The Second Inventory

The success is not in dispute

It is worth starting with what went right, because the argument that follows is easy to misread as an indictment and is not one.

Over roughly two decades, the industry put fiber and high-capacity access infrastructure into places that had nothing comparable before. Coverage expanded, delivered speeds rose by orders of magnitude, capital arrived in volume, and the engineering and construction capability required to absorb that capital was built more or less from scratch. Measured against the objectives of the build era — passings, coverage, take rate, time to market — this is one of the more successful infrastructure expansions of the period.

That capability was hard to acquire and is still real. Nothing in the following pages suggests the industry cannot build.

The question this essay takes up concerns a different capability, which is the one required after the building stops. Long-lived physical infrastructure does not become cheaper to own once it is in the ground; it becomes a different kind of liability, one whose costs are dispersed, recurring, and mostly invisible until they are not. An industry that develops extraordinary competence at deployment does not automatically develop the corresponding competence at custody. Whether broadband did is an open question, and it is the one worth asking.

Build economics and operating economics are different disciplines

Build and operate are frequently described as phases of the same business. They behave more like two businesses with different physics.

Build economics are legible. The scope is knowable in advance: a route is designed, footage is counted, materials are ordered, crews are scheduled against a plan. Variance exists but is bounded and priced. Progress is measurable in units that accumulate — miles, passings, sites — and the whole enterprise is oriented toward completion. Capital markets understand this shape well, because it resembles construction, which they have financed for a very long time.

Operating economics have almost none of those properties. The scope is not knowable in advance, because the work is generated by the condition of a plant that is deteriorating unevenly across a large geography. Volume is not forecastable in the way build volume is, since it depends on weather, third-party damage, equipment age, installation quality from years earlier, and the accumulated consequences of previous deferrals. Nothing accumulates: a year of excellent maintenance produces a network in approximately the condition it was in, which is a genuinely difficult thing to present as an achievement. And the unit of work resists standardization, because two jobs with identical descriptions can differ by an order of magnitude in cost.

The transition between these two regimes is not a milestone that gets crossed on a particular date. It happens gradually, asset by asset and market by market, and it requires different instrumentation, different vendor relationships, different labor models, and a different definition of what good performance looks like. An organization can be excellent at the first and structurally unprepared for the second without anyone noticing, because the metrics that signal the change are the ones it does not yet collect.

What the build period rewarded

The most common explanation for how the industry arrived here is that it moved too fast. That explanation is not quite right, and getting it wrong leads to the wrong remedy.

Speed mattered during the expansion, and the competitive conditions are not in dispute. New entrants appeared, incumbents moved to defend territory, and the differentiating claim migrated from availability to bandwidth to symmetrical gigabit service in a compressed period. An operator that moved slowly risked arriving at an address behind someone else, in a market where that is an expensive place to arrive.

What can be observed from that period is what the industry measured itself by. Coverage, passings, footprint and delivered bandwidth were the figures that were counted, reported, financed and compared. They were the terms in which progress was described, and they are the terms in which the era is still described now. Records, maintainability, training capacity, vendor depth and preventive regimes were not counted in the same way, and in most cases were not counted at all.

An organization tends to develop the capabilities it can see itself developing, and measurement is a large part of how it sees. That is worth offering as an explanation of the mismatch rather than as a finding about it, because the mismatch is the part that is visible. Build capability was measured closely and matured quickly. Sustaining capability was not measured in the same way. Whether anyone weighed one against the other and chose is not something the record shows, and it is not required for the rest of this essay to hold.

The consequences of a mismatch of this kind arrive late. Deployment capability is tested immediately, by whether the network gets built. Sustaining capability is tested years afterwards, by whether the network can be kept working at acceptable cost — which is a slower examination, and one that an organization can be some way into before it notices it has begun.

The pioneer penalty

One of the more specific consequences of that sequencing falls on the operators who moved earliest, and it runs contrary to the usual expectation that early movers accumulate advantage.

The first substantial wave of builders made architecture, tooling, labour and training decisions using the best information available at the time. Those decisions hardened. Engineering standards were written around them, field practices were institutionalized around the standards, and organizations were staffed and trained against the practices. All of that is what competence looks like while it is being built.

What an operator accumulates in that period is two things at once. One is plant: cable, conduit, electronics, cabinets, the physical network that appears in the asset register. The other is the operating capability assembled around that plant — the test equipment calibrated for it, the procedures written for its failure modes, the technicians who can diagnose it, the vendors who have tooled up to work on it, the spares held against it, and the institutional memory of why it is the way it is. The second set is what makes the first set serviceable. Only the first set is counted.

Both are specific to an architecture, which is easy to miss while the architecture is stable. A test set that characterizes one transmission medium may say nothing useful about another, and a technician's accumulated pattern recognition is recognition of particular patterns. This is not a criticism of the capability, which is what made the network work. It does mean that a change of architecture puts both inventories at risk at once, and that the organization can see only one of them.

The United States access network offers a documented sequence. It should be read as specific to that market, and to carriers with a large installed copper base: it does not describe cable operators, greenfield fiber builders, or markets outside the United States, whose regulatory structures and starting conditions differ substantially.

Incumbent carriers first operated a copper access network built for telephony and adapted for data. From 2004, that was extended rather than replaced: fiber was pushed out to remote terminals in the field, with the final segment to the premises still delivered over existing copper, in the architecture launched commercially as a fiber-to-the-node service in 2006. This was a substantial engineering commitment, and it brought its own equipment in the field, its own installation and maintenance practices, its own test regime and its own vendor arrangements. Some years later the direction changed again, toward fiber all the way to the premises. Sales of the copper-fed consumer product were withdrawn in 2020.

What that transition cost in physical terms is a matter of record. AT&T recorded a non-cash pre-tax charge of $2,883 million in 2017, stating that it was "a result of planned fiber deployment" and covered copper assets the company "did not plan to utilize to support network activity." A further $1,290 million was recorded in the fourth quarter of 2019, this time attributed to "the pace at which our customers have migrated to fiber, which exceeded previous forecasts." Neither charge arose because the copper stopped working. One followed an architectural decision; the other followed an estimate that turned out to be wrong.

There is no comparable figure for the capability. Nothing in any filing records the test procedures that ceased to describe the work, the training that no longer applied to the network being built, the vendor relationships organized around a method that was being retired, or the diagnostic experience that had taken a decade to accumulate and was specific to a medium now being removed. It is worth being careful about what that silence does and does not show. It is not evidence that the capability was stranded, and it would be circular to treat an absence as proof of the argument this essay is making. What can be said is narrower: the physical half of the transition was measured to the million, the capability half was not measured at all, and an organization can only weigh what its instruments report.

The effect, where it operates, is that experience does not compound into advantage in the way it does in most industries. An operator who learned the most, by building first and at scale, may end up carrying both a physical estate and an operating capability shaped around conditions that have moved. A later entrant designs against current conditions and carries neither. This is worth stating plainly because it explains a pattern that otherwise looks like mismanagement: some of the cost differential between operators reflects when they built, and what they had to become good at in order to build it.

Nothing in this requires that competitive entry forced anyone's hand. The burden described here is the burden of transforming a live operating system while continuing to run it, and it exists whether or not a competitor is applying pressure.

Five clocks on the same asset

A length of fiber in the ground is measured by several different clocks at once, and they are related without being the same. There is how long it remains physically serviceable. There is the period over which its cost is recognized in the accounts. There is how long it earns enough to justify keeping it. There is how long the architecture it belongs to remains the right one. And there is the horizon over which somebody decides whether to spend money on it this year.

The second of those is an estimate, and it is worth seeing one revised. In 2024 Cogent Communications reviewed the estimated useful lives of fiber it had acquired with the Sprint business and concluded that "the actual lives of owned fiber... were longer than the estimated useful lives used for depreciation purposes." Effective 1 July 2024 it increased those lives from an average of fourteen years to an average of forty. The fiber did not change. The estimate did, on the basis of operating experience, and the FY2024 effect was to reduce depreciation expense by $8.7 million, increase net income by $6.5 million, and increase earnings per share by $0.14.

The same company's policy states the basis for such estimates directly: useful lives "are determined based on historical usage with consideration given to technological changes and trends in the industry that could impact the asset utilization." Technological relevance is therefore already inside the accounting estimate, rather than sitting outside it as a separate consideration. Its published depreciation periods range from two to eight years for network equipment up to five to forty for system infrastructure — a spread of more than an order of magnitude within a single network, reflecting genuinely different expectations about how long different things stay useful.

Estimates move in both directions. AT&T's copper abandonment charges are the same mechanism running the other way: assets written off while still physically capable, because the architecture they served was being replaced faster than the accounting had assumed. One company found its fiber lasting longer than expected; another found its copper becoming useless sooner. Both were estimating the same difficult thing, which is how long a piece of infrastructure will go on being worth operating.

This is where the operating question sits. Expenditure that protects the long serviceable life of infrastructure — preventive work, accurate records, maintainable design, the capability to do the work at all — produces its value across a period measured in decades. The decision about whether to fund it is usually taken inside a planning cycle measured in a year or two, using instruments built for that cycle.

It is worth being exact about what a decision to defer actually does. It reduces current expenditure, and that reduction is real: cash not spent this year is genuinely not spent, which is sometimes the correct result and occasionally the only available one. Whether it reduces cost over the life of the asset is a separate question, and the answer depends on what happens next. The physical condition stays in place and continues to develop, and where that leads is genuinely open — the condition may prove stable, or it may produce more frequent intervention, or it may shorten the interval before replacement. The cash, the accounting expense, the carrying value and the physical state of the plant are four different things moving on four different schedules, and a decision that looks settled in one of them may be unresolved in the others.

None of this requires a particular ownership structure, and it is not an argument about investors. Broadband operators are financed through corporate cash flow, public equity, debt, private capital, government programmes and combinations of all of them, and the horizon problem appears under each. What varies is how visible the long clock is from wherever the decision is being made.

Operations was treated as the elastic part

As build volumes taper and attention moves to operating performance, a particular set of moves becomes available to an operator looking for margin. Maintenance cost can be compressed. Vendors can be re-tendered at lower rates. Field organizations can be leaned out. Judgment can be standardized into process and therefore staffed less expensively. How widely any of this was done across the industry is not something this essay can establish; what follows is about what happens in an organization that does it.

For a period, each of these can appear to work. Margins improve and service metrics hold. What the reporting cannot show is the source of the stability it is recording.

Resilience in these systems is rarely a designed property. It tends to be an emergent one, produced by people who know the plant: the technician who recognizes a site's failure signature, the supervisor who understands which vendor can actually perform a particular class of work, the engineer who remembers why a route was built the way it was and what will happen if it is touched. That knowledge substitutes for documentation that was never created, for records that were never reconciled, and for design margin that was never funded.

Compressing operating cost against that backdrop does not remove the resilience. It transfers the load onto the remaining people, who absorb it because absorbing it is what competent operators do. The system continues to perform, and its performance is read as confirmation that the compression was safe — which makes the next increment easier to approve.

Resilience of this kind was not designed. It was extracted, and extraction can be sustained for a considerable period before it registers anywhere.

The maintenance-economics hypothesis

There is a specific mechanism worth setting out, with a clear statement of its status: it is an operating hypothesis consistent with field experience, not an established industry-wide causal finding, and it should not be read as one.

The proposition is that the loss of experienced capability changes maintenance economics directly, through the labor content of the work rather than through the labor rate.

A technician with deep familiarity with a plant diagnoses faster. They arrive with a shorter list of probable causes, test the right things first, and recognize conditions they have seen before. A less experienced technician on the same fault performs more tests, escalates more often, returns to site more often, and more frequently requires a second resource. The hourly rate may be lower. The labor consumed per completed job is higher, and so is the elapsed time.

If that holds at scale, a maintenance budget of constant size buys a declining quantity of completed work. Backlog grows against a budget that has not been cut. Growing backlog forces triage, triage defers the non-urgent, deferred work degrades further, and degraded plant generates more faults — which consume more of the same constrained capacity. The loop closes, and it closes without any single decision that looks wrong.

Whether this operates at industry scale is an empirical question that the available public data does not obviously answer. It is offered here as a mechanism to test, and specifically as an alternative to the more common explanation that maintenance organizations have simply become less disciplined.

What "infrastructure" was taken to include

Underneath all of this sits a definitional question, and it is the one the title refers to.

Through the expansion, infrastructure was understood in substantially physical terms: fiber, conduit, electronics, huts, power, passings. That inventory was what got financed, what got counted, what appeared in the asset register, and what a buyer was understood to be buying.

Sustaining that plant over its actual life requires a second inventory, and it is one that is considerably harder to enumerate as a coherent whole. It includes accurate records of what was built and where it is. It includes physical evidence of conditions that cannot be re-derived later without excavation — depth, route, splice location, as-built deviation. It includes a labor pipeline capable of replacing technical capability as it retires, which requires training capacity that produces no output for a period. It includes maintainability as a design property: access, sectionalization, spares commonality, the ability to work on one part without taking down another. It includes vendor capacity that can survive a slow year. It includes preventive work, which by definition produces no visible result. And it includes knowledge continuity — the organizational memory of why the system is the way it is.

None of these are exotic. All of them require investment, some of it before the network carries a single customer and much of it continuously afterwards — a training pipeline has to keep producing, vendor capacity has to be kept viable between peaks, preventive work is by definition never finished, and knowledge continuity survives only if it is actively maintained. What they mostly do not do is announce their value while they are working. Their contribution becomes legible at the point where the capability is missing, degraded, or suddenly needed, which is a poor moment to discover the size of it.

This is the distinction the essay turns on: the price of the network and the cost of the capability required to keep the network are two different quantities. One of them has been measured with real precision for decades. The other has mostly been left to be discovered.

Erosion converges

A final observation about why the consequences arrived when they did rather than gradually.

Where an operating system deteriorates, it rarely does so along one dimension at a time. Experienced technical capability thins, maintenance backlogs accumulate, documentation drifts further from physical reality, and the expectations applied to returns harden — each on its own schedule.

Taken individually, each has been survivable, and each has in fact been survived for years. The systemic risk is not in any one of them. It is that the organization's capacity to absorb any one of them depends on the others still being intact — a thinning workforce is manageable while documentation is good, and poor documentation is manageable while experienced people remain. Compensating mechanisms are not independent, and the tolerance of the system is therefore much narrower than a list of individually tolerable conditions suggests.

What looks like a sudden deterioration in operating performance is more often several slow ones reaching the same point at once.

Was resilience priced?

The question this essay proposes as the serious one is narrower than did broadband fail, and more answerable.

An operator selling a network today can say with precision what plant is included. It is considerably harder to say what condition it is in, how accurate the records are, what proportion of the technical knowledge required to run it resides in individuals rather than systems, how much deferred work is embedded in it, or what it would cost to bring the operating capability to a level matching the physical asset. Those figures are not usually produced, and there is no settled method for producing them. These dimensions have not historically been measured together as one inventory, and assembling them consistently remains difficult.

That is the gap worth closing, and it is a technical and commercial problem rather than a moral one. An industry that can specify a network to the strand can learn to specify the capability required to sustain it. Until it does, the price of broadband infrastructure will continue to reflect what was built, and the cost of owning it will continue to reflect what was not.