For years, fiber was one of the more forgiving elements of data center site selection. Developers could secure land, line up power, begin planning the facility and then work with carriers to establish the connectivity required by tenants. In a traditional multi-tenant data center, that model generally worked.
At AI scale, Scott Bergs says it increasingly does not.
“The architecture of those original communications service provider networks just don’t meet the latency and/or capacity needs” of today’s high-density compute environments, said Bergs, CEO of Kirkwood Infrastructure Group, during a recent episode of the Data Center Frontier Show.
The result is a significant change in the data center development stack: network infrastructure can no longer be treated as something that gets solved after the site is chosen. For hyperscalers and neo-cloud providers, fiber route diversity, latency, physical security and future capacity increasingly need to enter the conversation alongside power and land.
And as data center campuses follow available power farther from established digital infrastructure hubs, the scale of the network challenge is expanding with them.
A connection between data center campuses that might once have extended two or 30 miles can now stretch 250 miles or more, Bergs said. What would traditionally have been considered a long-haul fiber route is increasingly becoming another piece of inter-campus infrastructure.
That change is helping drive Kirkwood’s own expansion.
From DF&I to Kirkwood
Bergs previously led DF&I, a dark-fiber infrastructure platform concentrated in Northern Virginia and Maryland. Kirkwood Infrastructure Group is not simply DF&I under a new name, he said. Rather, it represents what Bergs described as a second phase in a broader infrastructure investment strategy developed originally through IPI Partners.
IPI, an investment platform focused on digital infrastructure, backed DF&I after identifying communications infrastructure serving dense compute environments as an area requiring greater direct investment. Its portfolio has also included companies such as STACK Infrastructure, RadiusDC and other digital infrastructure businesses.
IPI was subsequently acquired by Blue Owl Capital. According to Bergs, the investment funds and portfolio companies remained intact, while the acquisition created access to a broader capital base.
Kirkwood emerged as the team and operating platform were separated from the underlying network assets, creating more flexibility as individual markets and investment funds mature.
The company is now working across Florida, Georgia, Mississippi, Alabama and Louisiana while evaluating projects in Kansas and Missouri.
The basic mission has remained consistent: solve connectivity problems for large, high-density compute environments — preferably before those problems materialize.
That “before” is becoming increasingly important.
Fiber Can No Longer Come Later
Bergs contrasted the current hyperscale model with his experience developing traditional multi-tenant facilities.
For a roughly 5 MW data center, a developer could build the facility and then create an ecosystem of communications providers around it. Tenants could select carriers and services appropriate for their workloads.
Hyperscale and neo-cloud campuses present a much different problem.
Existing carrier networks may not have been engineered for the capacity and latency requirements associated with dense AI compute. Even when suitable capacity exists nearby, physical route diversity requirements can force additional construction.
Bergs said high-density facilities typically require four diverse network paths.
That means an area with ample capacity on one or two existing routes may still require two or three new physical builds before a campus meets its connectivity requirements.
And waiting until late in the development cycle to discover that problem can create much more than a scheduling inconvenience.
Large parcels can be physically landlocked. Developers may require private easements. Public rights-of-way may not reach the property in useful directions. Natural barriers can eliminate potential routes altogether.
Meanwhile, construction of the data center itself can make fiber deployment harder.
Power infrastructure and road improvements frequently disturb public rights-of-way. That work can trigger construction moratoriums that temporarily prevent fiber providers from entering the same corridors.
“If you do that late in the game, the way it had been done historically,” Bergs said, developers can extend the project schedule — or find that a necessary route is simply impossible to build.
His conclusion is straightforward: connectivity planning increasingly needs to begin before the ultimate tenant is even known.
That allows fiber construction to be coordinated with power, transportation and other infrastructure while protecting the campus’s target ready-for-service date.
Power Is Changing the Network Map
The problem becomes more consequential as AI data centers spread beyond traditional hubs.
Power availability remains one of the dominant forces shaping data center geography. Increasingly, however, developers are examining sites not only for available utility capacity but also for their ability to support behind-the-meter generation.
That means evaluating proximity to natural gas pipelines, whether those pipelines have sufficient available capacity, and whether local zoning permits onsite generation.
There is another layer to the analysis: whether the equipment required to build that generation system can actually arrive when the developer expects it.
“Have they been ordering turbines?” Bergs said. “Have they been able to get some of those necessary components?”
A site with theoretical access to natural gas is not necessarily a site with executable power. And a location with an executable power strategy is not necessarily well connected.
That is creating a new class of potential data center markets: locations with the power and land required for large-scale development but considerably less communications infrastructure than established hubs such as Northern Virginia.
For network developers, following compute into those areas means extending infrastructure farther and planning it earlier.
From Fiber Strands to Dedicated Conduit
The changing physical requirements become even clearer at the campus level.
Bergs has spent more than three decades estimating network capacity requirements. His record, he joked, has been remarkably consistent.
“I’ve been wrong every time and I’ve never been high.” The progression in network consumption helps explain why.
Enterprise and carrier networks initially relied heavily on lit telecommunications services. As bandwidth requirements increased, some users moved toward dark fiber. High-density compute has pushed the requirement further, from individual fiber strands toward dedicated cables and now, increasingly, dedicated conduit.
Bergs said Kirkwood is seeing demand — and in some cases mandates — for conduit capacity rather than simply a specified number of fiber strands. The reasons extend beyond raw bandwidth.
Dedicated conduit can provide hyperscalers with greater physical-layer security. It allows the operator to control routing and therefore latency. And it creates room to expand the network without disturbing traffic already running over existing cables.
“If I’m a hyperscaler that’s going to be driving traffic from my training site out to various different inference nodes, I want to have the security,” Bergs said. “I also want to have the latency control.”
Fiber counts inside those routes are increasing rapidly as well. Bergs said Kirkwood is seeing conventional single-mode fiber cables ranging from a minimum of roughly 864 fibers to as many as 6,912, depending on the application.
Those numbers make the underlying conduit system a strategic asset in its own right.
Building for the Fiber After Next
That becomes particularly important when considering emerging technologies such as hollow-core fiber.
Instead of transmitting light primarily through glass, hollow-core fiber guides it through an air-filled core. The approach has attracted attention for its potential latency and performance advantages.
But Bergs cautioned that those benefits come with practical engineering tradeoffs.
Current hollow-core implementations have a larger physical form factor than conventional single-mode fiber. That matters when developers are attempting to maximize capacity inside existing conduit.
An inch-and-a-half conduit, Bergs noted, may not accommodate enough hollow-core fiber to create a compelling capacity advantage over placing a conventional 6,912-count cable in that same space.
Installation presents another issue. Matching and connecting the core requires a different skill set from conventional single-mode fiber splicing.
“Doing so with a hollow core is just a completely different process,” Bergs said.
He expects the industry to adapt if hollow-core achieves widespread adoption. But for the foreseeable future, Bergs believes conventional single-mode fiber still has enough available capacity to remain the predominant technology.
That makes the larger planning lesson more interesting than any prediction about which fiber technology ultimately prevails.
“We’re just trying to create enough conduit capacity so that regardless whether it’s hollow core, whether something else comes up in the interim … we’re able to accommodate it.”
In other words, the safest infrastructure strategy may be to avoid betting too heavily on the fiber of tomorrow and instead make sure the physical pathway built today can carry it.
Permitting Becomes a Shared Infrastructure Bottleneck
More conduit does not solve every problem. The sheer volume of infrastructure development surrounding AI campuses is placing pressure on the agencies that permit it.
Bergs pointed to the U.S. Army Corps of Engineers, levee districts and state and local transportation departments as examples of authorities confronting growing volumes of applications.
Fiber is only one source of that demand. Natural gas infrastructure, water systems and other utility projects can require crossings of the same waterways, roads and regulated areas.
In some regions, Bergs said, permitting authorities are simply struggling to keep up. That makes the quality of an application consequential.
Kirkwood’s approach is to engage permitting bodies early enough to understand their technical requirements and submit applications correctly the first time.
The goal is partly project efficiency. But it is also about avoiding repeated filings that consume already limited agency resources.
The fiber supply chain presents a similar capacity question.
Manufacturers have seen demand spikes before, including those associated with government-backed broadband programs. Those cycles create a difficult investment decision: build additional manufacturing capacity to address a surge that may disappear several years later, or risk shortages when demand remains strong.
Bergs sees the current cycle differently. The data-transfer requirements associated with high-density compute, he said, are appearing on planning horizons of 10 to 15 years, with little indication that bandwidth requirements will retreat.
He compared the current period less to a temporary subsidy-driven broadband surge and more to the late-1990s expansion of the commercial Internet, when increasing numbers of businesses and transactions moved online and the underlying infrastructure simply had to grow with them.
Bergs hopes that distinction will encourage additional domestic fiber manufacturing investment.
When 250 Miles Becomes Inter-Campus
The changing geography of AI infrastructure may be the clearest evidence of how different the network problem has become.
Large campuses are increasingly appearing in places where power is available but communications infrastructure is relatively sparse.
Historically, Kirkwood might have connected several sites within a market through a ring architecture and then extended those campuses back toward established long-haul networks.
Those individual inter-campus links might have covered two miles, 10 miles or 30 miles. Now they can reach 250 miles.
“What might have traditionally been thought of as a long-haul segment or path for us is just another inter-campus connectivity corridor,” Bergs said.
Training campuses need to connect with other training environments. Those systems also need access to inference infrastructure and, ultimately, the networks serving the users consuming and interacting with AI applications.
That is forcing network developers to reconsider routes they might previously have ignored.
In the past, Bergs said, the presence of an AT&T, Lumen or Zayo route with available capacity might have been enough to remove a corridor from consideration.
Not anymore. If a high-density campus requires three or four independent, high-capacity, low-latency paths, one incumbent fiber route — even a very good one — solves only part of the problem.
Existing connectivity therefore has to be evaluated not simply on whether fiber is present, but on whether the complete network architecture required by the campus can actually be delivered.
Infrastructure Still Needs a Community
The same early-planning logic increasingly extends to community acceptance.
Bergs said the industry has historically been reasonably effective at responding to questions and opposition after they arise. That approach is becoming inadequate as data center development attracts greater public scrutiny.
He pointed to Georgia, where state and business organizations are trying to bring stakeholders together earlier, identify communities interested in hosting data center development and provide factual information before opposition hardens.
Projects that might once have moved relatively smoothly through local approvals can now encounter intense community resistance, sometimes based on information that does not apply to the project being considered.
Bergs does not claim the industry has solved that problem. “I wish I had a crystal ball and a really, really perfect answer for that,” he said.
But he does see one lesson emerging.
“It’s very difficult to fight emotion with facts,” Bergs said. “But you can sometimes prevent negative emotion with positive facts if they’re presented first.”
That observation applies surprisingly well to the broader infrastructure problem.
The AI data center boom is forcing developers to think earlier about almost everything: power equipment, gas capacity, road construction, permitting, conduit, network diversity and community acceptance.
Fiber is becoming part of that front-end calculation.
AI demand may be pulling compute toward wherever power can be assembled fastest. But those campuses still have to connect — to other training campuses, to inference infrastructure and ultimately to the users and networks consuming the output.
At that scale, connectivity can no longer wait for the data center to arrive. Increasingly, it has to be part of the reason the data center can arrive at all.
