Fiber Route Diversity and Latency for Data Center Sites

Route diversity is the property that two or more fiber paths can fail independently. Two cables in the same trench are not diverse. Neither are two cables that leave a site in opposite directions but share a bridge crossing, a railroad underpass or the same carrier central office a few miles away. The test is physical: at every point between the building and the network hub, could one backhoe, fire, flood or vehicle strike take out both paths?

For a data center, diversity is usually discussed at three levels:

• **Building entry.** Separate entrance points, ideally on opposite sides of the building or campus, into separate entrance rooms.
• **Lateral and metro path.** Separate conduit routes from the property line to the carriers’ networks, with no shared structures.
• **Regional path.** Separate long-haul routes toward different hubs, so a regional event does not isolate the site. See long-haul and subsea fiber routes.

Engineering standards for critical facilities point the same way. Los Alamos National Laboratory’s communications standard, for example, recommends loop diversity for computer centers and other critical facilities.⁸ Our broader guide to fiber connectivity for data center sites covers route types and lateral basics; this guide goes deeper on how diversity fails, how latency is calculated and what rural sites face.

Most fiber outages are mundane. The Common Ground Alliance’s 2024 DIRT report analyzed 196,977 unique damage reports to buried utilities and found damages moving in the wrong direction, with utility work, particularly water and sewer and telecommunications, behind most of the top root causes.⁴⁵ A site whose two paths share any stretch of trench or roadway shoulder is exposed to exactly this kind of event.

The more insidious failure is loss of diversity over time. Carriers groom circuits, reroute around construction and consolidate facilities, and a path that was diverse at contract signing can quietly converge later. Federal rules for 911 service providers address this directly: covered providers must certify that they audited the physical diversity of critical circuits, tagged those circuits to reduce the chance of inadvertent loss of diversity between audits, and eliminated single points of failure or explained the alternative measures they took.² Data center buyers are not covered by that rule, but it is a useful model for what to ask of a carrier.

Large events show the regional version. In September 2025, multiple subsea cables in the Red Sea, including SEA-ME-WE 4 and IMEWE, were cut, and Microsoft rerouted Azure traffic, warning of higher latency for traffic that crossed the region.⁹ The traffic kept flowing because other paths existed; the cost was latency, because the alternative paths were longer.

Latency over fiber is dominated by propagation delay. Glass has a refractive index around 1.5, so light travels at about two-thirds of its speed in a vacuum, which works out to roughly 5 µs per km of fiber in one direction.¹ Round-trip time doubles that. Optical and switching equipment add a little more, but over long distances the glass dominates.

Two practical points follow. First, use route kilometers, not straight-line distance. Fiber follows highways, railroads and utility corridors, and a path that zigzags between towns can be meaningfully longer than the map distance. Ask carriers for route-mile figures and measured latency on existing circuits. Second, the diverse path is often the longer one. A primary route that meets a latency target may have a backup that does not, which is the trade-off the Red Sea rerouting illustrated at continental scale.⁹

How much latency matters depends on the workload. AI training clusters tolerate distance from users; inference, trading and interactive applications do not. The latency requirements by workload guide sets out typical tolerances, and AI training vs. inference site needs explains why the two are being sited differently.

Brokers and site packages often lead with the number of carriers near a site. The count is a starting point, not an answer. Several points matter more:

• **On-net vs. nearby.** A carrier with fiber passing the parcel is different from one whose network is in the county. Ask which carriers can deliver service without new construction, and how far each would need to build.
• **Independent plant.** Two brands can share one underlying cable through leases or swaps. Count physically separate routes, not logos.
• **Direction.** Carriers that all run along one highway add choice but little diversity.
• **Product.** Some carriers sell only lit services; large campuses often want dark fiber. See dark fiber vs. lit services.
• **Reach.** What matters is how directly each carrier reaches the hubs, cloud regions and carrier hotels the end user cares about.

For a hyperscale or AI campus, the end user frequently builds or leases its own long-haul capacity and cares less about local retail carriers than about rights-of-way for several independent laterals. For colocation and enterprise sites, a market with multiple competing carriers on site keeps pricing and service options open. In both cases, a credible diversity design names the routes, not just the carriers.

Many of the sites with available power are rural, where existing fiber may be a single regional route along a highway. Getting to two or more diverse paths usually means building new laterals, sometimes tens of miles, to reach a second long-haul route or a different carrier hub.

At those medians, a mile of underground fiber runs on the order of $95,000 before splicing, electronics and permits, and aerial builds about $42,000, though rock, river crossings and railroad permits can push individual miles far higher. The same study notes rural builds tend to cost less per foot than urban ones.³ Aerial routes are cheaper but share poles with power lines, so they are exposed to storms and vehicle strikes; buried routes cost more and take longer but are usually preferred for a campus’s critical paths.

Public and private investment is changing the map. NTIA’s Enabling Middle Mile program awarded nearly $980 million across 40 states and territories for projects to deploy more than 12,500 miles of new fiber.¹⁰ Carriers are also building toward AI demand: Zayo announced more than 5,000 new long-haul route miles in January 2025 and, by July, had begun building routes including a 521-mile Chicago to Minneapolis dark fiber path and a 385-mile Chicago to Columbus route.⁶⁷ A rural site near a planned route can look very different in three years, but plans are not plant; underwrite what exists or is contracted.

New laterals also need rights-of-way. Public road rights-of-way, railroad crossings and private easements each have their own permitting path, and a lateral across neighboring land may need easements from several owners. See mineral rights and easements.

A diversity claim should be checked with documents, not assurances. The sample screen below shows how a site can pass on carrier count and still have problems.

1. 01Request route maps (KMZ or GIS) for each proposed path from the building to the carrier hub, including entry points.
2. 02Overlay the routes and mark every shared structure: trench, conduit bank, bridge, underpass, pole line and central office.
3. 03Ask each carrier whether its route uses its own cable or leased strands, and from whom.
4. 04Get measured round-trip latency on comparable circuits, not only calculated figures.
5. 05Price and schedule any new laterals, including rights-of-way and railroad or highway permits.
6. 06Write diversity into contracts: named routes, notice of reroutes and periodic route audits, borrowing the logic of the 911 rules.²

Fiber rarely decides a site alone, but a convergence point found late can be expensive to fix. BlackForge screens fiber routes alongside power and land; to have a parcel’s connectivity checked early, get a site reviewed.