‘A Hurricane's Coming’: The AI Boom Is Making It Harder To Build Data Centers Quickly

The rapid pace of innovation in artificial intelligence is making it harder for developers and contractors to build new data centers as fast as their Big Tech customers demand. 

In a data center industry where speed to market is critically important amid unprecedented demand, builders were already struggling with yearslong supply chain and labor constraints that made costly delays difficult to avoid. Now, AI is making the situation much worse. 

With the computing infrastructure supporting AI evolving at a breakneck pace and tenants constantly shifting their AI adoption road maps, fundamental elements of a data center’s design are changing during development, throwing carefully orchestrated logistics into chaos.

Several developers and construction professionals at the forefront of the digital building boom, speaking this month at Bisnow’s National DICE Construction, Design and Development - Central event at The Westin Galleria Dallas, said these constant changes are forcing major adjustments in how they approach the construction process to deliver new data centers on time.

“What’s changed over the last 12 months is that with the implementation of AI, the design changes almost daily. You’ll get drawings, and the next thing you know, the client will say they want to utilize a different method of delivery for power,” said Tyler Stevenson, vice president and project executive at Clune Construction. “That has a real impact on schedule and price.” 

As an AI arms race has the world’s largest tech companies scrambling to snap up new data center inventory as fast as it can be built, speed to market has never been more important for data center builders. Contracts with major tenants carry severe financial penalties for missing delivery deadlines, and that is on top of the revenue providers lose with each day a project is delayed.  

Yet for the past four years, supply chain delays for critical equipment and an industrywide labor shortage have made it difficult for developers and contractors to adhere to construction timelines. 

The global supply chain chaos that struck many industries at the height of the pandemic has persisted for the data center sector. Wait times for equipment like generators and transformers are routinely measured in years, not weeks. At the same time, the skilled, experienced labor needed to install the specialty equipment prevalent in data centers has been in ever-shorter supply, as the digital infrastructure boom coincided with a wave of retirements across the data center workforce. 

These problems made the planning required to deliver new data centers on time vastly more precarious than what the industry had experienced previously, executives said at the DICE event. Advanced planning became critical. Equipment has to be stockpiled and ordered years in advance, with labor locked into contracts on specific schedules to coincide with equipment availability long before projects get underway.

Data center designs became increasingly standardized and modular to limit the time skilled workers had to be on site. Data center construction became a meticulously choreographed dance to avoid costly delays. 

“Speed is what everybody’s pushing for, but every headwind out there pushes against speed,” said Cory Plunk, director of real estate operations for AT&T. “Lead times, manpower for contractors, even permitting issues — all that makes speed hard, and the way you have to compensate for that is with proper planning on the front end.” 

But now, the emergence of AI has made this kind of planning vastly more complicated.

The high-performance computing equipment powering AI requires data centers that are designed differently from those hosting traditional information technology infrastructure. Powerful graphics processing units produced by firms like Nvidia use far more power per square foot than standard servers and often need to be cooled using pumped refrigerant instead of air, while other design elements ranging from network infrastructure to the structural weight limits of data halls differ from traditional data centers. 

The rapid yet uncertain pace of AI adoption means tenants are often telling developers to build for traditional servers at the start of a project, but they may decide midconstruction that some or all of that capacity needs to support AI, DICE panelists said. Developers are forced to adapt to a tenant’s long-term AI plans, even if those change during a project. 

“The way the industry is going, every quarter you need to be ready for some sort of pivot,” AT&T Technology Director Artecia Wilson said.

The designs of the AI computing infrastructure itself are evolving at a breakneck pace, complicating things further.

Whether it is Nvidia or other chipmakers unveiling new, more powerful processors or the development of new network technology or cooling systems, data center designers say they now find themselves routinely going back to the drawing board well into the development process to accommodate technology that may not have existed when the project got underway. 

For data center builders already struggling to deliver projects on time, these constant changes and uncertainty have made the planning and preparation needed for speed to market far more challenging. This limits their ability to do advanced procurement, utilize standardized designs and execute other speed-to-market strategies to navigate supply chain and labor constraints. 

The flexibility that constant AI-driven design changes demand of data center builders stands in direct conflict with the standardization and advanced planning needed to get to market quickly, DICE panelists said. And these challenges are only expected to become more pronounced as AI demand accelerates in the months ahead. 

“A hurricane’s coming, and it’s AI,” Shockey Precast Vice President Marshall Sorenson said. “We’re going to have to get projects from Point A to Point B when almost nothing is figured out, which is a huge challenge. How are we going to avoid things like supply chain constraints that we’re all afraid of?” 

Navigating this uncertainty requires unprecedented collaboration between all parties involved in a data center project from the moment it gets underway, DICE panelists said. Such multilateral early engagement with contractors, tenants, equipment manufacturers and other stakeholders allows developers to better anticipate design changes, identify areas where these changes are likely, and develop a construction plan that accounts for different scenarios. 

It is a shift that Microsoft Senior Program Manager Hani Noshi said the industry’s largest players are employing effectively.

“We're getting the general contractor, architectural and engineering firms, and the trade partners all together really early on to help accelerate time to market,” Noshi said.

“Having them engaged hand in hand in the design process, walking the path of construction and understanding what are the optimal scenarios from a path of construction perspective has been absolutely huge for all of us.”

Manufacturers of the AI computing equipment driving data center design changes are leaning into this kind of collaboration with data center builders. Nvidia, by far the largest producer of GPUs for AI, keeps customers updated on design parameters and infrastructure requirements for products that may not be available for years, allowing the customers to plan their data centers accordingly.

Nvidia account executive Arjuna Pandit said the firm has established internal teams with expertise in data center design to work with customers on how to build the facilities to host its products and make design changes so they can be more easily incorporated into the customers’ existing infrastructure. He said this is an investment to ensure that data center construction isn’t a bottleneck slowing AI’s growth. 

“We have blueprints with guidance not just on how to design your compute but how to design your data center, and we're starting to bring in personnel in our own organization from the data center space who understand cooling and heating,” Pandit said. “We're trying to be very collaborative to our ecosystem partners because ultimately, they're the ones driving AI out in the field.”