The AI Buildout's Other Two-Year Lead Time [2026]

The Pipe

The Second Constraint: Everyone Is Watching the Power. The Fiber Takes Just as Long.

A new long-haul route takes one to two years to build. So does a data center. One of those is on every slide.

Written by Scot Free

Every piece written about the limits of the AI buildout is about electricity. Grid interconnect queues, gas turbines, nuclear restarts, substations. It is a real constraint and the coverage is correct.

There is a second one, it has roughly the same lead time, and it shows up in almost none of that coverage: the data centers have to be connected to each other.

Not to the internet. To each other. Training runs split across sites, inference distributed across regions, checkpoints and datasets moving between campuses. That is a transport problem, it is measured in route miles and strand counts, and the network that exists today was not built for it.

What's actually being bought

The demand signal here is unusually clean, because the companies selling the capacity publish what they sold.

• Long-haul dark fiber demand doubled in a year. Zayo analyzed purchasing across nearly 6,000 customers during 2025 and reported long-haul dark fiber demand up 2x year over year, metro dark fiber demand up as much as 20x in some AI-driven markets, and wavelength capacity purchases among data center customers up 2.6x.
• And the builds followed. Zayo is constructing more than 8,000 route miles — six new long-haul routes at roughly 3,000 miles, plus overbuilds across high-capacity corridors covering more than 5,000 — which the company calls the largest single network investment in new build and overbuild miles in its history. NVIDIA is the anchor customer.
• Lumen is going from 16.6 million to 47 million intercity fiber miles by 2028, against nearly $9 billion in AI and cloud connectivity contracts, installing more than 22,000 route miles of Corning fiber across 20-plus new routes.
• And the supply chain is being locked up. Lumen reserved 10% of Corning's global fiber capacity for each of two years under a supply agreement — its largest cable purchase ever. Zayo expanded its own Corning agreement in August 2026 to secure a significant portion of the cable for its expansion.

When buyers start reserving a tenth of a global supplier's output two years out, that is not a forecast. That is someone who has done the arithmetic on lead times.

The finding: it's a topology problem, not a capacity problem

Here is the part that reframes the whole thing, and it comes from the people building the routes.

Existing fiber routes were designed around legacy data center topologies. The biggest gaps are emerging along secondary corridors where high-capacity long-haul fiber is limited or does not yet exist.

Work out why that follows. A conventional data center was sited near population, peering, and existing fiber. An AI training campus is sited near power — cheap, available, and permittable at the scale of hundreds of megawatts. Those are different maps.

So the new campuses are landing in places the long-haul network was never routed through, and the fix is not "pull more strands through the conduit you have." The conduit does not go there. It is a new build, through new right of way, with new permits.

The scale of the split is published. One breakdown of what the AI buildout requires puts it at roughly 109 million fiber miles for upgrading existing routes, 66 million for new laterals connecting campuses to the backbone, and 38 million for entirely new long-haul routes. The upgrade half is the easy half. The other 104 million miles involve putting something in the ground where nothing is.

The lead time nobody puts on the slide

A new long-haul fiber route takes more than a year to build, approaching two — which is, as the people building them point out, roughly the same order of magnitude as constructing the data center it connects.

That equivalence is the whole argument of this piece. Two projects, comparable duration, and only one of them is in the capital plan at the start.

Three things compound it:

• The design life assumption broke. Long-haul builds were historically sized to last about ten years. That capacity is being absorbed far faster than planned, which means routes completed recently are already being overbuilt.
• Optics are their own constraint. Coherent optical equipment supply is tight enough to limit how fast existing routes can be upgraded, which pushes more of the demand toward new builds — the slow option.
• Nobody else is building at this scale. The complexity of long-haul construction has kept most providers out of it, which concentrates the capacity in a small number of hands and makes the lead time less negotiable, not more.

How the capacity is actually contracted

Worth understanding if you are reading these announcements, because the structure tells you how confident the buyer is.

Dark fiber is unlit strand — the customer leases or buys exclusive use and lights it with their own optics, controlling performance and capacity end to end. It gets contracted two ways: an indefeasible right of use, typically 20 to 30 years, or a shorter lease of five years or less.

The IRU is the tell. A twenty-to-thirty-year commitment on a route that does not exist yet is a buyer saying they expect to need this corridor for a generation. Verizon signed an agreement valued at more than $1 billion with Google for dedicated dark fiber to its data centers. Those are not capacity purchases. They are infrastructure positions.

What to do with this if you're building anything

• Price transport on the same timeline as power. If your site selection process has an interconnect queue analysis and no fiber route analysis, you have modeled one of two constraints with comparable duration.
• Ask what's in the ground, not what's on the map. Route maps show corridors. The questions that matter are strand count, conduit capacity, diversity of path, and whether the route was built for the topology you need or the one that existed when it was trenched.
• Treat "we'll add capacity later" as a two-year statement. On an existing route with available conduit and available optics, later is months. On a corridor that does not exist, later is the same horizon as the building itself.

The Underground Take

This is the site's thesis showing up at national scale. The model gets the credit, the pipe does the work, and right now the pipe has a two-year lead time that almost nobody is putting next to the power number.

The detail I keep returning to is the topology point, because it is the kind of thing that only shows up if you talk to the people doing the construction. Everyone frames this as a capacity question — more bandwidth, more strands, more wavelengths. The builders describe it as a routing question. The old network connected the places data centers used to be. The new campuses are where the power is. Those maps do not overlap, and no amount of capacity on the wrong route fixes a route that does not go there.

There is also a labor half to this, and it runs on the other site: the federal projection for the people who install and splice this fiber says the occupation shrinks 3% over the next decade, because the model is still watching the last build. Someone has to pull 104 million miles of new fiber, and the forecast has not noticed.

Sources

• Zayo, 2026 Bandwidth Report, as reported by Data Center Frontier (Aug. 2026) — purchasing activity across nearly 6,000 customers during 2025; long-haul dark fiber demand doubling year over year, metro dark fiber up to 20x in AI-driven markets, wavelength purchases among data center customers up 2.6x.
• Zayo newsroom and Fierce Network (Apr. 2026), Data Center Knowledge (Aug. 2026) — 8,000+ route miles under construction; six new long-haul routes (~3,000 miles) plus overbuilds across high-capacity corridors (5,000+ miles); largest single network investment in new build and overbuild miles in company history; NVIDIA as anchor customer; executive comments on legacy topologies, secondary-corridor gaps, optical supply constraints, build duration, and capacity being absorbed faster than the ten-year design assumption.
• Zayo, January 2025 release — US data center capacity at 25 GW, estimated 65–148 GW by 2030; more than $1 billion in AI-related long-haul deals in 2024 with $3 billion in pipeline.
• Lumen — 16.6 million to 47 million intercity fiber miles by 2028; nearly $9 billion in AI and cloud connectivity contracts; 22,000+ route miles of Corning fiber across 20+ new routes; reservation of 10% of Corning's global fiber capacity for each of two years.
• Fiber mile requirement breakdown (109 million for upgrades, 66 million for new laterals, 38 million for new long-haul routes) and dark fiber contracting structures (IRUs of 20–30 years; shorter leases of five years or less): Compute Law analysis, 2026.
• PwC, US Deals 2026 Midyear Outlook, Telecommunications — AI infrastructure as a second demand engine for fiber; 400G optical transport deployment; dark-fiber corridors between AI hubs. Verizon–Google dark fiber agreement valued at more than $1 billion, July 2026.

A note on these sources: most of the demand figures here come from the companies selling the capacity, and a vendor's own bandwidth report is a sales document as much as a research one. They are cited because the underlying commitments — supply agreements, route-mile construction, and signed IRUs — are expensive and verifiable in a way that a forecast is not. The fiber-mile requirement breakdown is a third-party analysis and should be treated as an estimate.

Corrections and additions: Scott@IoTunderground.com.