Consider a data center campus that did everything the current playbook rewards. The developer chased inexpensive power to an exurban substation, locked in the megawatts, and energized the buildings on schedule. Everything the site needs in order to operate is in place, with a single exception: it cannot yet reach the wider network of carriers and internet exchanges that a tenant expects to connect to on the first day of a lease.

The fiber that would carry that traffic is still unbuilt, sitting somewhere between the campus and the nearest point where the broader internet can actually be reached, and until that connection is finished the campus earns nothing.

This is the failure that power-led development keeps running into, and it is worth understanding why, because the connection to the site is not a minor procurement detail. It is strategic infrastructure, and it belongs on the plan alongside power, land, and compute rather than on a purchase order at the very end of the process.

Part of the reason it slips is financial. On a site where land and construction can run to a billion dollars, the connection itself might cost tens of millions, and measured against the revenue a fully leased campus generates, that figure can look small enough to postpone. So it tends to get treated like a small purchase, ordered late and expected quickly, even though the price of the connection has almost nothing to do with the time it takes to build. On a remote site, that inexpensive connection is frequently one of the last elements to finish and the single thing standing between a powered campus and its first paying tenant.

For years, this rarely caused trouble, because data centers clustered in established markets where fiber was already in the ground. The AI build-out has redrawn that geography, sending developers after power and affordable land in places chosen precisely because they were undeveloped, which is the very reason the fiber routes that were once a given are not there.

It is tempting to assume the problem disappears wherever a network already exists, yet that is seldom the case. Across much of the country, there is fiber in the ground already, but it was built for another era: telephone service and consumer broadband, older long-haul routes, cable networks, and none of it was designed to carry the volume of traffic that a hyperscaler, or one of the newer AI-focused cloud providers known as neoclouds, now demands.

Serving a modern AI campus takes purpose-built, high-count fiber on diverse routes, with interconnection capacity measured in terabits per second and rising quickly as AI workloads and autonomous agents multiply. Whether a site sits in true greenfield or in a market the legacy carriers already cover, the developer is almost always building new, high-capacity connectivity rather than tapping into something already adequate.

Building the connection is its own construction project, and like any construction project, it depends on approvals and access that the developer does not control. On a site selected for power rather than connectivity, there may be no usable fiber for miles, so the route has to be engineered from scratch, parcel by parcel, until it reaches a point on the existing carrier network. Every jurisdiction along the way issues its own permits, and easements have to be negotiated one at a time with landowners who are under no obligation to move quickly.

Some of the longest delays come from physical infrastructure the developer has no authority over. Where the route needs to pass beneath an active railroad, the railroad has to review and permit the crossing, and on the major Class I freight lines that approval commonly takes six to eighteen months, with some individual crossings documented at more than a year. Where the route runs overhead on utility poles instead of underground, the company that owns those poles, usually the local electric utility, has to survey them and shift its existing lines to make room, a step the industry calls make-ready. Federal rules envision that work taking roughly five months, but because of utility backlogs it regularly takes nine to fourteen months in practice.

The fiber cable itself has become a constraint as well, and in this case the bottleneck is the supply chain rather than the route. AI data centers consume far more fiber than conventional facilities, and that surge in demand has outrun what manufacturers can produce, pushing lead times for the specialized cable close to a full year. Layered on top of all of this are the ordinary unknowns of any excavation, from wetlands reviews and archaeological surveys to the simple chance of hitting rock once crews begin to trench.

None of this makes connectivity the longest task on a campus, since power, with its multi-year interconnection queues, still holds that distinction. Connectivity is, however, the task most likely to be unfinished when everything else is ready, precisely because it usually starts last and relies on the parties least invested in the developer's deadline. A cost that looks small on the balance sheet turns out to carry one of the longest and least controllable schedules on the entire site.

This is why a promised light-up date, offered on its own, is worth very little. No developer or carrier controls a railroad's permitting queue or an electric utility's make-ready backlog, so a date is only as reliable as the approvals standing behind it. The commitments that actually hold are the ones with evidence attached: easements already secured, permits already filed, crossings already granted, and fiber already built close to the site, together with route diversity that can be demonstrated rather than merely asserted. The developers who avoid being stranded tend to be the ones who ask what has genuinely been secured, and who ask it before they close on the land.

The opposite mistake can be just as expensive. Developers who have been caught out once sometimes overcorrect, building every remote site to the highest specification a hypothetical hyperscaler might someday want and stranding capital that way instead, committing thirty million dollars to a connection where five million would have comfortably served the tenant they could realistically win. Over-building too early and starting too late are two versions of the same error: capital committed against the wrong variable, and it is worth remembering that connectivity alone cannot rescue a poorly chosen site.

In a market where capacity is routinely leased before it is even built, the connection is part of what the developer is selling. A site with a real, evidenced light-up date and a diverse fiber path can be marketed with confidence, while a powered site backed only by a promise cannot. Power determines whether the campus gets built; connectivity determines whether anyone can actually move in and use it. Planning the two together from the start and treating the connection as infrastructure to be evidenced rather than a purchase order filed at the end is what keeps an expensive, fully powered campus from sitting empty.