The Underground Layer That Decides Which AI Data Centers Get Built
In commercial real estate, the data center conversation almost always starts with power. How many megawatts can the site pull, how fast can the utility deliver it, and how close is the nearest substation. Those questions matter, but they obscure a second variable that quietly determines whether a site closes or collapses: fiber connectivity. A campus can have a gigawatt of committed power and still be functionally worthless if it cannot connect to the rest of the digital world along enough diverse, redundant routes. As one fiber operator put it on a recent episode of The Real Finds Podcast, you can build a brand new, beautiful, shiny data center and end up with an island connected to nothing.
That framing comes from Bruce Garrison, CEO of BIG Fiber, a metro dark fiber provider whose 100% underground network spans the San Francisco Bay Area, Greater Portland, and Greater Atlanta. Garrison brings more than 20 years in the fiber infrastructure industry, including leadership roles at Kansas Fiber Network and GTS Central Europe. His vantage point sits at the intersection of two trends most brokers and developers are only beginning to underwrite: the connectivity demands of hyperscale AI campuses, and the coming shift toward a far more distributed compute map. For practitioners working anywhere near data center, industrial, or land deals, the implications are worth understanding now rather than at diligence.
We're Here To Help
Why Fiber Is Built in Rings, and Why Redundancy Is the Whole Point
Garrison explains the basic unit of network design simply. A fiber ring connects an A end and a Z end with two separate paths, so traffic always has a second way to reach its destination. Larger rings add more endpoints, but the principle holds: the topology exists to deliver near-perfect uptime. Data centers commit to uptime service level agreements with their customers, and the fiber feeding them has to meet the same standard. Construction happens, road crews dig, and fiber gets cut. Rings are the redundancy that lets an operator honor an SLA despite all of it.
For hyperscale campuses, redundancy scales up sharply. According to Garrison, the largest training campuses being built today, in the 500-megawatt to 1-gigawatt range, generally require three to four diverse routes back to the interconnect sites, the downtown carrier hotels where networks meet and the public actually consumes the product. That requirement is driven largely by a handful of the biggest bandwidth buyers in the world, the hyperscalers, whose latency and uptime commitments mean the connection can never go down. A smaller user, like a gaming company linking two facilities, might be fine with two paths. The number of routes scales with how much compute is concentrated at the site.
The Fiber Itself Has Changed
Two things have shifted in the underground layer, in Garrison’s telling. The first is fiber count. Infrastructure built fifteen years ago was roughly a quarter of the count being deployed now, and Garrison argues the legacy networks simply will not solve for AI workloads at scale. Most of what BIG Fiber has built is from the last seven to eight years: higher count, purpose-built, and concentrated in dense metro routes. The second shift is the chase for power. Because machine learning and training chips draw an order of magnitude more power than traditional CPUs, campuses now have to go wherever the power is, which stretches the geography of a market well outside the dense urban core. As the metro footprint grows, fiber count climbs with it.
This is the same dynamic we have tracked in our analysis of how AI data centers and 1-gigawatt supercomputing will reshape commercial real estate, where the gravitational pull of power availability reorders where development can physically happen. Connectivity is the constraint that rides alongside it.
Why Fiber Gets Harder the Moment You Leave the City
Garrison is direct about where the difficulty concentrates. In a central business district, there are more roads, more rights of way, and more options, which makes diverse routing easier to solve, even though much of the existing legacy fiber lacks the capacity for AI-scale demand. BIG Fiber, he notes, builds an estimated 80 to 90 percent of its network on rights of way that competitors are not using, precisely because true diversity requires not sharing the same side of the same street.
The problem flips in rural and expanded-metro areas. Push 60 to 100 kilometers outside the urban core to find land and power for a gigawatt campus, and the road network thins out. There are only so many county roads, only so many paths back to the interconnect. So the site selection trade-off becomes counterintuitive: in the expanded metro, power gets easier to solve while fiber gets harder. Garrison frames the emerging hierarchy of site selection as power and acceptable distance first, then fiber optionality, route count, and expansion capability close behind. He expects fiber to move earlier in that hierarchy as inference workloads proliferate over the next three to five years.
Garrison also stresses that fiber is a deeply local business in ways data center construction is not. A data center is a box you control; you can manage the building and coordinate the generators with near-total ownership. Building fiber at scale means navigating state, county, city, and federal jurisdictions, sometimes four different jurisdictions on a single point-to-point segment. He cites BIG Fiber’s Bay Area buildout, where the original 207-route-mile network for a single customer required more than 100 permits, with timelines varying widely by municipality. Doing it on the promised schedule, he argues, takes outside-plant teams with long-standing local relationships and decades of accumulated knowledge about who built what, and where, years ago.
The 10x Endpoint Thesis
The most consequential idea in the conversation, for anyone underwriting future demand, is Garrison’s view that the map is about to get far denser. Commercial cloud was historically built in a few highly centralized US regions, about as non-distributed as infrastructure gets. He argues that the training and, increasingly, inference workloads coming next will be very distributed, and that the number of dots on the map could rise roughly tenfold.
The logic runs through latency. Training compute is latency-tolerant; asking a model a question and getting an answer a few minutes later is fine, and that work can sit in a remote, power-rich location like the campuses rising around Abilene, Texas, where cheap land, access to natural gas pipelines, and abundant power converge. Real-time applications are not latency-tolerant. Autonomous vehicles, connected manufacturing, real-time analytics, and robotic surgery all demand compute close to the user. That pulls inference back toward population centers and forces it to spread out.
What that distribution looks like physically is where the real estate implications get interesting. Garrison expects smaller points of compute and interconnect to appear in places the market does not currently think of as data centers: third-party facilities, yes, but also modular pods on hospital campuses, AI infrastructure at the base of cell towers, converted manufacturing facilities chosen for their existing power commitments, and on-premise expansions inside enterprise buildings. He goes as far as to suggest we may not call them data centers in ten years. This connects directly to the demand story we examined in why power quality matters for data centers, where adjacency to power and infrastructure reshapes value for buildings that are not themselves the data center.
From a Few Mega-Campuses to High Volume
For brokers, the structural takeaway is a likely change in deal composition. Today, Garrison observes, the real estate community is largely hunting large parcels for large compute: bigger sites, fewer of them. He expects the next decade to tilt toward volume, finding buildings with existing power and siting modular data centers that reach the market faster than the roughly two years a conventional 50-megawatt third-party build can take. One 500-megawatt marketing assignment, in his framing, becomes twenty smaller assignments replicated for inference. Transactions get smaller individually but more numerous in aggregate, and entirely new site types enter the picture.
That shift is already visible in markets like ours. Lake County’s emergence as a data center corridor, which we covered in how data centers are powering Grayslake’s industrial development, reflects exactly the dynamic Garrison describes: development-ready land with strong power and utility access absorbing demand that the saturated core can no longer hold. Illinois’s broader push into advanced compute, including the quantum and microelectronics development at the former U.S. Steel South Works site, points to the same widening of where high-density infrastructure can land.
The Community Conversation CRE Is Having Too Late
Asked what the industry is not talking about enough, Garrison points to local education. Much of the municipal pushback against data centers, he notes, centers on generator emissions or the optics of a multi-year tax abatement. What gets lost is the long-horizon math. These buildings, especially downtown interconnects, can operate for thirty or forty years, and the cumulative tax revenue over that span tends to dwarf the value of an initial abatement.
The Loudoun County, Virginia example Garrison references bears this out. Data centers there occupy roughly 4 percent of the county’s commercial parcels yet, by the county’s own accounting, generate a disproportionate share of general fund revenue and have allowed the Board of Supervisors to lower the real property tax rate every year for a decade while funding schools, the Sheriff’s Office, and Fire and Rescue. A 2026 analysis of Northern Virginia estimated that, without data center contributions, the residential property tax bill on a median-valued Loudoun home would rise by roughly $5,800 a year. Garrison’s point for the industry is one of timing: that case should be made proactively, in front of city councils and residents, before restrictions force a reactive conversation that has already turned adversarial.
What Practitioners Should Take From This
The through-line of Garrison’s argument is that connectivity is moving from an afterthought to a primary underwriting variable, and that the geography of compute is about to fragment in ways that expand the addressable market for real estate professionals who understand it. Brokers who can speak to route diversity, interconnect proximity, and the difference between latency-tolerant training sites and latency-sensitive inference sites will be equipped to market buildings that would have drawn little interest a few years ago. The closing advice Garrison offered younger listeners applies just as well to the industry adapting around him: swallow your pride, ask for help sooner, and accept that the learning never stops.
At Van Vlissingen and Co., we have tracked the data center buildout and its ripple effects across industrial, office, and land markets throughout Chicagoland and the Midwest, because the infrastructure reshaping where compute lives is reshaping where every adjacent asset class finds its value. If you are evaluating a site, underwriting power and connectivity, or weighing how the next wave of distributed compute could affect your portfolio, our team has helped clients navigate exactly these decisions since 1879.