The Invisible Dependency in Your Enterprise AI Plan: Fiber Infrastructure Between Your Facility and the Cloud

The paper’s core argument is simple: the five constraints on AI infrastructure growth are power, land, water, advanced chips, and fibre-optic interconnection between data centres — and the first four receive most of the attention. Fibre interconnection, the fifth, is described as the forgotten backbone of the entire system. Without the fibre routes that link hyperscale training and cloud compute facilities to each other and to users, the other four constraints become irrelevant. A data centre with available power, suitable land, adequate water and sufficient chips is worthless if it cannot exchange data with the rest of the AI infrastructure at the speed and volume that AI workloads require.

The scale of what needs to be built

The FBA white paper quantifies the interconnection gap with more precision than most treatments of AI infrastructure deployment. RVA’s baseline estimate puts current US long-haul fibre capacity at approximately 95,000 unique route miles and 159 million fibre miles. The forecast requirement by 2029 — to support a projected 3x increase in hyperscale data centre capacity and the addition of 573 new hyperscale facilities over the five-year period — is 187,000 route miles and 373 million fibre miles: a near doubling of route miles and a 2.3x increase in fibre miles within four years.

That growth breaks down across three distinct infrastructure categories. Approximately 109 million new fibre miles are needed to upgrade capacity on existing long-haul routes, primarily by replacing older, lower-density cables with newer cables carrying 864 to 1,728 strands per cable — and, in some cases, cables already announced with 3,456, 6,912 or even 16,000 strands. A further 66 million fibre miles are required for lateral connections from new data centre sites to existing Internet backbone access points, at an average of 135 route miles per new facility. The remaining 38 million fibre miles must come from entirely new long-haul routes where existing rights-of-way are either insufficient or unavailable.

The annual installation rate required to hit this target roughly doubles between 2025 and 2029: from around 30 million new fibre miles per year at the start of the period to nearly 60 million by its close. For context, the white paper notes that even the first US hyperscale data centre — Google‘s facility in The Dalles, Oregon, opened in 2006 — was sited specifically because pre-existing long-haul fibre backbone ran through the Columbia River corridor. Fibre availability has been a site-selection factor since the beginning of the data centre era; what has changed is the scale and urgency of the need.

Capacity, latency, and the edge inference threshold

The paper’s technical section on interconnection requirements has more direct relevance to enterprise AI planning than its framing as a telecom infrastructure story might suggest. Three performance parameters determine whether a given fibre connection adequately serves AI workloads: capacity, latency, and security/uptime. Each maps onto a different aspect of how industrial enterprises access AI compute.

On capacity, dense wavelength division multiplexing (DWDM) is the primary tool for increasing the data throughput of existing fibre. Long-haul DWDM capacity has already been upgraded four to six times in recent years, from 100 gigabits per second to 400–600 gigabits per second per wavelength, with some providers now planning systems capable of 1.2 terabits per second on shorter runs. The white paper notes that even DWDM has physical limits — Shannon’s Law ultimately defines the maximum information density achievable over any channel — meaning the strand-count upgrades described above are not optional as the throughput ceiling of existing fibre approaches.

On latency, the white paper describes hyperscale operators’ preference for at least one direct, low-latency connection from each data centre site with approximately 5 milliseconds of latency for primary backup operations. Secondary paths typically run at 7 to 10 milliseconds. These numbers matter for enterprise AI planning because 5 milliseconds defines a rough connectivity threshold for cloud-based AI inference in latency-sensitive applications. Facilities within roughly 500 to 600 kilometres of a well-connected data centre can reach that threshold on a direct route. Facilities further away, or connected via indirect or heavily switched paths, face latency that may push certain real-time AI use cases toward edge inference rather than cloud inference — a distinction with direct capital and operational cost implications. Hollow-core fibre, which routes light through an air-filled core rather than glass, is being developed to reduce latency by 20 to 30% on long-distance routes, though commercial deployment at scale remains some years away.

On security and uptime, major hyperscale operators now require three to four physically redundant fibre routes from each site, sourced from at least two different providers where possible, with routes hardened through additional undergrounding, steel or concrete encasement, and increasingly through continuous fibre sensing — optical cable deployed as a distributed sensor capable of detecting digging activity or seismic events along the route. For enterprise buyers assessing the resilience of their connectivity to cloud AI infrastructure, these redundancy standards offer a useful benchmark against which to evaluate the architecture their internet service providers or managed network partners actually deploy.

What this means for enterprise AI planning in industrial environments

For enterprises running or planning private 5G networks in manufacturing, mining, ports, airports or utilities, the FBA analysis surfaces an infrastructure dependency that rarely appears in AI deployment roadmaps: the quality of the fibre backhaul between a facility and the nearest well-connected data centre.

Most enterprise AI planning focuses on the AI application layer — which use cases to prioritise, which models to deploy, which operational processes to automate — and on the private network layer that connects devices and machines within the facility. The connectivity layer between the facility and the cloud compute running AI training or inference is typically treated as a commodity input, procured from an ISP or carrier and assumed to be adequate. The FBA paper’s argument is that this layer is neither a commodity nor guaranteed to be adequate, particularly for facilities in locations that are not currently well-served by long-haul fibre and are unlikely to be connected to new routes in the next four years.

This has practical implications for how industrial AI deployments get structured. Facilities with poor connectivity to major data centre clusters face a genuine choice between accepting higher latency for cloud-dependent AI applications, investing in edge compute infrastructure to run inference locally, or factoring connectivity improvement into the total cost of the AI deployment. None of these is a simple or inexpensive option, and all three should be evaluated explicitly rather than deferred to the implementation phase after vendor contracts are signed.

The dark fibre preference described in the white paper — where hyperscale operators prefer to lease or own the fibre capacity they use for critical interconnection rather than relying on managed services from carriers — points toward a procurement principle that larger industrial operators could reasonably apply to their own backhaul strategy. Where AI compute dependency is high and latency is a genuine operational constraint, treating backhaul as a managed service procured from whoever offers the lowest per-megabit price is a risk that may only become visible when a time-critical AI application fails to perform.

The policy dependency and its commercial implications

The white paper’s recommendations section addresses a structural problem that individual enterprise buyers cannot solve but should track: permitting, right-of-way access, and construction capacity are the binding constraints on whether the 2X route-mile target gets hit by 2029. New long-haul routes require continuous rights-of-way across land parcels that may involve dozens of separate easement negotiations, plus federal and state permitting processes that can take years. The FBA submitted a filing to the National Science Foundation in March 2025 requesting permitting relief, more favourable capital gains treatment for infrastructure investment, and expanded access to public land for fibre deployment.

Whether those policy requests translate into regulatory change on a timeline that matches the 2029 deployment target is uncertain. The commercial implication for enterprise buyers is straightforward: AI infrastructure connectivity is going to be distributed unequally across geographies for the foreseeable future, and the difference between a well-connected and a poorly-connected site will become more consequential as AI applications move from analytical to operational. Including connectivity quality in the site assessment and infrastructure planning process for any new industrial AI deployment is not premature caution — it is a due diligence step that the FBA’s numbers make difficult to justify skipping.