In 2001, in the trough of the dotcom bust, a generation of telecom engineers had just finished laying down what may be the single largest infrastructure overbuild in modern American history. By the most widely cited industry estimates from the early 2000s — Telegeography, Tier1 Research, FCC reports of the period — somewhere upwards of eighty percent of the long-haul fiber installed during the 1998–2002 buildout had never been lit. Hyperscaler usage since roughly 2010 has consumed a meaningful chunk of it, but even now, by every credible reconstruction I can find, long-haul utilization across the US remains well below half.

The cable is still in the ground. Most of it still works. We are running the entire commercial, consumer, and AI internet on a fraction of the capacity we paid to install twenty-five years ago.

What “dark” means

A fiber-optic cable is a bundle of glass strands. Each strand can carry information only if there is laser equipment at both ends to inject light, and amplifier or regeneration equipment at intervals to keep the signal alive. The cable in the ground is passive. The transmission electronics — the lasers, the modulators, the amplifiers, the wavelength-division multiplexing gear, the long-haul switching fabric — are active and expensive.

When telecoms talk about “lighting” fiber, they mean hooking up the active equipment at both ends. A “lit” strand is in service. A “dark” strand is sitting idle, ready to carry traffic if anyone bothers to plug in the lasers.

The buildout from 1998 to 2002 — driven by Level 3, Qwest, Williams Communications, MCI, Sprint, Global Crossing, and a dozen smaller carriers — laid an enormous quantity of fiber strand-equivalent across the United States. The rosy demand projections that justified the buildout did not materialize. By 2002, most of those carriers were in bankruptcy. The cable, however, didn’t go anywhere. It was sold off, consolidated, and rolled into balance sheets at fractions of construction cost. And it stayed dark.

Telecoms don’t publish utilization. The figures that get cited are reconstructions from filings, retired engineering disclosures, and industry survey data. But every rigorous reconstruction of the asset base I’ve found lands in the same neighborhood for the long-haul intercity capacity. Metro and last-mile utilization is much higher. The dark fiber is, almost entirely, the long-distance trunk capacity between major peering points.

Why it isn’t lit

The naive answer is that supply outran demand. That answer was correct in 2002 and has been wrong ever since. Demand for long-haul capacity has grown roughly thirty to sixty percent annually for two decades — driven first by streaming, then by mobile, then by cloud, now by AI training and inference traffic. By any rational projection from 2002, all of that fiber should be lit by now.

The structural answer is that lighting fiber is cheap and connecting lit fiber to anything useful is expensive. The cost stack of an active transcontinental link decomposes roughly like this:

The cable itself, amortized over forty years of expected life, is a small fraction of the total cost of operating the link.

The optical transmission gear at each end — modern coherent transceivers, ROADMs, amplifiers — has come down in price by orders of magnitude since 2002 and is now a manageable cost.

The expensive part is everything that turns a lit fiber into a useful network: cross-connects at carrier hotels, peering arrangements at major exchanges, IP routing equipment to terminate the optical signal into a switched network, last-mile distribution to the customers who actually pay for the bits, and — most expensively — the operations and engineering labor to keep all of that running with five-nines reliability.

A new fiber pair, lit between two unconnected points, is not a network. It is a fast path to nowhere. To make it useful, you need to extend it to where the customers are, peer it with the networks they’re on, terminate it at facilities that can route IP traffic, and operate it under the SLAs that enterprise and hyperscale customers demand. The dark fiber isn’t sitting idle because lighting it is hard. It’s sitting idle because the adjacent infrastructure required to monetize a lit pair has not been built out at the same pace.

The result is a peculiar economic equilibrium. The fiber owner can light a pair for a few hundred thousand dollars in optical gear. They cannot, for that price, turn it into revenue. So they wait — often for years, sometimes for decades — until a customer materializes who is willing to pay for end-to-end provisioning, including the entire stack of adjacent infrastructure. Hyperscalers (Google, Meta, Microsoft, Amazon) have been the dominant buyers since around 2010, lighting strands point-to-point between their own datacenters, where they own both ends of the connection and don’t need to peer with anyone.

Outside the hyperscaler corridors, most of the cable is still dark.

Why it matters now

The current AI infrastructure buildout is being framed in the press as a once-in-a-generation capital expansion — datacenters, transformers, water cooling, transmission lines, fab capacity. A chunk of the discussion is about whether the buildout will produce another 2002-style overbuild.

It is worth noting that we are running the largest distributed compute boom in history while sitting on top of an unfinished overbuild from the previous one. The dark fiber from 2001 is structurally adjacent to the GPU buildout in a way that is mostly going unspoken. Hyperscalers are using it. Independent inference operators, regional cloud providers, and AI startups mostly are not — because the inability to monetize a lit pair without owning both ends of it is exactly what kept it dark in 2002. The economics didn’t change.

There is a second-order question that interests me more: if the large majority of the previous generation’s strategic infrastructure can sit idle for twenty-five years despite continuously growing demand for the service it provides, what does that tell us about how to value strategic infrastructure overbuilds in general? The standard “Field of Dreams” narrative — build it and they will come — is empirically falsified by the dark fiber. Capacity does not summon use. The summoning requires the adjacent infrastructure (peering, switching, last-mile, operations) to be ready to absorb it. Capacity ahead of complementary capacity creates stranded assets that take a generation to monetize.

Apply that lens to the GPU buildout. Hyperscaler datacenters are vertically integrated — same problem-solving as hyperscaler dark-fiber lighting. The independent capacity (regional datacenters, GPU rental marketplaces, sovereign-AI national clusters) faces the same problem: GPUs are valuable only if the surrounding stack of inference orchestration, model hosting, customer pipelines, and compliance infrastructure is ready to monetize them. If that adjacent stack is not built at the same pace, expect a long dark-GPU period to follow — capacity in racks, cold, waiting.

The lesson

The dark fiber is not an artifact of irrational exuberance. It is an artifact of a deeper pattern: the cost of building infrastructure is dominated by the wrong line item. We measure the cable, the cement, the silicon. The cost that determines whether the infrastructure gets used is somewhere else — in the operations, the peering, the last mile, the contracts, the labor, the regulatory permissions. When the adjacent stack is missing, the visible asset goes dark.

This shows up everywhere. The reason most rural broadband subsidies underperform is not that the cable doesn’t get laid. It’s that nobody can sustain the peering and operations cost on a thin customer base. The reason high-speed rail in the US has stalled is not that we lack steel. It’s that the adjacent stack — eminent domain, freight-passenger track sharing, station integration, urban transit connection — is missing. The reason new nuclear power has been so slow to deploy is not the reactor. It’s the adjacent stack of regulatory permitting, transmission interconnection, and grid operator contracts.

The dark fiber is the cleanest example of the pattern because the visible asset worked, perfectly, the entire time. It just wasn’t enough. And nobody who is selling the next big infrastructure buildout — AI datacenters, fusion plants, mineral refining, whatever’s next — is talking about the adjacent stack as the binding constraint.

The fiber is in the ground. The lasers are cheap. The peering is hard.

That sentence is the entire story, and it is twenty-five years old.