'The Next Big Game': AI Is Putting Adaptive Reuse Back In The Data Center Development Playbook

Amid this decade's boom in data center development, conversions of existing buildings have become more rare than they were previously. But now, an emerging shift in demand for artificial intelligence computing may trigger more adaptive reuse projects. 

With other CRE sectors suffering from higher vacancy, the idea of converting underperforming warehouses, offices and retail properties into data centers has been frequently floated but rarely executed.

Such adaptive reuse projects may sound intuitive, but a consensus emerged across the data center sector over the past half-decade that, with few exceptions, demolishing an existing structure and building from scratch is faster, cheaper and easier than attempting a retrofit. 

Those conversion calculations appear to be changing, thanks to a surge in adoption of AI products and services that is reshaping the data center development landscape. 

Data center tenants increasingly need facilities that are smaller and closer to major population centers than the massive projects that have driven the data center building boom. And according to some industry leaders, this nascent shift in demand toward “edge” data centers will soon make adaptive reuse a more attractive option for bringing new capacity to market. 

“Adaptive reuse is not that popular of a term, but I think it's going to become more and more pertinent in the discussion as the edge market evolves,” Kirk Mettam, senior vice president at engineering firm TYLin, said at Bisnow’s DICE National event in Virginia last month.

“We see a renewed interest in existing buildings.”

Before this decade's construction boom, adaptive reuse was a common approach to data center development. Industrial buildings like the Chicago Sun-Times printing facility and a Prince Macaroni factory in Massachusetts plus warehouses, telecom buildings and even office complexes were regularly repurposed as data centers.

But as demand for new data center capacity skyrocketed with the growth of the cloud industry and accelerated further with the emergence of AI, such retrofits have all but disappeared from the data center development playbook. 

Plenty of office parks, warehouse complexes and industrial properties have been redeveloped into data centers over the past five years, but developers have almost universally opted to raze the structures and build from the ground up.

The value in these brownfield sites was entirely in their access to power, fiber and other critical infrastructure. When considering a site, existing buildings were viewed less as value opportunities and more as inconveniences that needed to be removed.

“The first question I’d ask is, are you willing to scrape that site and build new?” Jonathan Mesik, a principal in the mission-critical group of design firm DLR Group, said at the event.

More often than not, converting existing buildings into the kinds of data centers driving the digital building boom is impractical, expensive and slow, industry executives said. Office, retail and residential buildings are typically unable to support the weight of the servers, cooling systems and other critical equipment used in data centers and have ceiling heights that are too low to accommodate today’s server rack designs. 

Even on properties with industrial facilities more structurally suited to become data centers, the growing size of hyperscale data center projects and Big Tech tenants' insistence on proprietary designs tailored to their infrastructure meant that reusing buildings became like squeezing a square peg into a round hole.

An industry consensus emerged that the juice was rarely worth the squeeze when it came to adaptive reuse.  

“A common piece of feedback I've gotten is that once a company does an industrial-to-data center conversion, that will be the last one they do,” JLL Head of Data Center Capital Markets Carl Beardsley told Bisnow in January. “It ends up being a little bit more of a headache,”

Yet today, an ongoing shift in AI demand is changing where and how data centers are being built in a way that could make conversions more attractive.

The data center industry is in the early days of a wave of demand for the inference computing through which users interact with AI models. While it is only a fraction of the AI computing deployed today, inference is expected to account for the majority of AI demand by the end of the decade. 

Instead of the massive, centralized clusters of computing power developed for AI training that have arisen in recent years, inference will often be deployed in smaller facilities near major metro areas or specific end users. Industry executives anticipate a “flood” of demand for “last-mile” data centers — facilities near population hubs that are measured in dozens of megawatts instead of hundreds.

Sites near major metro areas that already have the power, fiber connectivity and other infrastructure needed to build data centers quickly are often home to existing industrial or mission-critical assets. And data center executives say it will soon be far more common for retrofits of these facilities to be the best path to delivering edge capacity fast enough to meet demand.

“Inference is the next big game after large language models, so if you can find facilities near metropolitan areas that have industrial scale power, it would behoove you to capture them and retrofit them,” Sam Tabar, CEO of AI computing firm Bit Digital, told Bisnow last week after the company announced a new conversion project.

Bit Digital unveiled plans last week to deliver 24 megawatts of data center capacity through the conversion of a factory building on 96 acres in Madison, North Carolina. The firm’s leadership has touted the location as ideal for inference AI computing, near key data center hubs in Charlotte and Northern Virginia. 

Further ground-up construction is planned that will expand the site’s capacity to at least 96 megawatts, but the firm’s leadership says the reuse of the existing building will allow the first phase to be delivered by year-end.

This development strategy will soon become far more common, Tabar said. 

“There's going to be a lot of new life breathed into these old facilities from the era of manufacturing that is mostly behind us,” Tabar said. “The next era is really about [high-performance computing] and AI, and a lot of these facilities can be repurposed exactly for that.”

At the smaller scale needed for edge AI deployments, Tabar said such industrial sites often have infrastructure and equipment that can be repurposed at lower costs and get capacity online quickly. He said backup generators and fiber networks already in place at the North Carolina factory made retrofitting a better option in terms of cost and speed to market. 

“It’s not plug and play — we have to put in a heck of a lot of financial investment … but certainly there’s a litany of things we saw in there that we can repurpose,” Tabar said.  

Reuse of existing structures and critical equipment can reduce speed to market by cutting down construction timelines and avoiding supply chain constraints for items like generators that have wait times of a year or more.

In addition to manufacturing sites, data center conversions have been launched in recent months of life sciences complexes and medical buildings.

Talk of data center conversions still attracts skepticism across much of the data center sector. Even those bullish on conversions, like TYLin’s Mettam, acknowledge that such adaptive reuse projects will often not be the best option for developers, even at the edge. But he said firms that continue to reflexively reject adaptive reuse are making a mistake. 

As AI inference drives a surge of demand for edge data centers, Mettam says smart developers focused on the edge market should develop tools and criteria to evaluate whether retrofitting existing assets on a given site makes sense. Failing to do so invites missed opportunities to deliver edge data center capacity fast enough to meet the coming flood of demand.

“If you ask a developer or a builder if it’s more expensive to build ground up or to renovate, you always get the same answer. And while that’s not necessarily wrong, it’s become a predisposition,” Mettam said.  “There are many underutilized assets on the market right now … as we watch how the needs of the data center industry are evolving, we can find situations where it rightsizes.  We are starting to watch how these pieces come together.”