to $1 trillion in spending on data center development in the next five years, there aren’t enough electrical contractors, equipment providers, or wire providers to build properly and correctly to meet demand, Mulligan says.
The problem has become so pronounced that some data center developers are even conscripting experts outside the field to handle some electrical work.
"Electricians are really hard to come by," one industry consultant told me. The situation in electrical systems digital design is so dire, and the required skills so high, that contractors have to resort to pulling expert electricians from their construction crews back into offices to perform digital design.
In one instance, a mechanical engineer had to jump in to handle some electrical work to get a facility commissioned. "He had to jump in to fill a gap in electrical engineering," the consultant said. "He’s having to help them make changes to the voltage… [When] they have to do rework on the spot, he has to put that hat on, and then physically he has to put that hat on," to get the job done.
An AI to Solve AI’s Problems
For many data center developers, finding ways to automate the design process and enable more prefabrication for materials to accelerate construction times is the holy grail.
"Automation can help in so many different ways," says Diamond. The prime concern for developers is getting power to the cabinets, and the biggest new trend he sees is optimizing design to eliminate unnecessary components.
While many of the largest technology firms may have extensive project engineering staff to design these facilities, the next tier might not. For firms that also serve major technology companies, having tools to optimize designs and achieve the same level of prefabrication capabilities can be a significant advantage, says Diamond.
"Even if you’ve got the dream team sitting there over at Microsoft or Google or Apple – with proprietary systems that they’ve put together or software that they’ve purchased, that’s half-a-dozen companies in the world that can do that," Diamond tells me. "Then you shift down into the top-tier [co-located datacenters], and they’re designing and provisioning for the hyperscalers. They’re not going to have 20-30 [professional engineers] on staff. Their willingness to adopt software is going to be highly desirable."
Using artificial intelligence for generative design can help cut the costs and man-hours associated with the first step for these project developers, pushing design teams over their biggest obstacle to delivering to customers… personnel.
And as Diamond says, these designs make prefabrication easier, smoothing the way for accelerated timelines across projects.
As much of the industry moves to prefabrication, accelerating designs lowers construction costs. There’s less rework, less material wasted, and accelerated speeds for construction, according to Diamond.
Meeting $1 Trillion in Demand
As demand for computing power continues to accelerate, the electrical engineering challenges in data centers pose both a crisis and an opportunity for the industry.
The irony that artificial intelligence, which requires unprecedented power and engineering resources, may ultimately solve its own infrastructure problems isn't lost on developers. By automating complex electrical designs, optimizing power distribution systems, and enabling more efficient prefabrication, AI tools are poised to multiply the productivity of scarce engineering talent.
The data center industry is racing to deploy nearly $1 trillion earmarked for construction over the next five years, and the only way to meet the demand might be through AI assistance.
It gets even more meta when you realize how many of these projects are for Meta, owner of Facebook and Instagram.
Francesco "Frio" Iorio is the co-founder and CEO of Augmenta, a tech start-up that uses artificial intelligence to assist and optimized the electrical systems design process.