A new long-haul fiber route costs between fifty thousand and one hundred fifty thousand dollars per route mile to build, and the bottleneck is not the cost. It is the permits. Right-of-way negotiations with Class 1 railroads, state departments of transportation, and utility easement holders now consume eighteen to thirty months of the typical thirty-six month build window. The trenching and splicing are the easy part. The signatures are the constraint.
This matters because the consensus underwriting framework for AI data center campuses has three inputs: power, water, and land. Fiber gets assumed. The assumption was reasonable when long-haul backbone capacity sat idle through the post-dot-com overhang and the 2010s cloud buildout absorbed inventory at a measured pace. The assumption is breaking now. Between the Northern Virginia core, the Columbus growth ring, and the emerging Memphis and Phoenix corridors, the long-haul routes that carry AI training and inference traffic between metros are running at utilization levels that would have been considered red-line ten years ago. New builds are slow. Direct hyperscaler acquisitions are quietly accelerating. And the asset most analysts treat as a commodity is becoming a scarce, route-specific input to campus valuation.
The Map As Most Analysts Draw It
The public picture of US long-haul fiber is dominated by six names. Lumen Technologies, the inheritor of the CenturyLink and Level 3 footprints, operates roughly four hundred fifty thousand fiber route miles in North America. Zayo, taken private in 2020 by EQT and Digital Colony, sits at approximately one hundred forty thousand. Crown Castle, before its tower-fiber spinoff discussions, held around eighty thousand. Uniti Group, which absorbed the Windstream wholesale assets, runs a Southeastern-weighted footprint. GTT, post-restructuring, operates a transatlantic-anchored network. The carrier incumbents, AT&T and Verizon, hold their own long-haul assets that they neither break out publicly nor systematically lease into the wholesale market.
Route miles are the wrong unit when the question is AI training capacity between metros. A route mile of fiber in central Nebraska is not fungible with a route mile in the Loudoun-to-Columbus corridor. The published totals create an impression of abundance. The economically meaningful subset, lit and dark capacity on the routes that connect the actual AI compute clusters, is a fraction of the total and concentrated on a small number of physical paths.
The paths matter because of how AI workloads use the network. Training runs require high-bandwidth, low-latency interconnect between clusters that may be split across two or three campuses. Inference traffic requires reliable transit between metro edge sites and the regional core. Both patterns concentrate demand on the same handful of corridors: Ashburn to Columbus, Ashburn to Atlanta, Dallas to Phoenix, Dallas to Atlanta, Atlanta to Memphis, and the Northern Virginia southbound route into the Carolinas. These are the corridors where the published route mile totals collide with physical reality.
Where The Map Is Already Breaking
The Northern Virginia to Columbus corridor is the clearest case. Loudoun County remains the densest concentration of data center capacity in the world. Columbus has become the Midwest counterweight, drawing hyperscaler campuses to New Albany and Hilliard. The fiber routes between them follow a small number of physical paths, most of them along Interstate 70 or the Norfolk Southern rail right-of-way through Pittsburgh. The Tier 1 backbones on that corridor are not failing. They are tight. Wholesale dark fiber inventory on the route has compressed. Lead times for new wave services have stretched. Pricing on incremental capacity has firmed for the first time in a decade.
The Dallas to Phoenix corridor shows a similar pattern with a different cause. The build is constrained by terrain and by the limited number of viable rights-of-way through New Mexico and West Texas. Union Pacific and BNSF control the rail corridors that carry most of the existing fiber. New trenching along Interstate 10 and Interstate 20 is possible but slow. The Phoenix metro is absorbing data center capacity faster than the long-haul into it can scale.
The Atlanta to Memphis path is the third case worth naming. Memphis has emerged as a strategic AI campus location because of TVA power, river access, and a clean entitlement path through the surrounding counties. The fiber connecting Memphis into the Southeast core, however, is thinner than the campus pipeline implies. Existing routes run along the Norfolk Southern and CSX corridors. New capacity will require either expanded leases from those carriers or fresh DOT permits along Interstate 40.
A secondary effect compounds the primary constraint. The optical equipment that lights new fiber is itself supply-limited. The transition from 400G to 800G to 1.6 terabit coherent optics requires reconfigurable optical add-drop multiplexers, and the ROADM market is concentrated among Ciena, Nokia, and Infinera. Lead times for ROADM hardware have moved from twelve weeks to closer to forty in the past eighteen months. A new long-haul route is not lit the day the fiber is in the ground. It is lit when the optics arrive.
Hyperscaler Direct Fiber Acquisition
The response from the largest buyers has been to stop leasing and start owning. Meta, Microsoft, and Google have all moved to direct dark fiber acquisition along strategic corridors. The transactions are not always announced. When they are, the language is muted, structured as long-term IRU (indefeasible right of use) deals with thirty-year terms that economically resemble ownership. The pattern is consistent. The hyperscalers are buying the corridors that connect their largest campuses, and they are buying them outright.
The subsea cable layer shows the same pattern more visibly. The 2Africa cable, anchored by Meta with a consortium of carriers, lands across the African continent and connects into European hubs. Anjana, Microsoft and Telxius, runs from Virginia Beach to Spain. Bifrost, Meta and Keppel, crosses the Pacific from California to Singapore. Marea, Microsoft and Telefonica, connects Virginia Beach to Bilbao. Each of these is a direct hyperscaler infrastructure investment rather than a capacity lease from a traditional carrier. The economics are clear. At hyperscaler scale, the lease-versus-buy crossover happens early, and the strategic value of route control outweighs the capital intensity of the build.
Inland, the same logic applies. A hyperscaler operating four campuses on the Northern Virginia to Columbus corridor cannot tolerate carrier-driven scheduling on capacity additions. Owning the fiber, or holding an IRU economically equivalent to ownership, moves the constraint inside the buyer's operational control. The secondary effect, less discussed, is that hyperscaler direct acquisition removes capacity from the wholesale market. Every IRU that moves from carrier inventory to hyperscaler control reduces what is available for the rest of the market to lease.
Right-Of-Way Is The Hard Asset
The fiber itself is a commodity. Single-mode optical fiber is manufactured at scale by Corning, Prysmian, and a handful of Asian producers. The cable, conduits, and splice enclosures are commodity inputs. What is not a commodity is the right to put the fiber in the ground along a specific path.
Three categories of right-of-way dominate long-haul fiber in the United States. The first is Class 1 rail. BNSF, Union Pacific, Norfolk Southern, CSX, Canadian National, and Canadian Pacific Kansas City together control the rail corridors that carry most of the existing long-haul fiber footprint. Fiber along rail corridors benefits from continuous, cleared, graded right-of-way that is already permitted for utility installation. The railroads lease this right-of-way to fiber operators under long-term agreements, and those agreements are increasingly the binding constraint on new build.
The second category is state department of transportation right-of-way along the interstate system. State DOTs control the easements along Interstate 70, Interstate 40, Interstate 10, and the other long-haul highway corridors. Permitting is state-by-state, slow, and increasingly political as state governments recognize the strategic value of the asset they are licensing.
The third category is utility easement, held by electric transmission operators and pipeline companies. These corridors are physically less continuous than rail or highway, but they exist and are increasingly considered for fiber co-location.
The binding insight is that all three categories are scarce, are controlled by a small number of incumbent holders, and are not being created. A railroad does not lay new track to enable new fiber. A state does not build new interstates. A pipeline does not relocate. The right-of-way map is fixed. What changes is who holds the leases on that map, and at what price.
Municipal fiber, which receives significant public attention, does not solve this. Municipal networks are last-mile assets. They connect homes and businesses within a metro to local aggregation points. They do not carry inter-metro AI training traffic. The long-haul gap is not bridgeable by municipal investment.
The Western Union Parallel
The historical analog is not the dot-com fiber buildout. It is the telegraph along the rails between 1850 and 1875.
Telegraph wire was not laid as independent infrastructure. It ran along the railroad right-of-way. The rail companies owned the corridor, and the telegraph operators leased the right to string wire along it. In the early years, dozens of small telegraph operators held fragmented agreements with individual rail lines. The infrastructure was duplicative, the agreements were unstable, and the economics were poor.
Hiram Sibley, who built Western Union from a small Rochester operator into the dominant US telegraph monopoly by 1866, understood the structural insight. The constraint was not wire. The constraint was right-of-way. Sibley's strategy was systematic acquisition of small telegraph operators specifically to consolidate their rail right-of-way agreements. Once consolidated, Western Union was the only operator able to negotiate exclusively with the major rail lines. By 1866, Western Union had absorbed its two largest competitors and held effective monopoly on long-haul telegraph in the United States. The position lasted into the 1890s.
The parallel to long-haul fiber 2024 to 2028 is structural. Fiber is the wire. The Class 1 railroads, the interstate DOTs, and the utility easement holders are the right-of-way. The fiber operators, the hyperscalers, and the emerging infrastructure funds are the consolidators. The same logic that drove Sibley to buy operators for their right-of-way rather than their wire is driving Meta, Microsoft, and Google to acquire dark fiber for the corridor control rather than the strands themselves. The accumulation window in 1855 to 1866 was eleven years. The accumulation window now appears to be 2024 to 2028.
What distinguishes the current moment from the dot-com fiber buildout is that the dot-com cycle was a supply glut. Operators trenched fiber on speculation, overshot demand, and triggered a decade of price compression that left long-haul fiber valued well below replacement cost. The current cycle is the inverse. Demand from AI training and inference is outrunning supply on the corridors that matter, and the right-of-way constraint prevents the supply response that would normally clear the imbalance.
What Becomes True By 2028
Three shifts follow from the structural pattern.
First, dark fiber becomes a publicly tracked asset class. Today, dark fiber capacity on specific corridors is opaque, negotiated bilaterally, and reported in fragments by the carriers that hold it. By 2028, the combination of hyperscaler disclosure pressure, infrastructure fund LP demand for visibility, and the simple commercial reality of repeated transactions on the same corridors will push the asset class toward something resembling the public reporting standards that already exist for cell tower capacity. Route-level dark fiber availability and pricing on the top twenty US corridors will be tracked, published, and benchmarked.
Second, long-haul fiber route control becomes an explicit underwriting input for AI campus valuation. The current diligence framework treats fiber as a checkbox. By 2028, a campus thirty miles from the nearest lit long-haul corridor will be valued at a meaningful discount to a campus with on-site or adjacent fiber from two or more carriers. The discount will be priced, the methodologies will be published, and the input will sit alongside power, water, and entitlement status in the standard underwriting model.
Third, the railroad right-of-way fiber leases that were signed in the 1990s and early 2000s will be strategically renegotiated. Many of those agreements were long-dated, priced at the era's wholesale market rates, and structured before the railroads understood the strategic value of what they were licensing. As they roll, the new terms will reflect the current scarcity, and the multiples on the new agreements will set the benchmark for how the railroads value the rest of their corridor. New entrants attempting to build long-haul fiber from a blank slate will face right-of-way pricing that makes the build economics materially harder than they appear today.
The map of US long-haul fiber, drawn correctly, is a map of right-of-way. The right-of-way is fixed. The demand is not.