AI Is Changing A Key Data Center Site Consideration, But Not In The Way The Industry Expected

Artificial intelligence is making network connectivity a far more complicated — and far more critical — part of the data center site selection equation. 

One of the key considerations for a data center site’s viability is its network connectivity — proximity to key fiber optic cables and other telecommunications infrastructure that make up the information highways along which data travels around the world.

While AI was expected to make connectivity less important for many large-scale data center projects, industry leaders today say it matters more than ever.

Thanks to widespread corporate AI adoption, the fiber requirements for data center sites have become far more substantial and complicated. AI computing has multiplied the volume of data that needs to be transported and has forced the largest data center users to rethink how their networks are designed.

With so much traffic on these data highways, existing network infrastructure is often inadequate. 

Amid these emerging fiber infrastructure constraints, hyperscale users are rejecting sites because of developers’ failure to anticipate their network needs and deliver the required infrastructure quickly.

While access to power remains the largest source of predevelopment risk, industry leaders like ValorC3 CEO Jim Buie say connectivity can no longer be an afterthought. Even if power has been secured, developers need to have answers in hand for end users’ bandwidth and latency problems.

“Connectivity has become as important in site planning for businesses like ours as power,” Buie said. “There are a lot of businesses who have gone into hot markets and just gobbled up land and power without real regard to the fiber need, and they may find that they can't really turn that land and power over to a buyer because of the lack of fiber capacity. It's a real issue.”

When Big Tech’s AI arms race first gained momentum more than two years ago, the widespread belief across the industry was that connectivity would become less important in determining where data centers would be built. Tech giants were scrambling for capacity to develop new AI models, but that training didn't require lightning-fast connectivity with end users or other facilities. Demand was expected to surge in locations with cheap power that lacked the robust network infrastructure data centers traditionally required.

However, while some demand for latency-agnostic data centers did materialize, the AI boom’s overall impact has been to make connectivity a more important factor in where data centers are built. 

AI is dramatically increasing the network infrastructure requirements demanded by cloud providers, social media giants and other major data center users. As these companies aggressively pursue AI models, products and services that require them to move far more data at faster speeds and to more places than ever before, their connectivity needs are surpassing what existing telecommunications infrastructure can support.  

“The legacy telecommunications networks were not built for this,” said Patton Lochridge, chief commercial officer at fiber provider BIG Fiber. “They're not going to solve the new requirements brought to us by AI and the proliferation of high-density data centers and everything else we're seeing today. They don't go to the right places, and there's not enough capacity.” 

Bandwidth shortages and latency, once secondary concerns, are rapidly becoming major constraints for data center end users. According to Flexential's survey of major data center tenants published this month, bandwidth shortages affected 59% of respondents, up from 43% a year ago. The share of respondents reporting excessive latency rose from 32% to 53%. 

Network traffic was already growing before Big Tech’s AI pivot, but the technology is causing the amount of data flowing in and out of data centers to skyrocket. As more powerful chips and higher rack densities create petabytes of data, demand for data center interconnect bandwidth is expected to increase 500% over the next five years, according to fiber provider Ciena.

AI is also changing the routes along which end users need this unprecedented flood of data to travel. The rapid corporate adoption of AI means that individual data center sites must serve a growing number of training and inference use cases, requiring more low-latency connections to population hubs and other data centers.

This has been exacerbated by the power constraints facing the industry. The need for power has pushed large-scale development outside of dense data center hubs and into areas where the network infrastructure is less robust. 

Additionally, power limitations have forced many users to train AI models across multiple data center sites instead of in a single cluster, requiring new high-speed fiber to be built between those facilities. For more traditional data center workloads, data traveling between these facilities would likely have gone through existing telecom networks. But these “distributed training architectures” require direct fiber routes that allow multiple sites to effectively act as a single data center.  

“AI workloads are reshaping the entire data center landscape, from infrastructure builds to bandwidth demand,” Ciena International Chief Technology Officer Jürgen Hatheier said. “Historically, network traffic has grown at a rate of 20-30% per year. AI is set to accelerate this growth significantly, meaning operators are rethinking their architectures and planning for how they can meet this demand sustainably.”

As with the power shortages facing the industry, the largest data center end users are now looking to developers to proactively address the network infrastructure shortcomings stalling their AI rollouts.

Just as hyperscalers or other major tenants will only consider sites with a clear path to power, Oracle Senior Principal Data Center Engineer Luke Saladyga said that a site without a clear path to fiber is increasingly considered a dealbreaker. 

“As we're evaluating as the end user, one of the biggest reasons that we'll reject things is connectivity — fiber to the site,” Saladyga said at Bisnow’s DICE: Southeast event last month. “It’s either there’s not enough or because of the time frame.”

This is a new phenomenon, said Brian Gero, Flexential’s vice president of core and strategic engineering.

Until now, if a developer could get enough power to a site with proximity to an existing fiber backbone, it was generally assumed that any new connectivity infrastructure could be built by the tenant or third-party fiber providers within the project timeline. There would have been little concern that fiber would be the biggest problem when it came to delivering the project on time. 

“The thinking used to be that the fiber piece would figure itself out,” Gero said. 

But network infrastructure can no longer be an afterthought during site selection or in the earliest stages of predevelopment for data center firms looking to compete for AI deployments, BIG Fiber’s Lochridge said. 

Connectivity is now a substantial area of predevelopment risk. He says developers need to get a detailed understanding of the additional fiber needed to support potential end users’ AI growth strategies, and they need to engage with fiber providers just as they would with a power utility.

“Stop taking fiber for granted,” Lochridge said. “If you're in site selection — if you're a developer or if you're a data center operator — you're going to need to start the process to solve for fiber sooner than you ever had, you're going to need more of it than you thought, and you're going to need to do more diligence around what exists and what's going to be required than you'd ever have before.

 “People think their tenant will figure that out, but that's too late. This is a big thing that has to be addressed immediately.”

UPDATE, MAY 13, 11:30 A.M. ET: This story has been updated with the newly rebranded company name of BIG Fiber.